Ontology-Based Recommender Systems Stuart E. Middleton, David De Roure, and Nigel R. Shadbolt2 I IT Innovation Centre, University of Southampton, Southampton SO16 7NP, UF sem@it-innovation, soton, ac uk 2 Intelligence, Agents, Multimedia Group, Department of Electronics and Computer Science, University of Southampton, Southampton SO17 1B, UK, ddergecs, soton. ac uk. nrsgecs, soton. ac uk Summary. We present an overview of the latest approaches to using ontologies in recommender systems and our work on the problem of recommending on-line academic research papers. Our two experimental systems, Quickstep and Foxtrot create user profiles from unobtrusively monitored behaviour and relevance feedback representing the profiles in terms of a research paper topic ontology. A novel profile visualization approach is taken to acquire profile feedback. Research papers are clas- sified using ontological classes and collaborative recommendation algorithms used to recommend papers seen by similar people on their current topics of interest. Onto- logical inference is shown to improve user profiling, external ontological knowledge used to successfully bootstrap a recommender system and profile visualization em- ployed to improve profiling accuracy. In a specific case study we report results from two small-scale experiments, with ubjects over 3 months, and a large-scale experiment, with 260 subjects over an academic year, are conducted to evaluate different aspects of our approach. The over- all performance of our ontological recommender systems are favourably compared to other systems in the literature 1 Introduction The mass of content available on the World-Wide Web raises important ques- tions over its effective use. Information on the web is largely unstructured with web pages authored by many people on a diverse range of topics. This often makes simple browsing too time consuming to be practical. The emer- gence of e-commerce sites means many vendors are offering potentially great deals on very similar products. Web information filtering has thus become necessary for most web users in order to find the things they really need Recommender systems have emerged as one successful approach that can help tackle the problem of information overload. They exploit patterns in item metadata and reviews posted by groups of people to find new items that might S Staab and R. Studer(eds ) Handbook on Ontologies, International Handbooks 779 on Information Systems, DOI 10.1007/978-3-540-92673-3 C Springer-Verlag Berlin Heidelberg 2009
Ontology-Based Recommender Systems Stuart E. Middleton1, David De Roure2, and Nigel R. Shadbolt2 1 IT Innovation Centre, University of Southampton, Southampton SO16 7NP, UK, sem@it-innovation.soton.ac.uk 2 Intelligence, Agents, Multimedia Group, Department of Electronics and Computer Science, University of Southampton, Southampton SO17 1BJ, UK, dder@ecs.soton.ac.uk, nrs@ecs.soton.ac.uk Summary. We present an overview of the latest approaches to using ontologies in recommender systems and our work on the problem of recommending on-line academic research papers. Our two experimental systems, Quickstep and Foxtrot, create user profiles from unobtrusively monitored behaviour and relevance feedback, representing the profiles in terms of a research paper topic ontology. A novel profile visualization approach is taken to acquire profile feedback. Research papers are classified using ontological classes and collaborative recommendation algorithms used to recommend papers seen by similar people on their current topics of interest. Ontological inference is shown to improve user profiling, external ontological knowledge used to successfully bootstrap a recommender system and profile visualization employed to improve profiling accuracy. In a specific case study we report results from two small-scale experiments, with 24 subjects over 3 months, and a large-scale experiment, with 260 subjects over an academic year, are conducted to evaluate different aspects of our approach. The overall performance of our ontological recommender systems are favourably compared to other systems in the literature. 1 Introduction The mass of content available on the World-Wide Web raises important questions over its effective use. Information on the web is largely unstructured, with web pages authored by many people on a diverse range of topics. This often makes simple browsing too time consuming to be practical. The emergence of e-commerce sites means many vendors are offering potentially great deals on very similar products. Web information filtering has thus become necessary for most web users in order to find the things they really need. Recommender systems have emerged as one successful approach that can help tackle the problem of information overload. They exploit patterns in item metadata and reviews posted by groups of people to find new items that might S. Staab and R. Studer (eds.), Handbook on Ontologies, International Handbooks 779 on Information Systems, DOI 10.1007/978-3-540-92673-3, c Springer-Verlag Berlin Heidelberg 2009
780 S.E. Middleton et al e of interest to a user Ontologies are increasingly being used within the field of recommender systems, allowing knowledge-based techniques to supplement classical machine learning and statistical approaches 1.1 Recommender Systems People find articulating exactly what they want difficult, but they are good at recognizing it when they see it. This insight has led to the utilization of relevance feedback, where people rate items as interesting or not interesting and the system tries to find items that match the"interesting", positive examples and do not match the not interesting", negative examples. With sufficient positive and negative examples, modern machine learning techniques can classify new pages with impressive accuracy. Recommender systems can recommend many types of item, including web pages, new articles, music CDs and books Unobtrusive monitoring provides positive examples of what the user is looking