A Multilayer Ontology-based Hybrid Recommendation Model Ivan Cantador, Alejandro Bellogin, Pablo Castells Escuela Politecnica Superior Universidad Autonoma de madrid Campus de Cantoblanco, 28049, Madrid, Spain Hivan.cantador, alejandro. bellogin, pablo castells; @uames Abstract may enjoy similar items. However, in typical approaches, We propose a novel hybrid recommendation model in which the comparison between users is done globally, in such a user preferences and item features are described in terms of way that partial, but strong and useful similarities might be semantic concepts defined in domain ontologies. The missed. For instance, two people may have a highly exploitation of meta-information describing the coincident taste in cinema, but a very divergent one in recommended items and user profiles in a general, portable sports. The opinions of these people on movies could be way, along with the capability of inferring knowledge from highly valuable for each other, but risk to be ignored by he relations defined in the ontologies, are the key aspects of many collaborative recommender systems, because the he presented proposal. Taking advantage of the enhanced semantics representation, user profiles are compared at a global similarity between the users might be lot finer grain size than they are in usual recommender systems In our proposal we argue for the distinction of different More specifically, the concept, item, and user spaces are layers within the interests and preferences of users, as a clustered in a coordinated way, and the resulting clusters are useful refinement to produce better recommendations d to find similarities among individuals at multiple Depending on the current context, only a specific subset of semantic layers. Such layers correspond to implicit the segments (layers)of a user profile is considered Communities of Interest(Col), and enable collaborative order to establish her similarities with other people when a recommendations of enhanced precision. Our approach is recommendation has to be performed. Such models of tested in two sets of experiments: one including profiles nduced user networks or communities, partitioned at manually defi real users and another with different common semantic layers can be exploited in the automatically profiles based on data from the IMDb and movielens datasets recommendation processes in order to produce more accurate and context-sensitive results Keywords: hybrid recommender systems, communities of interest, ontology, user profiling o Our approach is based on an ontological representation the domain of discourse where user interests are defined The ontological space takes the shape of a semantic network of interrelated domain concepts and the user 1 Introduction profiles are initially described as weighted lists measuring Recommender systems emerged in the early nineties as a the user interests for those concepts. We propose here to thriving research area on its own, distinct from other exploit the links between users and concepts to extract related fields in Artificial Intelligence and Information relations among users according to common interests Retrieval. The area has undergone a considerable leap in Analysing the structure of the domain ontology and taking significance and potential value since then, with the boost of digital content and online businesses involving stocks of profiles, we cluster the domain concept space, and generate goods of different sorts. The volume, growth rate, ubiquity groups of interests shared by certain users. Thus, those of access, and to a large extent unstructured nature of users who share interests of a specific concept cluster are worldwide content challenge the limits of human connected in the corresponding community, where their processing capabilities and information access preference weights measure the degree of membership to technologies, putting at stake the effective utility of that cluster content, despite its actual value. It is in such settings where The rest of the paper has the following structure. Section recommender systems can make a great valuable 2 describes the different types of recommender systems contribution, by proactively scanning the space of choices. and their current limitations, and depicts which of them are and predicting the potential usefulness of items for each addressed by our proposal. Section 3 is dedicated to the particular user, without needing users to explicitly specify underlying ontology-based knowledge representation and their needs or query tor items of whose existence they item features and user preferences are expressed in terms based on the principle that users with common traits(in of domain ontologies, how they are extended using the their demographic data, behaviour, taste, opinions, etc. semantic relations of those structures, and how they are
A Multilayer Ontology-based Hybrid Recommendation Model Iván Cantador, Alejandro Bellogín, Pablo Castells Escuela Politécnica Superior Universidad Autónoma de Madrid Campus de Cantoblanco, 28049, Madrid, Spain {ivan.cantador, alejandro.bellogin, pablo.castells}@uam.es Abstract We propose a novel hybrid recommendation model in which user preferences and item features are described in terms of semantic concepts defined in domain ontologies. The exploitation of meta-information describing the recommended items and user profiles in a general, portable way, along with the capability of inferring knowledge from the relations defined in the ontologies, are the key aspects of the presented proposal. Taking advantage of the enhanced semantics representation, user profiles are compared at a finer grain size than they are in usual recommender systems. More specifically, the concept, item, and user spaces are clustered in a coordinated way, and the resulting clusters are used to find similarities among individuals at multiple semantic layers. Such layers correspond to implicit Communities of Interest (CoI), and enable collaborative recommendations of enhanced precision. Our approach is tested in two sets of experiments: one including profiles manually defined by real users and another with automatically generated profiles based on data from the IMDb and MovieLens datasets. Keywords: hybrid recommender systems, communities of interest, ontology, user profiling 1. Introduction Recommender systems emerged in the early nineties as a thriving research area on its own, distinct from other related fields in Artificial Intelligence and Information Retrieval. The area has undergone a considerable leap in significance and potential value since then, with the boost of digital content and online businesses involving stocks of goods of different sorts. The volume, growth rate, ubiquity of access, and to a large extent unstructured nature of worldwide content challenge the limits of human processing capabilities and information access technologies, putting at stake the effective utility of content, despite its actual value. It is in such settings where recommender systems can make a great valuable contribution, by proactively scanning the space of choices, and predicting the potential usefulness of items for each particular user, without needing users to explicitly specify their needs or query for items of whose existence they cannot be aware beforehand. Recommender systems are based on the principle that users with common traits (in their demographic data, behaviour, taste, opinions, etc.) may enjoy similar items. However, in typical approaches, the comparison between users is done globally, in such a way that partial, but strong and useful similarities might be missed. For instance, two people may have a highly coincident taste in cinema, but a very divergent one in sports. The opinions of these people on movies could be highly valuable for each other, but risk to be ignored by many collaborative recommender systems, because the global similarity between the users might be low. In our proposal we argue for the distinction of different layers within the interests and preferences of users, as a useful refinement to produce better recommendations. Depending on the current context, only a specific subset of the segments (layers) of a user profile is considered in order to establish her similarities with other people when a recommendation has to be performed. Such models of induced user networks or communities, partitioned at different common semantic layers can be exploited in the recommendation processes in order to produce more accurate and context-sensitive results. Our approach is based on an ontological representation of the domain of discourse where user interests are defined. The ontological space takes the shape of a semantic network of interrelated domain concepts and the user profiles are initially described as weighted lists measuring the user interests for those concepts. We propose here to exploit the links between users and concepts to extract relations among users according to common interests. Analysing the structure of the domain ontology and taking into account the semantic preference weights of the user profiles, we cluster the domain concept space, and generate groups of interests shared by certain users. Thus, those users who share interests of a specific concept cluster are connected in the corresponding community, where their preference weights measure the degree of membership to that cluster. The rest of the paper has the following structure. Section 2 describes the different types of recommender systems and their current limitations, and depicts which of them are addressed by our proposal. Section 3 is dedicated to the underlying ontology-based knowledge representation and basic content retrieval of our proposal. We describe how item features and user preferences are expressed in terms of domain ontologies, how they are extended using the semantic relations of those structures, and how they are
exploited for basic content-based recommendations. The mechanism to cluster the concept space in several layers of 'reference Profile shared semantic interests for building multi-level relations chant the uwer'spm between users is presented in Section 4. The exploitation of the derived communities to enhance collaborative filtering is described in Section 5. The empirical evaluation of that model is presented in Section 6. As already mentioned, two different experiments are described: one using user profiles manually defined by real users and other conducted with artificial user profiles built from data of the well-known IMDb and MovieLens repositories. Section 7summarises related work, and finally, we conclude with some discussions and future research lines in Section 8 Figure I General process followed by a recommendersystem 2. Background In this scenario, the main difficulty lies in that the utility function g is usually not defined in the entire uxI space, recommendation problem can be formulated as but only on some subset of it. In recommender systems, the ows [1]. Let l=(u, uz. uu) be the set of all users utility function is defined only on the items that have been registered in the system, and let I=(i i2 in) be the set previously rated by the users, and it has to be extrapolated to the whole uxt sPace of all possible items that can be recommended. Let g(ua is ) be a utility function that measures the gain or Thus based on the mechanism in which item ratings are usefulness of item i to user l4x工→R estimated for different users, the following two main types where R is a totally ordered set (e.g. non negative of recommender systems can be distinguished: 1)content integers or real numbers within a certain range). Then, for based