2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology Web Intelligence in Tourism User Modeling and Recommender System runo co o Martins GECAD-Knowledge Engineering GECAD-Knowledge Engineering GECAD-Knowledge Engineering and Decision Support Group and Decision Support Group and Decision Support Group Porto, Portugal Porto, Portugal brunocoelho(@dei isep. pp pt const(@dei isep. ipp.pt ana(@dei isep. tpp. pt Abstract-This paper presents a successful attempt at evolving web brief approach regarding UM in tourism, while section 3 refers ss, which is divided in three stages: the user areas: User modeling and Recommender systems. The first model itself, the set of reasoning components and the user a wide variety of techniques, such as stereotypes, keywords and customization stage. In section 4, the developed tourism psychological models. These techniques, besides presenting user application will be presented regarding its advantages and interests with great coherence and completeness, allow for the testing results. Finally, in section 5 some conclusions will be reduction of several current problems such as the cold start issue, outlined, along with future work perspectives gray sheep individuals and overspecialization. The recommender system, by making use of all user models'building IL. USER MODELING IN TOURISM n interesting, innovative and hybrid nature to as behavioral filtering, multi-techniq Despite recent evolutions in UM, current systems still do resourcefulness and on-the-fly suggestions. The architecture was not explore the huge potential of this area, namely in the already tested in the scope of a prototype regarding the city of tourism domain [18J[19]. User information is extensively Porto, in Portug requested at startup, without necessarily being used throughout the remaining life-cycle of the application. Most of Keywords-user modeling: user-adapted representations of the assumptions they make about users, which ends up in incoherent recommendations being . INTRODUCTION performed B3[15]. [61, [9], [18 and [19] present some he tourism domain, and in particular, the holiday choosing interesting systems which use UM techniques and / or are process, represent very complex decision-making matters. On pertinent to the domain at hand. Since the data available is not one hand, the user is faced with the obviously endless group of very extent, it was not chosen to perform a formal comparison existing options, and on the other, the heterogeneity of the between the different systems. Still, informal considerations places to visit. Furthermore, he still has to consider the can be extrapolated. Some of the referred systems [6]follow a inherent specificities of the holiday chosen destiny, like, for knowledge-approach, by making use of several kinds of example, the type of Points of Interest(POI)available, hosted knowledge management techniques(suppositions and beliefs, events and so on. Another reason that explains why this for instance), which, although dealing with certainty in process is so difficult is the fact that, besides user's own inferred data, are much more computational intensive. These interests and preferences, many times they also have to take techniques are too strict when we consider the final natural into consideration those of other people as well [5] task of UM application: the rs. The rs needs an extremel iven those premises, tourism is an area clearly electable well balanced relation between fast and reliable data, which is for the use of Artificial Intelligence(Al) and its benefits, and not achieved when using those complex knowledge-based in particular, Decision Support Systems(DSS)[1]. Such techniques. The use of stereotypes was positively detected systems require the use of a coherent model of the user in these applications [6], and will be an integral part of the work order for results, and the overall system, to be customized and to be described In a very broad statement, the current main targeted to him / her: thats where User Modeling(UM) flaw regarding tourism systems is the poor UM backing them techniques come into play [2]. After an assessment of the up. Most systems rest their efficiency on a single Ul current tourism platforms, as well as general systems which technique; even if such modeling is not incorrect,it Ise UM, there is the belief that significant work can still be certainly not enough, considering the complexity of the human done, regarding a more complex modeling of users and being in various aspects, such as behavioral. Another approach uch more useful and effective use of those models surfacing in the latest years is the overrated preference for 978-0-7695-4191-4/052600◎2010IEEE DOII0.I109/I-AT2010236
Web Intelligence in Tourism User Modeling and Recommender System Bruno Coelho GECAD - Knowledge Engineering and Decision Support Group Porto, Portugal brunocoelho@dei.isep.ipp.pt Constantino Martins GECAD - Knowledge Engineering and Decision Support Group Porto, Portugal const@dei.isep.ipp.pt Ana Almeida GECAD - Knowledge Engineering and Decision Support Group Porto, Portugal ana@dei.isep.tpp.pt Abstract-This paper presents a successful attempt at evolving web intelligence in the tourism scenario, namely throughout two main areas: User Modeling and Recommender Systems. The first subject deals with the correct modeling of tourists’ profiles using a wide variety of techniques, such as stereotypes, keywords and psychological models. These techniques, besides presenting user interests with great coherence and completeness, allow for the reduction of several current problems such as the cold start issue, gray sheep individuals and overspecialization. The recommender system, by making use of all user models’ building blocks, brings an interesting, innovative and hybrid nature to the area, with benefits such as behavioral filtering, multi-technique