for, without interfering with the user's normal work activity. Heuris- tics can also be applied to infer negative examples from observed behaviour although recommender systems, which unobtrusively watch user behaviour and recom mend new items that correlate with a user's profile Another way to recommend items is based on the ratings provided by other people who have liked the item before Collaborative recommender systems do this by asking people to rate items explicitly and then recommend new items that similar users have rated highly. An issue with collaborative filtering is that there is no direct reward for providing examples since they only help other people. This leads to initial difficulties in obtaining a sufficient number of ratings for the system to be useful, a problem known as the cold-start problem 15 Hybrid systems, attempting to combine the advantages of content-based and collaborative recommender systems, have also proved popular to-date The feedback required for content-based recommendation is shared, allowing collaborative recommendation as well 1.2 User Profiling User profiling is typically either knowledge-based or behaviour-based Knowledge-based approaches use static models of users and dynamically match users to the closest model. Questionnaires and interviews are often employed to obtain this user knowledge. Once a model is selected for a user, specialist domain knowledge for that user type can be applied to help the ser. Behaviour-based approaches use the user's behaviour as a model,com- monly using machine-learning techniques to discover useful patterns in the behaviour. Behavioural logging is employed to obtain the data necessary from which to extract patterns. Kobsa 9 provides a good survey of user modellin technique
780 S.E. Middleton et al. be of interest to a user. Ontologies are increasingly being used within the field of recommender systems, allowing knowledge-based techniques to supplement classical machine learning and statistical approaches. 1.1 Recommender Systems People find articulating exactly what they want difficult, but they are good at recognizing it when they see it. This insight has led to the utilization of relevance feedback, where people rate items as interesting or not interesting and the system tries to find items that match the “interesting”, positive examples and do not match the “not interesting”, negative examples. With sufficient positive and negative examples, modern machine learning techniques can classify new pages with impressive accuracy. Recommender systems can recommend many types of item, including web pages, new articles, music CDs and books. Unobtrusive monitoring provides positive examples of what the user is looking for, without interfering with the user’s normal work activity. Heuristics can also be applied to infer negative examples from observed behaviour, although generally with less confidence. This idea has led to content-based recommender systems, which unobtrusively watch user behaviour and recommend new items that correlate with a user’s profile. Another way to recommend items is based on the ratings provided by other people who have liked the item before. Collaborative recommender systems do this by asking people to rate items explicitly and then recommend new items that similar users have rated highly. An issue with collaborative filtering is that there is no direct reward for providing examples since they only help other people. This leads to initial difficulties in obtaining a sufficient number of ratings for the system to be useful, a problem known as the cold-start problem [15]. Hybrid systems, attempting to combine the advantages of content-based and collaborative recommender systems, have also proved popular to-date. The feedback required for content-based recommendation is shared, allowing collaborative recommendation as well. 1.2 User Profiling User profiling is typically either knowledge-based or behaviour-based. Knowledge-based approaches use static models of users and dynamically match users to the closest model. Questionnaires and interviews are often employed to obtain this user knowledge. Once a model is selected for a user, specialist domain knowledge for that user type can be applied to help the user. Behaviour-based approaches use the user’s behaviour as a model, commonly using machine-learning techniques to discover useful patterns in the behaviour. Behavioural logging is employed to obtain the data necessary from which to extract patterns. Kobsa [9] provides a good survey of user modelling techniques
Ontology-Based Recommender Systems 781 profiling approach used by most recommender systems is behavioural-based, commonly using a binary class model to represent what users find interesting and not interesting. Machine-learning techniques are then used to find potential items of interest in respect to the binary model recommending items that match the positive examples and do not match the negative examples. There are a lot of effective machine learning algorithms based on two classes. a binary profile does not, however, lend itself to sharing examples of interest or integrating any domain knowledge that might be available. Sebastiani [19 provides a good survey of current machine learning An ontology is a conceptualisation of a domain into a human-understandable, but machine-readable format consisting of entities, attributes, relationships and axioms [8. Ontologies can provide a rich conceptualisation of the working domain of an organisation, representing the main concepts and relationships of the work activities. These relationships could represent isolated information home phone number, or they could represent an activity such as authoring a document, or attending a conference. Part Ill contains examples of the types of ontology that are in use today, such as chapter COMM: A Core Ontology for Multimedia Annotation Ontologies help extend recommender systems to a multi-class environment allowing knowledge-based approaches to be used alongside classical machine learning algorithms. Section 2 provides an in-depth overview of how ontolo- gies are integrated into the techniques used for recommendation. Part IV of this book contains details on the current best practice for supporting infras- tructures and for ontologies, especially chapters"Ontology Repositories"and Ontology Mapping 1. 