recommender systems, in which the user is each user u∈ll, we aim to choose the item i"∈工 recommended items similar to those he preferred in the that maximises the users utility. More formally past, and, 2)collaborative filtering systems, in which the user is recommended items that people with similar tastes Vu∈l,= arg max g(un,) and preferences liked in the past. Due to the limitations of each of the above strategies, combinations of them have The utility of an item is usually represented by a rating, een investigated in the so-called hybrid recommender measuring how much a specific user is(or is predicted to systems, empirically demonstrating their better be)interested in a specific item. Depending on the effectiveness application, the ratings can either be specified by the users, Nowadays, the interest in recommender systems is on or computed by the application. Each element of the user rise, constituting an integral part of a number of spaceycanbedescribedwithprofilethatmightincluderecommendedAmazon.com[341.whererecommendations several demographic characteristics, such as gender, age, nationality, marital status, etc, or some information about of books, CDs, and other products are done, or Google set of characteristics. For example, in a movie recommender system, movies can be described not only by improvements to make the recommendation algorithms directors. etc more effective and able to a broader range of real Figure I shows a general schema of a recommendation world applications [1[10]. As we explain later, these process. Firstly, the system manually or automaticall improvements include, among others, the application of captures the target users preferences, building her personal strategies that address situations in which few ratings are features of the preferred items, evaluations or ratings of with major flexibility and interpretability for the users, and profile. These preferences are defined as explicit content available over certain items. the use of recommendations those items, or as implicit tastes/interests information the study of more scalable algorithms that allow to make acquired from the user's behaviour or utilisation of the recommendations not only for a single user, but also for a system. Once the user profile is created, it is somehow group of people with similar tastes and interests compared against the items stored in the system, and those The strategies to confront the previous and other aspects items which are most appropriate are recommended are currently open research issues in the field. Here, we Depending on the algorithm implemented to choose the propose the use of Semantic Web technologies to address most appropriate items, we shall distinguish several some of them. Specifically, we present a hybrid recommendation model based on ontologies which offers a ategorisations for the recommender systems, and identify novel contribution to the scientific community that works on recommender systems. The opportunity to add meta
exploited for basic content-based recommendations. The mechanism to cluster the concept space in several layers of shared semantic interests for building multi-level relations between users is presented in Section 4. The exploitation of the derived communities to enhance collaborative filtering is described in Section 5. The empirical evaluation of that model is presented in Section 6. As already mentioned, two different experiments are described: one using user profiles manually defined by real users and other conducted with artificial user profiles built from data of the well-known IMDb and MovieLens repositories. Section 7 summarises related work, and finally, we conclude with some discussions and future research lines in Section 8. 2. Background The recommendation problem can be formulated as follows [1]. Let U= ( ) 1 2 , ,..., M uu u be the set of all users registered in the system, and let I = ( ) 1 2 , ,..., N ii i be the set of all possible items that can be recommended. Let ( , ) m n gu i be a utility function that measures the gain or usefulness of item ni to user m u , i.e., g : U × →I R , where R is a totally ordered set (e.g. non negative integers or real numbers within a certain range). Then, for each user um ∈ U , we aim to choose the item ∈ I max, mu i that maximises the user’s utility. More formally: ∈ ∀∈ = I U max, , arg max ( , ) m n u m mn i u i gu i The utility of an item is usually represented by a rating, measuring how much a specific user is (or is predicted to be) interested in a specific item. Depending on the application, the ratings can either be specified by the users, or computed by the application. Each element of the user space U can be described with a profile that might include several demographic characteristics, such as gender, age, nationality, marital status, etc., or some information about the user’s tastes, interests and preferences. Analogously, each element of the item space I can be described with a set of characteristics. For example, in a movie recommender system, movies can be described not only by their titles, but also by their genres, principal actors, directors, etc. Figure 1 shows a general schema of a recommendation process. Firstly, the system manually or automatically captures the target user’s preferences, building her personal profile. These preferences are defined as explicit content features of the preferred items, evaluations or ratings of those items, or as implicit tastes/interests information acquired from the user’s behaviour or utilisation of the system. Once the user profile is created, it is somehow compared against the items stored in the system, and those items which are most appropriate are recommended. Depending on the algorithm implemented to choose the most appropriate items, we shall distinguish several categorisations for the recommender systems, and identify different subsets of items that can be retrieved. Figure 1 General process followed by a recommender system In this scenario, the main difficulty lies in that the utility function g is usually not defined in the entire U×I space, but only on some subset of it. In recommender systems, the utility function is defined only on the items that have been previously rated by the users, and it has to be extrapolated to the whole U×I space. Thus, based on the mechanism in which item ratings are estimated for different users, the following two main types of recommender systems can be distinguished: 1) contentbased recommender systems, in which the user is recommended items similar to those he preferred in the past, and, 2) collaborative filtering systems, in which the user is recommended items that people with similar tastes and preferences liked in the past. Due to the limitations of each of the above strategies, combinations of them have been investigated in the so-called hybrid recommender systems, empirically demonstrating their better effectiveness. Nowadays, the interest in recommender systems is on the rise, constituting an integral part of a number of important websites like MovieLens [27], where movies are recommended, Amazon.com [34], where recommendations of books, CDs, and other products are done, or Google News Personalization [20], a system for recommending news. In all of them, the use of recommendation methods has been very successful. However, the current generation of recommender systems still requires further improvements to make the recommendation algorithms more effective and applicable to a broader range of realworld applications [1][10]. As we explain later, these improvements include, among others, the application of strategies that address situations in which few ratings are available over certain items, the use of recommendations with major flexibility and interpretability for the users, and the study of more scalable algorithms that allow to make recommendations not only for a single user, but also for a group of people with similar tastes and interests. The strategies to confront the previous and other aspects are currently open research issues in the field. Here, we propose the use of Semantic Web technologies to address some of them. Specifically, we present a hybrid recommendation model based on ontologies which offers a novel contribution to the scientific community that works on recommender systems. The opportunity to add meta-
information to the descriptions of the recommended items More formally, and following the notation used in [11, and the preferences of the users, together with the let Content(i ) be the content description of item i,EI capability of inferring knowledge from the relations i.e., the set of content features characterising i, that are existent in the used domain ontologies are the key aspects used to determine the appropriateness of the item for the of the presented proposal different users. This description is usually represented as a Before introducing our recommendation model and its vector of real numbers(weights), in which each compone enefits, we briefly describe the characteristics of content measures the importance"(or"informativeness")of the based, collaborative filtering and hybrid recommender corresponding feature in the item content description systems, and explain their main current limitations Content(i 2. 1. Content-based recommender systems Since content-based recommender systems were Content-based approaches to recommendation making designed mostly to recommend textual items, the contents [7[8]][46] build on the conjecture that a person of the items are usually described with keywords. Hence, likes items with features similar to those of other items he for example, the content-based component of the Fab or she liked in the past [54]. Thus, the utility gain function system [5] represents web page contents in terms of the g(un) of item,∈ a for user u∈ u is estimated based on the utilities a(,, i,)assigned by user u, to items i 28 most informative words Analogously, let Content Based User Profile (u)be the that ggest movies to user u, a content-based recommender weighted item content features that describe the tastes, ystem would try to understand the commonalities among interests and needs of the user movies user u. has previously evaluated positively specific genres, preferred actors and directors, etc Content Based UserPr()=un=(,2…,1)∈R In content-based recommender systems, items are suggested according to a comparison between their The utility gain of item i, for user u is then calculated descriptions and the user profiles, which contain with a score function that combines the different item information about the users' tastes. interests. and needs description and user profile components Data structures for both of these components are created using features extracted from the content of the items. a g(u,i)=score( Content Based User Profile(u), Content(L ))ER weighting scheme is often used for providing high weights to the most discriminating features and preferences, and Different content-based recommendation approaches low weights to the less informative ones have been proposed in the literature to formulate the Figure 2 shows the general process followed by a previous expression. Basically, these techniques are content-based recommender system. Firstly, the users classified in heuristic-based and model-based approaches eferences are established according to the content The first ones calculate utility predictions based on features of those items preferred/selected by her. The heuristic formulas that are inspired mostly on information preferences existing in this profile are compared against retrieval methods, such as the cosine similarity measure the features of the items stored in the system choice set, The second ones, on the other hand, obtain utility and the items whose features are most similar to the user's predictions based on a model learned from the underlying content-based preferences are finally retrieved. Note that in statistical learning and machine learning this scenario only the items that share content-based models, such as Bayesian classifiers, clustering algorithms features with the user profiles can be suggested, reducin decision trees and artificial neural networks drastically the set of items that might be recommended to For both types of techniques, several limitations have each individual user been identified in the literature [15[10]. We describe some of them nex Relevence Nhb or Restricted content analysis. Content-based recommendations are restricted by the features that are explicitly associated with the items to be recommended. For example, content-based movie recommendations can only be based on written materials about a movie: actors' names, plot summaries, cinematographic genres, etc effectiveness of these techniques thus depends on the descriptive data available. Therefore, in order to have a sufficient set of features. the content should either Targtww be in a form that can be automatically parsed by a Figure 2 Content-based recommendations computer or in a form in which the features can be manually extracted in an easy way. In many cases these situations are very difficult to achieve. There