resourcefulness and on-the-fly suggestions. The architecture was already tested in the scope of a prototype regarding the city of Porto, in Portugal. Keywords-user modeling; user-adapted web systems; recommender systems; stereotypes; tourism I. INTRODUCTION The tourism domain, and in particular, the holiday choosing process, represent very complex decision-making matters. On one hand, the user is faced with the obviously endless group of existing options, and on the other, the heterogeneity of the places to visit. Furthermore, he still has to consider the inherent specificities of the holiday chosen destiny, like, for example, the type of Points of Interest (POI) available, hosted events and so on. Another reason that explains why this process is so difficult is the fact that, besides user’s own interests and preferences, many times they also have to take into consideration those of other people as well [5]. Given those premises, tourism is an area clearly electable for the use of Artificial Intelligence (AI) and its benefits, and in particular, Decision Support Systems (DSS) [1]. Such systems require the use of a coherent model of the user in order for results, and the overall system, to be customized and targeted to him / her: that’s where User Modeling (UM) techniques come into play [2]. After an assessment of the current tourism platforms, as well as general systems which use UM, there is the belief that significant work can still be done, regarding a more complex modeling of users and a much more useful and effective use of those models. This paper is organized as follows: section 2 will present a brief approach regarding UM in tourism, while section 3 refers to the UM process, which is divided in three stages: the user model itself, the set of reasoning components and the user customization stage. In section 4, the developed tourism application will be presented regarding its advantages and testing results. Finally, in section 5, some conclusions will be outlined, along with future work perspectives. II. USER MODELING IN TOURISM Despite recent evolutions in UM, current systems still do not explore the huge potential of this area, namely in the tourism domain [18][19]. User information is extensively requested at startup, without necessarily being used throughout the remaining life-cycle of the application. Most of the times, though, systems rely on single and / or poor representations of the assumptions they make about users, which ends up in incoherent recommendations being performed [3][15]. [6], [9], [18] and [19] present some interesting systems which use UM techniques and / or are pertinent to the domain at hand. Since the data available is not very extent, it was not chosen to perform a formal comparison between the different systems. Still, informal considerations can be extrapolated. Some of the referred systems [6] follow a knowledge-approach, by making use of several kinds of knowledge management techniques (suppositions and beliefs, for instance), which, although dealing with more certainty in inferred data, are much more computational intensive. These techniques are too strict when we consider the final natural task of UM application: the RS. The RS needs an extremely well balanced relation between fast and reliable data, which is not achieved when using those complex knowledge-based techniques. The use of stereotypes was positively detected in these applications [6], and will be an integral part of the work to be described. In a very broad statement, the current main flaw regarding tourism systems is the poor UM backing them up. Most systems rest their efficiency on a single UM technique; even if such modeling is not incorrect, it is certainly not enough, considering the complexity of the human being in various aspects, such as behavioral. Another approach surfacing in the latest years is the overrated preference for 2010 IEEE/WIC/ACM International Conference on Web Intelligence and Intelligent Agent Technology 978-0-7695-4191-4/10 $26.00 © 2010 IEEE DOI 10.1109/WI-IAT.2010.236 619
social filtering methods [18]. While the use of this kind of Personal: this component holds user's personal data such as recommendations is not uninteresting(we also embrace them name and e-mail: n this very work), it is clear that the user itself is still the most Demographics: this information will be used by knowle important and primary source of recommending material, one discovery techniques, such as Data Mining whose deep analysis has not yet been performed in such a algorithms, in order to find usage patterns which may be useful for the system as well. Demographic data may also be used as part of the rs; is with some surprise that systems do not rely on information Handicaps: this group of user physical information is related to the psychological behavioral side of users, a key responsible for avoiding suggestions that do not fit a certain component of decision-making processes user handicap Given this action space for improvement, the proposed Trips: this module contains work tries to evolve UM techniques and the way systems can trip in which the user might be involved in, as well as the benefit from different kinds of user information, by using a containing tours. This module also contains past trips complex UM framework. IIL. DEvELOPment B. User Reasoning /Inference Mechanisms As it was earlier implied, a powerful UM process is he project to be described here devises a UM methodology endorsed by several components that maximize the hich dictates how the system is designed and operates: accuracy User Modeling(UM)-it's the means by which a system of the necessary user information in a variety of ways. Since with the final purpose of improving and customizing user methodology must therefore be a collaborative effort of begins with a suitable representation of the