4 Chapter Structure In this chapter we show how ontologies are used in recommender systems to- day, providing an overview of the technology space and some further reading on specific approaches. We then examine in some depth a case study of two recommender systems that were among the first to adopt ontological tech- niques. In these case studies the problem domain, algorithms and results ar detailed along with a discussion that highlights some of the practical difficul ties experienced running a recommender system for real 2 Ontology Use in Recommender Systems Ontologies are now used routinely in recommender systems in combination with machine learning, statistical correlations, user profiling and domain spe- cific heuristics. Commercial recommender systems generally either maintain
Ontology-Based Recommender Systems 781 The user profiling approach used by most recommender systems is behavioural-based, commonly using a binary class model to represent what users find interesting and not interesting. Machine-learning techniques are then used to find potential items of interest in respect to the binary model, recommending items that match the positive examples and do not match the negative examples. There are a lot of effective machine learning algorithms based on two classes. A binary profile does not, however, lend itself to sharing examples of interest or integrating any domain knowledge that might be available. Sebastiani [19] provides a good survey of current machine learning techniques. 1.3 Ontologies An ontology is a conceptualisation of a domain into a human-understandable, but machine-readable format consisting of entities, attributes, relationships, and axioms [8]. Ontologies can provide a rich conceptualisation of the working domain of an organisation, representing the main concepts and relationships of the work activities. These relationships could represent isolated information such as an employee’s home phone number, or they could represent an activity such as authoring a document, or attending a conference. Part III contains examples of the types of ontology that are in use today, such as chapter “COMM: A Core Ontology for Multimedia Annotation”. Ontologies help extend recommender systems to a multi-class environment, allowing knowledge-based approaches to be used alongside classical machine learning algorithms. Section 2 provides an in-depth overview of how ontologies are integrated into the techniques used for recommendation. Part IV of this book contains details on the current best practice for supporting infrastructures and for ontologies, especially chapters “Ontology Repositories” and “Ontology Mapping”. 1.4 Chapter Structure In this chapter we show how ontologies are used in recommender systems today, providing an overview of the technology space and some further reading on specific approaches. We then examine in some depth a case study of two recommender systems that were among the first to adopt ontological techniques. In these case studies the problem domain, algorithms and results are detailed along with a discussion that highlights some of the practical difficulties experienced running a recommender system for real. 2 Ontology Use in Recommender Systems Ontologies are now used routinely in recommender systems in combination with machine learning, statistical correlations, user profiling and domain specific heuristics. Commercial recommender systems generally either maintain
782 S.E. Middleton et al simple product ontologies(e.g. books )that they can then utilize via heuristics or have a large community of users actively rating content(e. g. movies )suit- able for collaborative filtering. More research oriented recommender systems use a much wider variety of techniques that offer advantages such as improved accuracy coupled with constraints such as requiring explicit relevance feedback or intrusive monitoring of user behaviour over prolonged periods of time Recommendation of new items to users can be performed by looking at item to item similarity(content-based filtering), item reviews within a com munity of users(collaborative filtering), semantic relationships between items (heuristic-based recommendation) or a hybrid approach. In many cases the type of approach adopted will depend heavily on how much metadata is avail able about the items and how much user feedback is available, both implicit and explicit Content-based techniques work well if training data is available advance. Collaborative techniques work well when a system has a large community of users. There are, however, no definitive rules to decide on an approach and normally experience and expertise is required to pick the best a given problem doma 2.1 Content-Based Recommendation Early recommender systems used content-based binary classification ap- proaches looking at training sets of what was, and what was not interesting to a specific user. Machine learning techniques were employed to perfor supervised learning based on sets of observed training examples that a user labelled either as"good"or"bad. A classic example of a content-based rec- ommender system is Fab 1, which uses a binary class k-Nearest Neighbour classifier. Other binary class examples include personal assistant agents such as News Dude 2, using a naive Bayes classifier, and News Weeder [11,using a TF-IDF based classifier, which profile individual user interests and try to find items of interest To enhance binary classification domain ontologies were introduced alloy ing multi-class classification and hence multi-class recommendation. Typically the classes in a domain ontology, such as a product ontology defining all the products of an e-commerce website, would be used to classify the previously observed products/web pages a user had purchased / viewed. A good exam- le of multi-class recommendation is RAAP 4, which uses a simple set of categories to represent individual user profile Once a domain has been classified in terms of ontological concepts the relationships defined by the domain ontology can be used to infer interest and relevance of one concept from observed interest in another. A knowledge-based system can use expert system rules to infer probabilistic interest in classes of tem with a semantic connection to an observed item of interest. Typically the semantic distance(number of relationships away one topic is from another)is used to calculate semantic similarity, and this is used to weight likely interest Entre 3 is a restaurant recommender system that uses a knowledge