information to the descriptions of the recommended items and the preferences of the users, together with the capability of inferring knowledge from the relations existent in the used domain ontologies are the key aspects of the presented proposal. Before introducing our recommendation model and its benefits, we briefly describe the characteristics of contentbased, collaborative filtering and hybrid recommender systems, and explain their main current limitations. 2.1. Content-based recommender systems Content-based approaches to recommendation making [7][8][31][33][43][46] build on the conjecture that a person likes items with features similar to those of other items he or she liked in the past [54]. Thus, the utility gain function ( , ) m n gu i of item i n ∈ I for user um ∈ U is estimated based on the utilities ( , ) m l gu i assigned by user m u to items l i that are “similar” to item n i . For instance, in order to suggest movies to user m u a content-based recommender system would try to understand the commonalities among movies user m u has previously evaluated positively: specific genres, preferred actors and directors, etc. In content-based recommender systems, items are suggested according to a comparison between their descriptions and the user profiles, which contain information about the users’ tastes, interests, and needs. Data structures for both of these components are created using features extracted from the content of the items. A weighting scheme is often used for providing high weights to the most discriminating features and preferences, and low weights to the less informative ones. Figure 2 shows the general process followed by a content-based recommender system. Firstly, the user’s preferences are established according to the content features of those items preferred/selected by her. The preferences existing in this profile are compared against the features of the items stored in the system choice set, and the items whose features are most similar to the user’s content-based preferences are finally retrieved. Note that in this scenario only the items that share content-based features with the user profiles can be suggested, reducing drastically the set of items that might be recommended to each individual user. Figure 2 Content-based recommendations More formally, and following the notation used in [1], let ( ) n Content i be the content description of item i n ∈ I , i.e., the set of content features characterising n i that are used to determine the appropriateness of the item for the different users. This description is usually represented as a vector of real numbers (weights), in which each component measures the “importance” (or “informativeness”) of the corresponding feature in the item content description: ( ) == ∈ ( ,1 ,2 , , ,..., ) RK n n n n nK Content i i i i i Since content-based recommender systems were designed mostly to recommend textual items, the contents of the items are usually described with keywords. Hence, for example, the content-based component of the Fab system [5] represents web page contents in terms of the 128 most informative words. Analogously, let ( ) m ContentBasedUserProfile u be the content-based preferences of user um ∈ U , i.e., the weighted item content features that describe the tastes, interests and needs of the user. ( ) == ∈ ( ,1 ,2 , , ,..., ) RK m m m m mK ContentBasedUserProfile u u u u u The utility gain of item n i for user m u is then calculated with a score function that combines the different item description and user profile components: g u i score ContentBasedUserProfile u Content i ( mn m n , , ) = ( ( ) ( ))∈R Different content-based recommendation approaches have been proposed in the literature to formulate the previous expression. Basically, these techniques are classified in heuristic-based and model-based approaches. The first ones calculate utility predictions based on heuristic formulas that are inspired mostly on information retrieval methods, such as the cosine similarity measure. The second ones, on the other hand, obtain utility predictions based on a model learned from the underlying data using statistical learning and machine learning models, such as Bayesian classifiers, clustering algorithms, decision trees, and artificial neural networks. For both types of techniques, several limitations have been identified in the literature [1][5][10]. We describe some of them next. • Restricted content analysis. Content-based recommendations are restricted by the features that are explicitly associated with the items to be recommended. For example, content-based movie recommendations can only be based on written materials about a movie: actors’ names, plot summaries, cinematographic genres, etc. The effectiveness of these techniques thus depends on the descriptive data available. Therefore, in order to have a sufficient set of features, the content should either be in a form that can be automatically parsed by a computer or in a form in which the features can be manually extracted in an easy way. In many cases, these situations are very difficult to achieve. There
are some domains that have an inherent problem with as an approximate representation of her interests and automatic feature extraction, and it is often not needs in the domain of application practical to assign features by hand due to limitations matched against ratings submitted by =一。 ratings are of resources. For instance, it is much harder to apply obtaining the users set of "nearest neighbours". The items automatic feature extraction methods to multimedia hat were rated highly by the user's nearest neighbours and ta,e.g, graphical images, video streams, and audio were not rated by the user will finally be recommended records The way in which the user's"neighbours"are determined, and the strategy followed to combine the ratings of such Content overspecialisation. Content-based users will differentiate the existent CF approache recommender systems only retrieve items that score highly against a specific user profile. They cannot With the above ideas, the definitions of the user profile recommend items that are different from anything the and the item description given in this section for content ased recommender systems differ from those associated user has seen before. Thus, for example, a person to CF recommender systems. Specifically, let with no experience in Spanish cuisine would never Collaborative User Profile(u )=r=(a, ' aMer. be the receive recommendations for even the best Spanish restaurant n town collaborative profile of user u constituted by the set of ratings provided by the user to the N items stored in the Cold-start: new user problem. A user has to rate a system, and let Ratings()==(5…,4)∈R"behe sufficient number of items before a content-based of ratings r∈ R assigned to item i, by the M users recommender system can really understand her registered in the system. In both of the above definitions, if preferences and present him with reliable user u. has not rated item i, then r.=0. The utility recommendations. A new user having none or very ain of item i, for user u. is then computed by a score few ratings may not be suggested any accurate function that combines the different user profile and item description components Portfolio effect: non diversity problem. In certain ases, items should not be recommended if they are q(u,.)=score( Collaborative User Profile(u), Ratings()ER too similar to something the user has already seen. The different formulations given for the previous For example, it is not necessarily a good idea to expression [91[481511[] have lead to two main recommend all movies by Antonio Banderas to a user categories of CF techniques: user-based and item-based who liked one of them in the past, or it could not be approaches appropriate to recommend news articles describing active user s the same event ratings with those of other users to identify g a group of 2. 2. Collaborative filtering recommender systems be recommended to that user Collaborative filtering (CF) techniques [22][301 [47[48511[53] match people with similar preferences in order to make recommendations. Unlike content-based methods, collaborative recommender systems try to predict ⊙0◎ the utility of items for a particular user according to the items previously evaluated by other users. In other words, the utility gain function g(um, i, ) of item i, EI for user Eu is estimated based on the utilities g(un i )assigned to item i, by those users u, that are" similar to user u The great power of the CF approaches relative to Inside the b recommendation ability [101, i.e., the chance of recommending items that do not share content features A-y"for twg pressed in the user profiles. For example, it may be that listeners who enjoy free jazz also enjoy avant-garde classical music. but a content-based recommender trained Figure 3 User-based collaborative filtering recommendations on the preferences of a free jazz aficionado would not be The items preferred by the most similar users are recommended able to suggest items in the classical music realm to the active user none of the features(performers, instruments, repertories) associated with items in the different categories would be Item-based CF approaches, on the other hand, take each shared. Only by looking outside the preferences of the item of the active users list of rated items and recommend ndividual can such suggestions be made other items that seem to be similar to that item according to In CF systems, the users express other users ratings rating items. The ratings submitted by a user are thus used