782 S.E. Middleton et al. simple product ontologies (e.g. books) that they can then utilize via heuristics or have a large community of users actively rating content (e.g. movies) suitable for collaborative filtering. More research oriented recommender systems use a much wider variety of techniques that offer advantages such as improved accuracy coupled with constraints such as requiring explicit relevance feedback or intrusive monitoring of user behaviour over prolonged periods of time. Recommendation of new items to users can be performed by looking at item to item similarity (content-based filtering), item reviews within a community of users (collaborative filtering), semantic relationships between items (heuristic-based recommendation) or a hybrid approach. In many cases the type of approach adopted will depend heavily on how much metadata is available about the items and how much user feedback is available, both implicit and explicit. Content-based techniques work well if training data is available in advance. Collaborative techniques work well when a system has a large community of users. There are, however, no definitive rules to decide on an approach and normally experience and expertise is required to pick the best approach for a given problem domain. 2.1 Content-Based Recommendation Early recommender systems used content-based binary classification approaches looking at training sets of what was, and what was not interesting to a specific user. Machine learning techniques were employed to perform supervised learning based on sets of observed training examples that a user labelled either as “good” or “bad”. A classic example of a content-based recommender system is Fab [1], which uses a binary class k-Nearest Neighbour classifier. Other binary class examples include personal assistant agents such as NewsDude [2], using a naive Bayes classifier, and NewsWeeder [11], using a TF-IDF based classifier, which profile individual user interests and try to find items of interest. To enhance binary classification domain ontologies were introduced allowing multi-class classification and hence multi-class recommendation. Typically the classes in a domain ontology, such as a product ontology defining all the products of an e-commerce website, would be used to classify the previously observed products / web pages a user had purchased / viewed. A good example of multi-class recommendation is RAAP [4], which uses a simple set of categories to represent individual user profiles. Once a domain has been classified in terms of ontological concepts the relationships defined by the domain ontology can be used to infer interest and relevance of one concept from observed interest in another. A knowledge-based system can use expert system rules to infer probabilistic interest in classes of item with a semantic connection to an observed item of interest. Typically the semantic distance (number of relationships away one topic is from another) is used to calculate semantic similarity, and this is used to weight likely interest. Entre [3] is a restaurant recommender system that uses a knowledge-base
Ontology-Based Recommender Systen and heuristic rules for recommendation. Where users articulate queries via a web interface the query criteria can drive a knowledge-based decision tree for dvanced query refinement. The CWAdvisor 5 system is an example of such an approach where a finite state model is used to refine queries for available financial service products that match the users stated requirements 2.2 Clustering and Topic Diversification Some domains do not have well identified classes of item from which content can be classified In these cases recommender systems have employed cluster ing techniques to identify within groups of items potentially similar classes. Hierarchical clustering has been used to categorize document collections for recommender algorithms [18 and sub-divides into either distance-based clus- tering or concept-based clustering Distance-based clustering [21 takes either a top-down(partitioning)or bottom-up(agglomerative) approach to building a hierarchical class tree. A distance function is defined to compute similarity between documents, often based on the similarity of frequency of the words within the document. The clustering algorithm iterates, either dividing super-clusters or merging small clusters into larger ones, until the final concept tree is formed. Concept-based clustering takes items represented as attribute-pairs and builds relationships based on the probability of occurrence of attribute-pairs within nodes. An early example of concept-b ased clustering is the COBWeb [6 algorithm. Nodes are created in a top-down approach where nodes are split or merged according to a category utility value; category utility is a measure of differentiation power of that node Often recommender systems will recommend clusters of items that are very similar, or variants of the same item(e. g. different formats of the film/DVD) To avoid this topic diversification 22 can be employed to ensure each rec- ommendation is on a well defined concept, hopefully increasing the useful- ness of a set of recommendations to the e user Algorithms to perform topic diversification will compute a dissimilarity ranking and merge this with the recommendation ranking. Semantic distance and super-class relationships can e used to compute dissimilarity between item sets 2.3 Collaborative Filtering Collaborative filtering works by using the ratings provided by a community of users to recommend items for a specific user. There are two complementary approaches available, user-based or item-based collaborative filtering. User- based collaborative filtering is where similar users are found and items recom mended that these similar users also liked. Item-based collaborative filtering is where items are grouped if people rate them similarly. In order to perform collaborative filtering a user profile must be created rom the available historical records of what items people have reviewed and