are some domains that have an inherent problem with automatic feature extraction, and it is often not practical to assign features by hand due to limitations of resources. For instance, it is much harder to apply automatic feature extraction methods to multimedia data, e.g., graphical images, video streams, and audio records. • Content overspecialisation. Content-based recommender systems only retrieve items that score highly against a specific user profile. They cannot recommend items that are different from anything the user has seen before. Thus, for example, a person with no experience in Spanish cuisine would never receive recommendations for even the best Spanish restaurant in town. • Cold-start: new user problem. A user has to rate a sufficient number of items before a content-based recommender system can really understand her preferences and present him with reliable recommendations. A new user having none or very few ratings may not be suggested any accurate recommendations. • Portfolio effect: non diversity problem. In certain cases, items should not be recommended if they are too similar to something the user has already seen. For example, it is not necessarily a good idea to recommend all movies by Antonio Banderas to a user who liked one of them in the past, or it could not be appropriate to recommend news articles describing the same event. 2.2. Collaborative filtering recommender systems Collaborative filtering (CF) techniques [22][30] [47][48][51][53] match people with similar preferences in order to make recommendations. Unlike content-based methods, collaborative recommender systems try to predict the utility of items for a particular user according to the items previously evaluated by other users. In other words, the utility gain function ( , ) m n gu i of item i n ∈ I for user um ∈ U is estimated based on the utilities ( , ) l n gui assigned to item n i by those users l u that are “similar” to user m u . The great power of the CF approaches relative to content-based ones is its “outside the box” recommendation ability [10], i.e., the chance of recommending items that do not share content features expressed in the user profiles. For example, it may be that listeners who enjoy free jazz also enjoy avant-garde classical music, but a content-based recommender trained on the preferences of a free jazz aficionado would not be able to suggest items in the classical music realm since none of the features (performers, instruments, repertories) associated with items in the different categories would be shared. Only by looking outside the preferences of the individual can such suggestions be made. In CF systems, the users express their preferences by rating items. The ratings submitted by a user are thus used as an approximate representation of her tastes, interests and needs in the domain of application. These ratings are matched against ratings submitted by all other users, obtaining the user’s set of “nearest neighbours”. The items that were rated highly by the user’s nearest neighbours and were not rated by the user will finally be recommended. The way in which the user’s “neighbours” are determined, and the strategy followed to combine the ratings of such users will differentiate the existent CF approaches. With the above ideas, the definitions of the user profile and the item description given in this section for contentbased recommender systems differ from those associated to CF recommender systems. Specifically, let ( ) == ∈ ( ,1 ,2 , , ,..., ) N m m m m mN CollaborativeUserProfile u r r r r R be the collaborative profile of user mu constituted by the set of ratings provided by the user to the N items stored in the system, and let ( ) == ∈ ( 1, 2, , , ,..., ) M Ratings i r r r n n n n Mn r R be the set of ratings r m n, ∈R assigned to item n i by the M users registered in the system. In both of the above definitions, if user m u has not rated item n i , then r m n, = ∅ . The utility gain of item n i for user m u is then computed by a score function that combines the different user profile and item description components: g u i score CollaborativeUserProfile u Ratings i ( mn m n , , ) = ( ( ) ( ))∈R The different formulations given for the previous expression [9][48][51][53] have lead to two main categories of CF techniques: user-based and item-based approaches. User-based CF approaches compare the active user’s ratings with those of other users to identify a group of similar people. The highest rated items of that group will be recommended to that user. Figure 3 User-based collaborative filtering recommendations. The items preferred by the most similar users are recommended to the active user Item-based CF approaches, on the other hand, take each item of the active user’s list of rated items, and recommend other items that seem to be similar to that item according to other users’ ratings
available it might become very difficult to categorise 888 the user's interests Cold-start: new item problem. Collaborative filtering recommender systems only rely on users eferences to make recommendations and do not ke use of content information of the existing items Until a new item is rated by a substantial number of users, the recommender system would not be able to recommend it. a recent item that has not obtained many ratings cannot be easily recommended. This problem shows up in domains such as news articles where there is a constant stream of new items and each user only rates a few [52] Gray sheep problem. For the user whose tastes are Figure 4 Item-based collaborative filtering unusual compared to the rest of the population, there The items which have been most similarly evaluated are will not be any other users who are particularly recommended to the active user recommendations Collaborative recommenders work best for a user Pure collaborative filtering recommender systems who fits into a cluster with many neighbours of confront some of the weaknesses existing in content-base similar tastes. However, the techniques do not work approaches. Since collaborative strategies make use of well for the so-called "Gray sheep", those people who other users recommendations(ratings), they can deal with fall on a border between two cliques of users. This is any kind of content and recommend any item, even the also a problem for demographic systems that attempt ones that are dissimilar to those seen in the past. However to categorise users according to personal collaborative techniques suffer from their own limitations Portfolio effect: non diversity problem. Since Sparse rating problem. In collaborative filtering collaborative filtering systems' knowledge about systems, the number of ratings already obtained is content is purely derived from usually very small compared to the number of ratings recommendations are strongly based toward what has needed to be predicted. In practice, many commercial been chosen(or recommended) in the past, resulting systems, such as Amazon. com which recommends in frequent recommendations of just the most popular CDNow. com which recommends music This could make collaborative filtering a po albums, are used to evaluate very large datasets where cry tool for the end user, often failing to even active users may have rated well under 1% of produce an interesting diversity of recommended the existent items [ 50]. The success of collaborative content filtering recommendations depends on the availability of a critical mass of users. They are based on the 2.3. Hybrid recommender systems verla in ratings across users and have difficulties when the space of ratings is sparse, i.e., few users Hybrid recommender systems 3151[17[251301 38][4555] combi have rated the same items. There might be many filtering techniques under a single framework, mitigating ems that have been rated by only a few people and inherent limitations of either paradigm. Thus, hybrid these items would be recommended very rarely, even recommendations are generated taking into account both if those few users gave high ratings to them Moreover, if the set of items changes too rapidly old descriptive features and ratings ratings will be of little value to new users who will Numerous ways for combining content-based and ot be able to have their ratings compared to those of collaborative filtering information are conceivable [1[10] the existing users Among them, the most widely adopted is the so-called collaborative via content" paradigm [451, where content- Cold-start: new user problem. Collaborative based profiles are built to detect similarities among users ltering strategies learn the users' preferences only Based on the taxonomy of hybridization methods given from the ratings they have given. When a new user [10], hybrid recommender systems can be classified utilises the system no personal ratings are available follows for her, and no proper recommendations can be made Because recommendations follow from a comparison Weighted hybrid recommenders. These systems between the target user and other users, based solel suggest items with scores that are obtained from the on the accumulation of ratings, if few ratings are results of all their individual recommendation techniques. Those results are usually merged by linear combination or vote consensus schemes. The benefit
Figure 4 Item-based collaborative filtering recommendations. The items which have been most similarly evaluated are recommended to the active user Pure collaborative filtering recommender systems confront some of the weaknesses existing in content-based approaches. Since collaborative strategies make use of other users’ recommendations (ratings), they can deal with any kind of content and recommend any item, even the ones that are dissimilar to those seen in the past. However, collaborative techniques suffer from their own limitations [1][5][10], as described next. • Sparse rating problem. In collaborative filtering systems, the number of ratings already obtained is usually very small compared to the number of ratings needed to be predicted. In practice, many commercial systems, such as Amazon.com which recommends books or CDNow.com which recommends music albums, are used to evaluate very large datasets where even active users may have rated well under 1% of the existent items [50]. The success of collaborative filtering recommendations depends on the availability of a critical mass of users. They are based on the overlap in ratings across users and have difficulties when the space of ratings is sparse, i.e., few users have rated the same items. There might be many items that have been rated by only a few people and these items would be recommended very rarely, even if those few users gave high ratings to them. Moreover, if the set of items changes too rapidly, old ratings will be of little value to new users who will not be able to have their ratings compared to those of the existing users. • Cold-start: new user problem. Collaborative filtering strategies learn the users’ preferences only from the ratings they have given. When a new user utilises the system no personal ratings are available for her, and no proper recommendations can be made. Because recommendations follow from a comparison between the target user and other users, based solely on the accumulation of ratings, if few ratings are available it might become very difficult to categorise the user’s interests. • Cold-start: new item problem. Collaborative filtering recommender systems only rely on users’ preferences to make recommendations, and do not make use of content information of the existing items. Until a new item is rated by a substantial number of users, the recommender system would not be able to recommend it. A recent item that has not obtained many ratings cannot be easily recommended. This problem shows up in domains such as news articles where there is a constant stream of new items and each user only rates a few [52]. • Gray sheep problem. For the user whose tastes are unusual compared to the rest of the population, there will not be any other users who are particularly similar, leading to poor recommendations. Collaborative recommenders work best for a user who fits into a cluster with many neighbours of similar tastes. However, the techniques do not work well for the so-called “Gray sheep”, those people who fall on a border between two cliques of users. This is also a problem for demographic systems that attempt to categorise users according to personal characteristics. • Portfolio effect: non diversity problem. Since collaborative filtering systems’ knowledge about content is purely derived from user choices, recommendations are strongly based toward what has been chosen (or recommended) in the past, resulting in frequent recommendations of just the most popular items. This could make collaborative filtering a poor discovery tool for the end user, often failing to produce an interesting diversity of recommended content. 2.3. Hybrid recommender systems Hybrid recommender systems [3][5][17][25][30] [38][45][55] combine content-based and collaborative filtering techniques under a single framework, mitigating inherent limitations of either paradigm. Thus, hybrid recommendations are generated taking into account both descriptive features and ratings. Numerous ways for combining content-based and collaborative filtering information are conceivable [1][10]. Among them, the most widely adopted is the so-called “collaborative via content” paradigm [45], where contentbased profiles are built to detect similarities among users. Based on the taxonomy of hybridization methods given in [10], hybrid recommender systems can be classified as follows: • Weighted hybrid recommenders. These systems suggest items with scores that are obtained from the results of all their individual recommendation techniques. Those results are usually merged by linear combination or vote consensus schemes. The benefit