Ontology-Based Recommender Systems 783 and heuristic rules for recommendation. Where users articulate queries via a web interface the query criteria can drive a knowledge-based decision tree for advanced query refinement. The CWAdvisor [5] system is an example of such an approach where a finite state model is used to refine queries for available financial service products that match the user’s stated requirements. 2.2 Clustering and Topic Diversification Some domains do not have well identified classes of item from which content can be classified. In these cases recommender systems have employed clustering techniques to identify within groups of items potentially similar classes. Hierarchical clustering has been used to categorize document collections for recommender algorithms [18] and sub-divides into either distance-based clustering or concept-based clustering. Distance-based clustering [21] takes either a top-down (partitioning) or bottom-up (agglomerative) approach to building a hierarchical class tree. A distance function is defined to compute similarity between documents, often based on the similarity of frequency of the words within the document. The clustering algorithm iterates, either dividing super-clusters or merging small clusters into larger ones, until the final concept tree is formed. Concept-based clustering takes items represented as attribute-pairs and builds relationships based on the probability of occurrence of attribute-pairs within nodes. An early example of concept-based clustering is the COBWEB [6] algorithm. Nodes are created in a top-down approach where nodes are split or merged according to a category utility value; category utility is a measure of differentiation power of that node. Often recommender systems will recommend clusters of items that are very similar, or variants of the same item (e.g. different formats of the film/DVD). To avoid this topic diversification [22] can be employed to ensure each recommendation is on a well defined concept, hopefully increasing the usefulness of a set of recommendations to the user. Algorithms to perform topic diversification will compute a dissimilarity ranking and merge this with the recommendation ranking. Semantic distance and super-class relationships can be used to compute dissimilarity between item sets. 2.3 Collaborative Filtering Collaborative filtering works by using the ratings provided by a community of users to recommend items for a specific user. There are two complementary approaches available, user-based or item-based collaborative filtering. Userbased collaborative filtering is where similar users are found and items recommended that these similar users also liked. Item-based collaborative filtering is where items are grouped if people rate them similarly. In order to perform collaborative filtering a user profile must be created from the available historical records of what items people have reviewed and
784 S.E. Middleton et al rated. Often a 5-point scale is used for ratings (very good to very bad) A common user profile representation is a weighted vector if interest with as many dimensions as the domain has classes. Vectors can also be used for item to item similarity. Domains where item metadata is not accessi- ble as ontological terms will usually apply pre-processing techniques to com- pute word/document/metadata term frequencies, remove common words and merge similar words using a thesaurus like WordNet User-based collaborative filtering is the most popular recommendation al- gorithm due to its simplicity and excellent quality of recommendation. First neighbourhoods are formed using a similarity metric, such as a statistical correlation metric like Pearson-r correlation. Second a set of rating predic tions are created using profiles that are within the same neighbourhood as the users own profile. Recommendations are created from the top-M The GroupLens project [10 is an early exploiter of user-based collaborative filtering Item-based collaborative filtering has become popular in the last 5 years ince it decouples the model computation from the prediction process; Amazon have used this technique successfully. Just as in user-based similarity items are compared on the basis of how many users rank them similarly. The neighbourhoods computed are therefore collections of items that are similar. This technique scales well since new items will be added to neighbourhoods s users rate them without the need for explicit ontology maintenance. Sometimes a recommender system will have to compare items from differ ent domain ontologies, such as two product lists. In these cases an ontology can be created for both domains in a common language(such as OWL)and the mapping between them formulated, either manually or using a automated technique [12 such as a Bayesian belief network. Once concepts are success- fully mapped the normal approaches for recommendation can be applied 2.4 Use of the Semantic Web and Web 2.0 Approaches Recent work has also used some of the emerging Web 2.0 resources from the Semantic Web to help identify classes of item. One such system [20 has used an internet movie database that contains extensive information about actors movies, etc, and mapped this semantic information to user behaviour on a movie recommendation website. Tag clouds are created based on the keyword frequencies behind the items they have rated. Data mining techniques [ can also be coupled with ontological knowledge to improve similarity matching 3 Case Study: Two Ontological Recommender Systems For a case study two experimental recommender syste nted, Quic step and Foxtrot, that explored the novel idea of using an ontological approach
784 S.E. Middleton et al. rated. Often a 5-point scale is used for ratings (very good to very bad). A common user profile representation is a weighted vector if interest with as many dimensions as the domain has classes. Vectors can also be used for item to item similarity. Domains where item metadata is not accessible as ontological terms will usually apply pre-processing techniques to compute word/document/metadata term frequencies, remove common words and merge similar words using a thesaurus like WordNet. User-based collaborative filtering is the most popular recommendation algorithm due to its simplicity and excellent quality of recommendation. First neighbourhoods are formed using a similarity metric, such as a statistical correlation metric like Pearson-r correlation. Second a set of rating predictions are created using profiles that are within the same neighbourhood as the user’s own profile. Recommendations are created from the top-N items. The GroupLens project [10] is an early exploiter of user-based collaborative filtering. Item-based collaborative filtering has become popular in the last 5 years since it decouples the model computation from the prediction process; Amazon [13] have used this technique successfully. Just as in user-based similarity items are compared on the basis of how many users rank them similarly. The neighbourhoods computed are therefore collections of items that are similar. This technique scales well since new items will be added to neighbourhoods as users rate them without the need for explicit ontology maintenance. Sometimes a recommender system will have to compare items from different domain ontologies, such as two product lists. In these cases an ontology can be created for both domains in a common language (such as OWL) and the mapping between them formulated, either manually or using a automated technique [12] such as a Bayesian belief network. Once concepts are successfully mapped the normal approaches for recommendation can be applied. 2.4 Use of the Semantic Web and Web 2.0 Approaches Recent work has also used some of the emerging Web 2.0 resources from the Semantic Web to help identify classes of item. One such system [20] has used an internet movie database that contains extensive information about actors, movies, etc., and mapped this semantic information to user behaviour on a movie recommendation website. Tag clouds are created based on the keyword frequencies behind the items they have rated. Data mining techniques [4] can also be coupled with ontological knowledge to improve similarity matching and recommendation within historical usage data. 3 Case Study: Two Ontological Recommender Systems For a case study two experimental recommender systems are presented, Quickstep and Foxtrot, that explored the novel idea of using an ontological approach to user profiling in the context of recommender systems. Representing user
Ontology-Based Recommender Systems 785 interests in ontological terms involves losing some of the fine grained informa- ion held in the raw examples of interest, but in turn allows inference to assist user profiling, communication with other external ontologies and visualization of the profiles using ontological terms understandable to users. Figure l shows the general approach taken by both our recommender systems. Quickstep im- plements only the basic recommendation interface, while Foxtrot implements ll the shown features A research paper topic ontology is shared between all system processes, al- lowing both classifications and user profiles to use a common terminology. The ontology itself contains is-a relationships between appropriate topic classes; a ction from the topic ontology is shown in Fig. 2. The Quickstep ontology was Proler Database Fig. 1. Quickstep and Foxtrot recommender system data flow Artificial Intelligen face Agents Multi-Agent-Systems owledge Representahon atural Language Processing Rule Learning Vision Pattern Recogition Theory (machine leaming Human-Computer Interaction permed .. Data Structures Algorthms& Theory..Graphics& Virtual Reality Operating Syster tecture Fig. 2. Section from the Foxtrot research paper topic ontology
Ontology-Based Recommender Systems 785 interests in ontological terms involves losing some of the fine grained information held in the raw examples of interest, but in turn allows inference to assist user profiling, communication with other external ontologies and visualization of the profiles using ontological terms understandable to users. Figure 1 shows the general approach taken by both our recommender systems. Quickstep implements only the basic recommendation interface, while Foxtrot implements all the shown features. A research paper topic ontology is shared between all system processes, allowing both classifications and user profiles to use a common terminology. The ontology itself contains is-a relationships between appropriate topic classes; a section from the topic ontology is shown in Fig. 2. The Quickstep ontology was Fig. 1. Quickstep and Foxtrot recommender system data flow Fig. 2. Section from the Foxtrot research paper topic ontology