of these methods is that all the recommendation Meta-level hybrid recommenders. These systems apabilities are straightforward incorporated in the combine two recommendation techniques by using recommendation process. However, they have the the entire model generated by one as the input for implicit assumption that the relative value of the another. The benefit of these methods, especially for a different techniques is more or less uniform across the ontent-based/collaborative hybrid approach, is that space of items fact that is not always true. For the learned (content-based)model is a com example, from the discussion on the limitations of representation of the users interests, and the collaborative filtering given previously, it is known (collaborative) recommendation mechan that a CF system will be weaker for those items with follows can operate on this information-dense more small number of rating Switched hybrid recommenders. These sy stems use Hybrid recommenders based on feature some criterion to switch between recommendation ugmentation. These systems, similarly to cascade techniques. The benefit of these methods is that the hybrids, involve a staged process. A first suggestions can be sensitive to the strengths and recommendation technique produces a rating or weakness of the constituent recommendation classification of each item. Afterwards. a second However, they introduce additional recommendation technique takes the obtained complexity into the recommendation process since information and incorporates it into Its the switching criteria must be determined with recommendation process. Note that these approaches another level of parameterization are different to cascade ones since in the latter the first recommendation technique has no influence over Mixed hybrid recommenders. These systems present together the suggestions given by the different the second. the benefit of these methods is that it recommendation techniques. The benefit of these offers a way to improve the performance of core methods is that they directly exploit the benefits of recommendation techniques, enriching their inputs collaborative and without modifying their internal model recommendations. However, they require ranking of items, or selection of a single best suggestion. part from the specific weaknesses of both content- entailing the development of some kind of based and collaborative recommendation approaches, there combination technique exist other general limitations in the current recommender Hybrid recommenders based on feature These Poor understanding of users and items. Most of the content/collaborative suggestions treatin recommender systems produce ratings that are based collaborative information as simply additional on a limited information about users and items as atures associated with each item, and using content captured by user and item profiles, and do not take based techniques over the augmented dataset. The ull advantage of information from users' behaviour, benefit of these methods is that collaborative data is transactional histories and other available data. For considered without relying on it exclusively, reducing example, classical collaborative filtering methods rely thus the sensitivity of the recommendations to the exclusively on the ratings information to make recommendations. Although there has been some gress made on incorporating user and item profiles Cascade hybrid recommenders. These systems into some of the methods since the early days of involve a staged process. A first recommendation recommender systems, these profiles tend to be quite technique produces a coarse ranking of candidates simple and do not utilise more advanced profiling Afterwards, a second recommendation technique uses techniques. In addition to using traditional pI the previous filtered candidate set, refining the final features such as keywords and simple demograp suggestions. The benefit of these methods is that they more advanced profiling techniques based or void employing the second, lower-priority technique mining could be used, finding recommendation rules on items that are well differentiated by the first behaviour and usage patterns, etc. technique, or are sufficiently poorly-rated that they will never be recommended. Doing this, cascade contextual formation recommenders perform efficient recommendation process. Traditional recommender recommendations than, for example, a weighted systems operate on the two-dimensional User x Item hybrid recommender that has to apply allits space. That is, they make their recommendations techniques to all items. In addition, the cascade based only on the user and item information, and do approach is by its nature tolerant to noise in the low- not take into consideration additional contextual iority technique, since recommendations given by information that may be crucial in some applications high-priority recommender can only be refined However, in many situations, the utility of a certain item to a user may depend significantly on time, the people with whom the item will be consumed or
of these methods is that all the recommendation capabilities are straightforward incorporated in the recommendation process. However, they have the implicit assumption that the relative value of the different techniques is more or less uniform across the space of items – fact that is not always true. For example, from the discussion on the limitations of collaborative filtering given previously, it is known that a CF system will be weaker for those items with a small number of ratings. • Switched hybrid recommenders. These systems use some criterion to switch between recommendation techniques. The benefit of these methods is that the suggestions can be sensitive to the strengths and weakness of the constituent recommendation techniques. However, they introduce additional complexity into the recommendation process since the switching criteria must be determined with another level of parameterization. • Mixed hybrid recommenders. These systems present together the suggestions given by the different recommendation techniques. The benefit of these methods is that they directly exploit the benefits of both content-based and collaborative recommendations. However, they require ranking of items, or selection of a single best suggestion, entailing the development of some kind of combination technique. • Hybrid recommenders based on feature combination. These systems merge content/collaborative suggestions treating the collaborative information as simply additional features associated with each item, and using contentbased techniques over the augmented dataset. The benefit of these methods is that collaborative data is considered without relying on it exclusively, reducing thus the sensitivity of the recommendations to the number of ratings. • Cascade hybrid recommenders. These systems involve a staged process. A first recommendation technique produces a coarse ranking of candidates. Afterwards, a second recommendation technique uses the previous filtered candidate set, refining the final suggestions. The benefit of these methods is that they avoid employing the second, lower-priority technique on items that are well differentiated by the first technique, or are sufficiently poorly-rated that they will never be recommended. Doing this, cascade recommenders perform more efficient recommendations than, for example, a weighted hybrid recommender that has to apply all its techniques to all items. In addition, the cascade approach is by its nature tolerant to noise in the lowpriority technique, since recommendations given by high-priority recommender can only be refined. • Meta-level hybrid recommenders. These systems combine two recommendation techniques by using the entire model generated by one as the input for another. The benefit of these methods, especially for a content-based/collaborative hybrid approach, is that the learned (content-based) model is a compressed representation of the user’s interests, and the second (collaborative) recommendation mechanism that follows can operate on this information-dense more easily than on the initial raw data. • Hybrid recommenders based on feature augmentation. These systems, similarly to cascade hybrids, involve a staged process. A first recommendation technique produces a rating or classification of each item. Afterwards, a second recommendation technique takes the obtained information and incorporates it into its recommendation process. Note that these approaches are different to cascade ones, since in the latter the first recommendation technique has no influence over the second. The benefit of these methods is that it offers a way to improve the performance of core recommendation techniques, enriching their inputs and without modifying their internal model. Apart from the specific weaknesses of both contentbased and collaborative recommendation approaches, there exist other general limitations in the current recommender systems. • Poor understanding of users and items. Most of the recommender systems produce ratings that are based on a limited information about users and items as captured by user and item profiles, and do not take full advantage of information from users’ behaviour, transactional histories and other available data. For example, classical collaborative filtering methods rely exclusively on the ratings information to make recommendations. Although there has been some progress made on incorporating user and item profiles into some of the methods since the early days of recommender systems, these profiles tend to be quite simple and do not utilise more advanced profiling techniques. In addition to using traditional profile features such as keywords and simple demographics, more advanced profiling techniques based on data mining could be used, finding recommendation rules, behaviour and usage patterns, etc. • No contextual information within the recommendation process. Traditional recommender systems operate on the two-dimensional User × Item space. That is, they make their recommendations based only on the user and item information, and do not take into consideration additional contextual information that may be crucial in some applications. However, in many situations, the utility of a certain item to a user may depend significantly on time, the people with whom the item will be consumed or
shared and under which circumstances. For example, 2.4. Our proposal a user can have significantly different preferences for the types of movies she wants to see when she is As explained in the introduction, and as shall be described going out to a movie theatre with her boyfriend on a in detail in the next sections, we propose a multilayered Saturday night, as opposed to watching a rental movie approach to hybrid recommendation, based on the at home with her parents on a Wednesday evenin automatic identification of Col from semantic user Using multidimensional settings, the inclusion of ferences stored in well-structured ontology-based user knowledge about the user's task/environment into the profiles. Our method builds and compares profiles of user recommendation algorithm can lead to better interests for semantic topics and specific concepts in order to find similarities among users. The issue of finding hidden links between users and items based on the Non flexible recommendations. In similarity of the user preferences/interests (expressed by recommendation methods are inflexible in the sense means of opinions, comparatives or ratings of items) and that they support a predefined and fixed set of the item content features is the essence of the already recommendations. Moreover, most of them only presented hybrid recommender systems. But in contrast recommend individual items to individual users, and to classic collaborative strategies, the comparison is done do not deal with aggregation of items and/or users in our approach by splitting the user profiles into clusters Group recommendations [36](371[44] are still open of cohesive interests, and based on this, several layers of investigation and innovations. Col are found. This provides a richer model of Non support for multi-criteria ratings. Most of the find common interests in real life. According to the criterion ratings. However, it is important to be able taxonomy of hybrid recommender systems given to provide aggregated recommendations in a number previousl ur approach adopts the so-called f applications, such as recommend brands or collaborative via content"paradigm [45] and can be categories of items to certain segments of users. In categorised as a meta-level hybrid recommender. The some applications, it is crucial to incorporate mult users' interests are represented as semantic concepts of criteria ratings into recommendation methods. Mult domain ontologies, and a collaborative recommendation criteria ratings have been extensively studied in the mechanism is then applied which takes into account the similarities between such content-based user profiles Operation Research community Our proposal addresses some of the limitations of Scalability problem. Nearest neighbour-based current recommender systems, including both content- algorithms require computation that grows with the based and collaborative filtering strategies. As we ill umber of users and the number of items. With show, the semantic relations between concepts and millions of users and items, a typical web-based instances of the knowledge ontologies, are exploited in our recommender system running existing algorithms will approach to reduce the impact of problems such as suffer serious scalability problems. For them, efficient restricted content analysis, preference/rating sparsity, cold clustering techniques are thus needed. There exist a start, content overspecialisation, or portfolio effects. number of dimensionality reduction techniques [501 Moreover, through our mechanism for identifying uch as Singular Value Decomposition (SVD) multilayered communities of interest, we are able to 11[32], and efficient clustering strategies, suc discover relations between users at different levels co-clustering 24 augmenting the possibilities of finding similarities for those users without very common/popular interests(gray Intrusiveness. Many recommender systems ar sheep problem). Moreover, our user profile representatio intrusive in the sense that they require explicit and content retrieval mechanism are open to new strategies feedback from the user and often at a significant level of user involvement. Some non-intrusive methods of for group-oriented, context-aware, query-driven and multi- criteria recommendations. research fields which we have getting user feedback have been presented in the iterature. However. non-intrusive ratings are often already started to investigate We shall show results obtained from empirical inaccurate and cannot fully replace explicit rating evaluations of the model. As we explain in last sections provided by the user. Therefore, the problem of we conducted experiments with two different repositories, minimising intrusiveness while maintaining certain manually obtained from real users, and automatically levels of accuracy of recommendations needs to be enerated merging information from IMDb and Movielens addresse Need of explainability. The recommender should have the ability of explaini causes,inferences performed from the user considered constraints. etc
shared and under which circumstances. For example, a user can have significantly different preferences for the types of movies she wants to see when she is going out to a movie theatre with her boyfriend on a Saturday night, as opposed to watching a rental movie at home with her parents on a Wednesday evening. Using multidimensional settings, the inclusion of knowledge about the user’s task/environment into the recommendation algorithm can lead to better recommendations. • Non flexible recommendations. In general, recommendation methods are inflexible in the sense that they support a predefined and fixed set of recommendations. Moreover, most of them only recommend individual items to individual users, and do not deal with aggregation of items and/or users. Group recommendations [36][37][44] are still open to investigation and innovations. • Non support for multi-criteria ratings. Most of the current recommender systems deal with single criterion ratings. However, it is important to be able to provide aggregated recommendations in a number of applications, such as recommend brands or categories of items to certain segments of users. In some applications, it is crucial to incorporate multicriteria ratings into recommendation methods. Multicriteria ratings have been extensively studied in the Operation Research community. • Scalability problem. Nearest neighbour-based algorithms require computation that grows with the number of users and the number of items. With millions of users and items, a typical web-based recommender system running existing algorithms will suffer serious scalability problems. For them, efficient clustering techniques are thus needed. There exist a number of dimensionality reduction techniques [50], such as Singular Value Decomposition (SVD) [21][32], and efficient clustering strategies, such as co-clustering [24]. • Intrusiveness. Many recommender systems are intrusive in the sense that they require explicit feedback from the user and often at a significant level of user involvement. Some non-intrusive methods of getting user feedback have been presented in the literature. However, non-intrusive ratings are often inaccurate and cannot fully replace explicit ratings provided by the user. Therefore, the problem of minimising intrusiveness while maintaining certain levels of accuracy of recommendations needs to be addressed. • Need of explainability. The recommender systems should have the ability of explaining the recommendations they present to the user [26]: causes, inferences performed from the user profile, considered constraints, etc. 2.4. Our proposal As explained in the introduction, and as shall be described in detail in the next sections, we propose a multilayered approach to hybrid recommendation, based on the automatic identification of CoI from semantic user preferences stored in well-structured ontology-based user profiles. Our method builds and compares profiles of user interests for semantic topics and specific concepts in order to find similarities among users. The issue of finding hidden links between users and items based on the similarity of the user preferences/interests (expressed by means of opinions, comparatives or ratings of items) and the item content features is the essence of the already presented hybrid recommender systems. But in contrast to classic collaborative strategies, the comparison is done in our approach by splitting the user profiles into clusters of cohesive interests, and based on this, several layers of CoI are found. This provides a richer model of interpersonal links, which better represents the way people find common interests in real life. According to the taxonomy of hybrid recommender systems given previously, our approach adopts the so-called “collaborative via content” paradigm [45] and can be categorised as a meta-level hybrid recommender. The users’ interests are represented as semantic concepts of domain ontologies, and a collaborative recommendation mechanism is then applied which takes into account the similarities between such content-based user profiles. Our proposal addresses some of the limitations of current recommender systems, including both contentbased and collaborative filtering strategies. As we will show, the semantic relations between concepts and instances of the knowledge ontologies, are exploited in our approach to reduce the impact of problems such as restricted content analysis, preference/rating sparsity, coldstart, content overspecialisation, or portfolio effects. Moreover, through our mechanism for identifying multilayered communities of interest, we are able to discover relations between users at different levels, augmenting the possibilities of finding similarities for those users without very common/popular interests (gray sheep problem). Moreover, our user profile representation and content retrieval mechanism are open to new strategies for group-oriented, context-aware, query-driven and multicriteria recommendations, research fields which we have already started to investigate. We shall show results obtained from empirical evaluations of the model. As we explain in last sections, we conducted experiments with two different repositories, manually obtained from real users, and automatically generated merging information from IMDb and MovieLens repositories