86 S.E. Middleton et al based on the open directory project's [7 computer science topic classificatio while the Foxtrot ontology was based on the CORA [14] digital library paper sification; manual enhancements were made to each ontology to better reflect some of the more specialist sub-topics researchers required. Reusing existing classifications saves time and provides a source for training examples especially with the COra digital library, which contained many pre-classified 3.1 Classification Using a Research Paper Topic Ontology much la taining examples, within the structure of an ontology, allows for much larger training sets than would be possible if a single user just pro- vided examples of personal interest. Larger training sets improve classifier accuracy. However, multi-class classification is inherently less accurate than binary class classification, so the increased training set size has to be weighed along with the reduction in accuracy that occurs with every extra class the system supports. Both the Quickstep and Foxtrot recommender systems use the research paper topic ontology to base paper classifications upon. A set of labelled example papers is manually provided for each class within the ontology, and then used by the classifier as a labelled training set. In the Quickstep system users can add new examples of papers as time goes by, allowing the training set to reflect the continually changing needs of the users In addition to larger training sets, having users share a common ontology enforces a consistent conceptual model, which removes some of the subjective nature of selecting categories for research papers. A common conceptual model also helps users to understand how the recommender system works, which helps form reasonable user expectations and assists in building trust and a feeling of control over what the system is do 3.2 Ontological Inference to Assist User Profiling Ontological inference is a powerful tool to assist user profiling. An ontology could contain all sorts of useful knowledge about users and their interests. such as related research subjects, technologies behind each subject area, projects people are working on, etc. This knowledge can be used to infer more interests than can be seen by just observation Our two experimental recommender systems both use ontological inference to enhance user profiles. Is-a relationships within the research paper topic ontology are used to infer interest in more general, super-class topics. We add 50% of the interest in a specific class to the super-class. This inference has the effect of rounding out profiles, making them more inclusive and attuning them to the broad interests of a user
786 S.E. Middleton et al. based on the open directory project’s [7] computer science topic classification, while the Foxtrot ontology was based on the CORA [14] digital library paper classification; manual enhancements were made to each ontology to better reflect some of the more specialist sub-topics researchers required. Reusing existing classifications saves time and provides a source for training examples, especially with the CORA digital library, which contained many pre-classified research papers. 3.1 Classification Using a Research Paper Topic Ontology Sharing training examples, within the structure of an ontology, allows for much larger training sets than would be possible if a single user just provided examples of personal interest. Larger training sets improve classifier accuracy. However, multi-class classification is inherently less accurate than binary class classification, so the increased training set size has to be weighed along with the reduction in accuracy that occurs with every extra class the system supports. Both the Quickstep and Foxtrot recommender systems use the research paper topic ontology to base paper classifications upon. A set of labelled example papers is manually provided for each class within the ontology, and then used by the classifier as a labelled training set. In the Quickstep system users can add new examples of papers as time goes by, allowing the training set to reflect the continually changing needs of the users. In addition to larger training sets, having users share a common ontology enforces a consistent conceptual model, which removes some of the subjective nature of selecting categories for research papers. A common conceptual model also helps users to understand how the recommender system works, which helps form reasonable user expectations and assists in building trust and a feeling of control over what the system is doing. 3.2 Ontological Inference to Assist User Profiling Ontological inference is a powerful tool to assist user profiling. An ontology could contain all sorts of useful knowledge about users and their interests, such as related research subjects, technologies behind each subject area, projects people are working on, etc. This knowledge can be used to infer more interests than can be seen by just observation. Our two experimental recommender systems both use ontological inference to enhance user profiles. Is-a relationships within the research paper topic ontology are used to infer interest in more general, super-class topics. We add 50% of the interest in a specific class to the super-class. This inference has the effect of rounding out profiles, making them more inclusive and attuning them to the broad interests of a user
Ontology-Based Recommender Systems 787 Interest value(n)/ days old(n) Paper browsed interest values Recommendation followed =2 50% of sub-class per instanc Fig. 3. Profiling algorithm The profiling algorithm used is shown in Fig 3. A time-decay function is applied to the observed behaviour events to form the basic profile Inference is then used to enhance the interest profile, with the 50% inference rule applied to all ontological is-a relationships, up to the root class, for each observe The event interest values were chosen to balance the feed back in favour of explicitly provided feedback, which is likely to be the most reliable. The 50% inference value was chosen to reflect the reduction in certainty you get the further away from the observed behaviour you move. Determining optimal values for these parameters would require further empirical evaluation 3.3 Bootstrapping with an External Ontology of initial behavioural information significantly reduces the accuracy of, CK Recommender systems suffer from the cold-start problem [ 15], where the lad profiles, and hence recommendations. This poor performance can deter users from adopting the system, which of course prevents the system from acquiring more behaviour data; it is possible that a recommender system will never be used enough to overcome its cold-start In one of our experiments we take an external ontology containing publi cation and personnel data about academic researchers and integrate it with the Quickstep recommender system. The knowledge held within the external ontology is used to bootstrap initial user profiles, with the aim of reducing the cold-start effect. The external ontology uses the same research topic ontology as the Quickstep system, providing a firm basis for communication. The exter- nal ontology contains publications and authorship relationships, projects and project membership, staff and their roles and other such knowledge. Knowl edge of publications held within the external ontology is used to infer histori- cal interests for new users, and network analysis of ontological relationships is used to discover similar users whose own interests might be used to bootstrap a new user's profile The two bootstrapping algorithms used in our experiment are shown in Figs 4 and 5. The new-system initial profile algorithm takes all the
Ontology-Based Recommender Systems 787 Fig. 3. Profiling algorithm The profiling algorithm used is shown in Fig. 3. A time-decay function is applied to the observed behaviour events to form the basic profile. Inference is then used to enhance the interest profile, with the 50% inference rule applied to all ontological is-a relationships, up to the root class, for each observed event. The event interest values were chosen to balance the feedback in favour of explicitly provided feedback, which is likely to be the most reliable. The 50% inference value was chosen to reflect the reduction in certainty you get the further away from the observed behaviour you move. Determining optimal values for these parameters would require further empirical evaluation. 3.3 Bootstrapping with an External Ontology Recommender systems suffer from the cold-start problem [15], where the lack of initial behavioural information significantly reduces the accuracy of user profiles, and hence recommendations. This poor performance can deter users from adopting the system, which of course prevents the system from acquiring more behaviour data; it is possible that a recommender system will never be used enough to overcome its cold-start. In one of our experiments we take an external ontology containing publication and personnel data about academic researchers and integrate it with the Quickstep recommender system. The knowledge held within the external ontology is used to bootstrap initial user profiles, with the aim of reducing the cold-start effect. The external ontology uses the same research topic ontology as the Quickstep system, providing a firm basis for communication. The external ontology contains publications and authorship relationships, projects and project membership, staff and their roles and other such knowledge. Knowledge of publications held within the external ontology is used to infer historical interests for new users, and network analysis of ontological relationships is used to discover similar users whose own interests might be used to bootstrap a new user’s profile. The two bootstrapping algorithms used in our experiment are shown in Figs. 4 and 5. The new-system initial profile algorithm takes all the
788 S.E. Middleton et al publications of a user and creates a profile of historical interests. The as sumption is that a user's previous publications indicate that user's interests The new-user initial profile algorithm takes a set of similar users, obtained via network analysis of the external ontologies project membership and inter-staff relationships, and includes these users interests into the bootstrap profile. His- torical publication interests from the new user are also added as before. The parametric values shown in Figs 4 and 5 were empirically determined after several experimental runs using test data. In addition to using the ontology to bootstrap the recommender system our experiment uses the interest profiles held within the red to continually update the external ontology Interest acquisition is a problem- atic task for ontologies that are based on static knowledge sources, and this synergistic relationship provides a useful source of personal knowledge about individual researchers topic interest(t) ∑ 1/ publication age(n) new-system initial profile(t, topic interest(t))* Interest value for 50% of sub-class super-class per topic 7∑ profile interest (u,t) ∑ rofile interest(u, t)-interest of user u in topic t"confidence est value for 50 of sub-cla Fig. 5. New-user initial profile algorithm
788 S.E. Middleton et al. publications of a user and creates a profile of historical interests. The assumption is that a user’s previous publications indicate that user’s interests. The new-user initial profile algorithm takes a set of similar users, obtained via network analysis of the external ontologies project membership and inter-staff relationships, and includes these users interests into the bootstrap profile. Historical publication interests from the new user are also added as before. The parametric values shown in Figs. 4 and 5 were empirically determined after several experimental runs using test data. In addition to using the ontology to bootstrap the recommender system, our experiment uses the interest profiles held within the recommender system to continually update the external ontology. Interest acquisition is a problematic task for ontologies that are based on static knowledge sources, and this synergistic relationship provides a useful source of personal knowledge about individual researchers. Fig. 4. New-system initial profile algorithm Fig. 5. New-user initial profile algorithm