3. Ontology-based recommendations Furthermore, ontology standards, such as RDF and OWL support inference mechanisms that can be used to In this section, we present our approach to the semantic enhance personalisation, so that, for instance, a user description of user preferences and items in terms of interested in animals (superclass of cat) is also concepts and instances defined in domain ontologies. We recommended items about cats. Inversely, a user interested also present a basic content-based recommendation model n lizards and snakes can be inferred with a certain that is used as the base line approach for the experiments confidence to be interested in reptiles. Similarly, a user performed with our hybrid recommendation models fascinated about the life of actors and actresses can be recommended items in which for example the name of Brad 3. 1. Knowledge representation Pitt appears, due to that person could be an instance of the class Actor. Also, a user keen on Spain can be assumed to In contrast to other strategies in personalised content like Madrid, through the locatedIn transitive relation. These retrieval, our approach makes use of explicit user profiles (as opposed to e.g. sets of preferred documents ). Working characteristics are exploited in our recommendation models within an ontology-based personalisation framework [56] user preferences are represented as vectors 3.2. Content-based recommendation model u=(um. w,,,Amw,x)where m E[0, 1] measures the With the presented knowledge representation, we use a (a class or an instance) in a do main ontology o, being that warks d two pmhasesnt lt im the trst on a form al the total number of concepts in the ontology. Similarly, the ontology-based query is issued by some form of query items d, ED in the retrieval space are assumed to be interface(e.g. NLP-based)formalising a user information notate vectors d,=(di, d need. The query is processed, outputting a set of ontology weights, in the same vector-space as user preferences concepts that satisfy it. From this point, the second phase is Based on this common logical representation, measures of based on an adaptation of the classic vector-space IR user interest for content items can be computed by model [41[49], where the axes of the space are the concepts comparing preference and annotation vectors, and these of O, instead of text keywords. The query and each item satisfaction of a query by an item can be computed bi. the measures can be used to prioritise, filter and rank contents are thus represented by vectors q and d, so tha (a collection, a catalogue, a search result) in a personal cosine measure pace ontology-based Semantic User Profile respectively er of users and items registered in the Weighted system. Users (MI Fina/ Ranked 88 88 8 Porsonal Annotations Repository Figure 6 Personalised ontology-based content retrieval Annotations 首/ The problem, of course, is how to build the q and d vectors. For more details, see [151[57]. Here we obviate this issue, and continue explaining our content retrieval Figure 5 Ontology-based user profiles and item descriptions Personalised Ranking in Figure 6). Once a user profile is e The ontology-based representation is richer and less obtained. our notion of content retrieval is based on a ambiguous than a keyword-based or item-based model. It matching algorithm that provides a personal relevance provides an adequate grounding for the representation of measure pref(d, u) of an item d for a user u.This coarse to fine-grained user interests(e.g. interest for items measure is set according to the semantic preferences of the such as a sports team, an actor, a stock value), and can be a user and the semantic annotations of the item based again key enabler to deal with the subtleties of user preferences on a cosine-based vector similarity An ontology provides further formal, computer-processable meaning on the concepts(who is coaching a team, an actors filmography, financial data on a stock), and makes it available for the personalisation system to take advantage of ResourceDescriptionfRameworkwww.w3.org/rdf 2WebOntologyLanguagewww.w3.org/2004/owl
3. Ontology-based recommendations In this section, we present our approach to the semantic description of user preferences and items in terms of concepts and instances defined in domain ontologies. We also present a basic content-based recommendation model that is used as the base line approach for the experiments performed with our hybrid recommendation models. 3.1. Knowledge representation In contrast to other strategies in personalised content retrieval, our approach makes use of explicit user profiles (as opposed to e.g. sets of preferred documents). Working within an ontology-based personalisation framework [56], user preferences are represented as vectors ,1 ,2 , ( , ,..., ) m m m mK u = uu u where [ ] , 0,1 m k u ∈ measures the intensity of the interest of user m u ∈ U for concept k c ∈O (a class or an instance) in a domain ontology O , K being the total number of concepts in the ontology. Similarly, the items n d ∈D in the retrieval space are assumed to be annotated by vectors ,1 ,2 , ( , ,..., ) n n nK = dd d dn of concept weights, in the same vector-space as user preferences. Based on this common logical representation, measures of user interest for content items can be computed by comparing preference and annotation vectors, and these measures can be used to prioritise, filter and rank contents (a collection, a catalogue, a search result) in a personal way. Figure 5 shows our twofold-space ontology-based knowledge representation, in which M and N are respectively the number of users and items registered in the system. Figure 5 Ontology-based user profiles and item descriptions The ontology-based representation is richer and less ambiguous than a keyword-based or item-based model. It provides an adequate grounding for the representation of coarse to fine-grained user interests (e.g. interest for items such as a sports team, an actor, a stock value), and can be a key enabler to deal with the subtleties of user preferences. An ontology provides further formal, computer-processable meaning on the concepts (who is coaching a team, an actor’s filmography, financial data on a stock), and makes it available for the personalisation system to take advantage of. Furthermore, ontology standards, such as RDF1 and OWL2 , support inference mechanisms that can be used to enhance personalisation, so that, for instance, a user interested in animals (superclass of cat) is also recommended items about cats. Inversely, a user interested in lizards and snakes can be inferred with a certain confidence to be interested in reptiles. Similarly, a user fascinated about the life of actors and actresses can be recommended items in which for example the name of Brad Pitt appears, due to that person could be an instance of the class Actor. Also, a user keen on Spain can be assumed to like Madrid, through the locatedIn transitive relation. These characteristics are exploited in our recommendation models. 3.2. Content-based recommendation model With the presented knowledge representation, we use a retrieval model (component ‘Item retrieval’ in Figure 6) that works in two phases [16]. In the first one, a formal ontology-based query is issued by some form of query interface (e.g. NLP-based) formalising a user information need. The query is processed, outputting a set of ontology concepts that satisfy it. From this point, the second phase is based on an adaptation of the classic vector-space IR model [4][49], where the axes of the space are the concepts of O , instead of text keywords. The query and each item are thus represented by vectors q and d , so that the satisfaction of a query by an item can be computed by its cosine measure. Figure 6 Personalised ontology-based content retrieval The problem, of course, is how to build the q and d vectors. For more details, see [15][57]. Here we obviate this issue, and continue explaining our content retrieval process with its personalisation phase (component ‘Personalised Ranking’ in Figure 6). Once a user profile is obtained, our notion of content retrieval is based on a matching algorithm that provides a personal relevance measure pref d u ( , ) of an item d for a user u . This measure is set according to the semantic preferences of the user and the semantic annotations of the item based again on a cosine-based vector similarity: 1 Resource Description Framework, www.w3.org/RDF 2 Web Ontology Language, www.w3.org/2004/OWL
pref(d, u=cos(d,)di performance when using ontology-based profiles, instead ld of classical keyword-based preferences representations We have conducted several experiments showing that In order to bias the result of a search(the ranking)to th the performance of the personalisation system is preferences of the user, the above measure has to be considerably poorer when the spreading mechanism is not combined with the query-based enabled. Typically, the basic user profiles without personalisation sim(d, q) defined previously, to produce a expansion are too simple. They provide a good combined ranking [16] representative sample of user preferences, but do not In real scenarios, user profiles tend to be very scattered reflect the real extent of user interests which results in low especially in those recommender systems where user verlaps between the preferences of different users profiles have to be manually defined. Users are usually not Therefore, the extension is not only important for the willing to spend time describing their detailed preferences performance of individual personalisation, but it is to the system, even less to assign weights to them. essential for the clustering strategy described in the especially if they do not have a clear understanding of the following sections effects and results of this input. On the other hand applications where an automatic preference learning algorithm is applied tend to recognise the main 4. Multilayered Communities of Interest characteristics of user preferences, thus yielding profiles In social communities, it is commonly accepted that people that may entail a lack of expressivity. To overcome this who are known to share a specific interest are likely to problem, we propose a semantic preference spreading mechanism, which expands the initial set of preferences who share interests in travelling might be also keen on stored in user profiles through explicit semantic relation with other concepts in the ontology(see Figure 7). Our topics related in photography, gastronomy or languages. In fact, this assumption is the essence of the collaborative approach is based on the Constrained Spreading Activation filtering systems. We assume this hypothesis here as well (CSA)strategy [18 19]. The expansion is self-controlled by applying a decay factor to the intensity of preference preferences shared by several users Space in groups of in order to cluster the concept each time a relation is traversed Taking advantage of the relations between concepts, and the (weighted) preferences of users for the concepts, we correlation of concepts appearing in the preferences of individual users. After this, user profiles are partitioned by =/。0 Q projecting the concept clusters into the set of preferences of each user. Then, users can be compared on the basis of the resulting subsets of interests, in such a way that several rather than just one, weighted) links can be found betweer 88a Specifically, a vector c,=(Ck,I,CK, 23,, M) is assigned to each concept vector C, present in the preferences of at 89°9 least one user. where c in the semantic profile of user u. Based on these 888 vectors a classic hierarchical clustering strategy [623 is applied. The clusters obtained(Figure 8)represent the 86 groups of preferences( topics of interests)in the concept- Ontology concept user vector space shared by a significant number of users Users(M 8③ Figure 7 Semantic preference extension 88/ Thus, the system outputs ranked lists of content items taking into account not only the preferences of the current M+ od Ontology concepts user, but also a semantic spreading mechanism through the user profile and the domain ontology. In fact, experiments were done without the semantic spreading process and poor results were obtained. The profiles were very simple Figure 8 Semantic concept clustering based on the shared and the matching between the preferences of different interests of the users users was low. This observation shows a better
() () · pref d u, cos , = = d u d u d u In order to bias the result of a search (the ranking) to the preferences of the user, the above measure has to be combined with the query-based score without personalisation sim d q ( , ) defined previously, to produce a combined ranking [16]. In real scenarios, user profiles tend to be very scattered, especially in those recommender systems where user profiles have to be manually defined. Users are usually not willing to spend time describing their detailed preferences to the system, even less to assign weights to them, especially if they do not have a clear understanding of the effects and results of this input. On the other hand, applications where an automatic preference learning algorithm is applied tend to recognise the main characteristics of user preferences, thus yielding profiles that may entail a lack of expressivity. To overcome this problem, we propose a semantic preference spreading mechanism, which expands the initial set of preferences stored in user profiles through explicit semantic relations with other concepts in the ontology (see Figure 7). Our approach is based on the Constrained Spreading Activation (CSA) strategy [18][19]. The expansion is self-controlled by applying a decay factor to the intensity of preference each time a relation is traversed. Figure 7 Semantic preference extension Thus, the system outputs ranked lists of content items taking into account not only the preferences of the current user, but also a semantic spreading mechanism through the user profile and the domain ontology. In fact, experiments were done without the semantic spreading process and poor results were obtained. The profiles were very simple and the matching between the preferences of different users was low. This observation shows a better performance when using ontology-based profiles, instead of classical keyword-based preferences representations. We have conducted several experiments showing that the performance of the personalisation system is considerably poorer when the spreading mechanism is not enabled. Typically, the basic user profiles without expansion are too simple. They provide a good representative sample of user preferences, but do not reflect the real extent of user interests, which results in low overlaps between the preferences of different users. Therefore, the extension is not only important for the performance of individual personalisation, but it is essential for the clustering strategy described in the following sections. 4. Multilayered Communities of Interest In social communities, it is commonly accepted that people who are known to share a specific interest are likely to have additional connected interests. For instance, people who share interests in travelling might be also keen on topics related in photography, gastronomy or languages. In fact, this assumption is the essence of the collaborative filtering systems. We assume this hypothesis here as well, in order to cluster the concept space in groups of preferences shared by several users. Taking advantage of the relations between concepts, and the (weighted) preferences of users for the concepts, we propose to cluster the semantic space based on the correlation of concepts appearing in the preferences of individual users. After this, user profiles are partitioned by projecting the concept clusters into the set of preferences of each user. Then, users can be compared on the basis of the resulting subsets of interests, in such a way that several, rather than just one, (weighted) links can be found between two users. Specifically, a vector ck k k kM = (cc c ,1 ,2 , , ,..., ) is assigned to each concept vector k c present in the preferences of at least one user, where km mk , , c u = is the weight of concept k c in the semantic profile of user m u . Based on these vectors a classic hierarchical clustering strategy [6][23] is applied. The clusters obtained (Figure 8) represent the groups of preferences (topics of interests) in the conceptuser vector space shared by a significant number of users. Figure 8 Semantic concept clustering based on the shared interests of the users
Once the concept clusters are created, each user is The figure represents a situation where four user clusters assigned to a specific cluster. The similarity between a are obtained. Based on them, user profiles are partitioned user's preferences um=(um, 1, l/m 2,,"m, k ) and a cluster in four semantic layers On each layer, weighted relations C。 Is computed by among users are derived, building up different communities of interest ∑nk The resulting communities have many potential applications. For one, they can be exploited to the benefit of collaborative filtering and recommendation, not only because they establish similarities between users, but also where ck represents the concept that corresponds to the because they provide powerful means to focus on different Umu component of the user preference vector, and Cis semantic contexts for different information needs. The the number of concepts included in the cluster. The clusters design of two recommendation models in this direction is with highest similarities are then assigned to the users, thus explored in next section creating groups of users with shared interests(Figure 9) Users (M 868 5. Multilayered hybrid recommendations 86 In this section, we present two hybrid recommendations models that exploit the similarities existing between the 三。 Ontology concepts ontology-based descriptions of the items to be retrieved and the multilayered user similarities obtained in the community of interest identification mechanism described in the previous section Figure 9 Groups of users obtained from the semantic concept 51 Recommendation models We believe that exploiting the relations of the underlying The concept and user clusters are then used to find social networks which emerges from the users'interests emergent, focused semantic social networks. The information can have an important benefit in collaborative measures between clusters are used to find relations multilayered community of interest proposal explained in between two distinct types of social items: individuals and the previous section, we present here two recommendation models that generate ranked lists of items in different groups of individuals On the other hand, using the concept clusters user generated social networks. The first model(that we shall label as UP)is based on the whole semantic profile of the these segments corresponds to a concept cluster and user to whom a unique ranked list is delivered. The second represents a subset of the user interests that is shared by the model(labelled UP-q) outputs a ranking for each semantic users who contributed to the clustering process. By thus cluster C introducing further structure in user profiles, it is now possible to define relations among users at different levels, The two strategies are formalised next. In the following for a user profile u., an information object vector d, and obtaining a multilayered network of users. Figure 10 illustrates this idea a cluster Ca, we denote by um and di the projection of the corresponding concept vectors onto cluster Cg,1.e. the k-th component of um and dg m, and respectively, if c, E Ca, and 0 otherwise Model up The semantic profile of a user u is used by the system to return a unique ranked list. The preference score of an item 中 d. is computed as a weighted sum of the indirect preference values based on similarities with other users in each cluster, where the sum is weighted by the similarities with the clusters. as follows: layered semantic Communities of Interest built pref(dn,un)=∑nm(dc)∑mim,(mn)sm,(d,x)
Once the concept clusters are created, each user is assigned to a specific cluster. The similarity between a user’s preferences ,1 ,2 , ( , ,..., ) m m m mK u = uu u and a cluster Cq is computed by: ( ) , , k q m k c C m q q u sim u C C ∈ = ∑ where k c represents the concept that corresponds to the um,k component of the user preference vector, and Cq is the number of concepts included in the cluster. The clusters with highest similarities are then assigned to the users, thus creating groups of users with shared interests (Figure 9). Figure 9 Groups of users obtained from the semantic concept clusters The concept and user clusters are then used to find emergent, focused semantic social networks. The preference weights of user profiles, the degrees of membership of the users to each cluster and the similarity measures between clusters are used to find relations between two distinct types of social items: individuals and groups of individuals. On the other hand, using the concept clusters user profiles are partitioned into semantic segments. Each of these segments corresponds to a concept cluster and represents a subset of the user interests that is shared by the users who contributed to the clustering process. By thus introducing further structure in user profiles, it is now possible to define relations among users at different levels, obtaining a multilayered network of users. Figure 10 illustrates this idea. Figure 10 Multilayered semantic Communities of Interest built from the obtained clusters The figure represents a situation where four user clusters are obtained. Based on them, user profiles are partitioned in four semantic layers. On each layer, weighted relations among users are derived, building up different communities of interest. The resulting communities have many potential applications. For one, they can be exploited to the benefit of collaborative filtering and recommendation, not only because they establish similarities between users, but also because they provide powerful means to focus on different semantic contexts for different information needs. The design of two recommendation models in this direction is explored in next section. 5. Multilayered hybrid recommendations In this section, we present two hybrid recommendations models that exploit the similarities existing between the ontology-based descriptions of the items to be retrieved, and the multilayered user similarities obtained in the community of interest identification mechanism described in the previous section. 5.1. Recommendation models We believe that exploiting the relations of the underlying social networks which emerges from the users’ interests, and combining them with semantic item preference information can have an important benefit in collaborative filtering and recommendation. Using our semantic multilayered community of interest proposal explained in the previous section, we present here two recommendation models that generate ranked lists of items in different scenarios taking into account the links between users in the generated social networks. The first model (that we shall label as UP) is based on the whole semantic profile of the user to whom a unique ranked list is delivered. The second model (labelled UP-q) outputs a ranking for each semantic cluster Cq . The two strategies are formalised next. In the following, for a user profile m u , an information object vector n d , and a cluster Cq , we denote by q m u and q n d the projection of the corresponding concept vectors onto cluster Cq , i.e. the k-th component of q m u and q n d are m k, u and n k, d respectively, if k q c C ∈ , and 0 otherwise. Model UP The semantic profile of a user m u is used by the system to return a unique ranked list. The preference score of an item n d is computed as a weighted sum of the indirect preference values based on similarities with other users in each cluster, where the sum is weighted by the similarities with the clusters, as follows: ( , , , · (,) ) ( ) ( ) nm n q q mi q ni q i pref d u nsim d C nsim u u sim d u = ∑ ∑