user (or user of part of user data. Furthermore, it is also believed that model, which can be the sum of a wide variety of different knowing a certain user information space by using more than techniques); then, that information is used to infer valuable one method simultaneously successfully increases confidence knowledge that can either be added to the user model or used in existent assumptions and divides responsibility amongst by the last phase of the process, the system adaptation. Fig. 1 various techniques, which ultimately results in a more backed- shows such as cycle up system with more solutions [344][5] User Model Figure 1. User Modeling Process The rest of this section will follow this approach, " UM as a Process", namely with the specifics of both the first and second stage of the platform. A. User model The User Model is the root of the UM process and pertains Figure 3. Knowledge Discovery Mechanisms Architecture to the broader user architecture that it features. In a certain point in time, the user model photogram is the user image as Fig. 3 allows the easy inference that tourism applications perceived by the system. The created model got part of its upper-level core functions, such as the RS, gather information that data into a coherent the information hierarchy that's comprehended in the proposed user profile, in order to generate new information User Model Community Models(CM). CM are an adaptation of etworks(NN), specifically simple form of previously fed into its mechanisms [8]. The first CM is about system navigation(user sessions, clickstream analysis, etc. ) Figure 2. User Model Architecture commitment, such as the generating a route. Plus, CM still display the following advantages: (1)act as a clustering
social filtering methods [18]. While the use of this kind of recommendations is not uninteresting (we also embrace them in this very work), it is clear that the user itself is still the most important and primary source of recommending material, one whose deep analysis has not yet been performed in such a complete manner as will be presented. To finish, and in such a domain where both personal and social interests are at stake, it is with some surprise that systems do not rely on information related to the psychological / behavioral side of users, a key component of decision-making processes. Given this action space for improvement, the proposed work tries to evolve UM techniques and the way systems can benefit from different kinds of user information, by using a complex UM framework. III. DEVELOPMENT The project to be described here devises a UM methodology which dictates how the system is designed and operates: User Modeling (UM) - it’s the means by which a system keeps user data and uses that information in a variety of ways with the final purpose of improving and customizing user experience within that system. It pertains to a process that begins with a suitable representation of the user (or user model, which can be the sum of a wide variety of different techniques); then, that information is used to infer valuable knowledge that can either be added to the user model or used by the last phase of the process, the system adaptation. Fig. 1 shows such as cycle. Figure 1. User Modeling Process The rest of this section will follow this approach, “UM as a Process”, namely with the specifics of both the first and second stage of the platform. A. User Model The User Model is the root of the UM process and pertains to the broader user architecture that it features. In a certain point in time, the user model photogram is the user image as perceived by the system. The created model got part of its influence from Benyon’s UM architecture [15]. Fig. 2 presents the information hierarchy that’s comprehended in the proposed user model. Figure 2. User Model Architecture • Personal: this component holds user’s personal data such as name and e-mail; • Demographics: this information will be used by knowledge discovery techniques, such as Data Mining (DM) algorithms, in order to find usage patterns which may be useful for the system as well. Demographic data may also be used as part of the RS; • Handicaps: this group of user physical information is responsible for avoiding suggestions that do not fit a certain user handicap. • Trips: this module contains information about the current trip in which the user might be involved in, as well as the containing tours. This module also contains past trips. B. User Reasoning / Inference Mechanisms As it was earlier implied, a powerful UM process is endorsed by several components that maximize the accuracy of the necessary user information in a variety of ways. Since user data can’t all be retrieved in the same manner, our UM methodology must therefore be a collaborative effort of several sub-systems, each of them responsible for the retrieval of part of user data. Furthermore, it is also believed that knowing a certain user information space by using more than one method simultaneously successfully increases confidence in existent assumptions and divides responsibility amongst various techniques, which ultimately results in a more backedup system with more solutions [3][4][5]. ! " Figure 3. Knowledge Discovery Mechanisms’ Architecture Fig. 3 allows the easy inference that tourism applications’ upper-level core functions, such as the RS, gather information throughout all sub-systems and merge that data into a coherent user profile, in order to generate new information. 1) Community Models Our UM architecture makes use of a group of two Community Models (CM). CM are an adaptation of Neural Networks (NN), specifically a rather simple form of those. CM generate two-dimensional representations of the data previously fed into its mechanisms [8]. The first CM is about system navigation (user sessions, clickstream analysis, etc.), while the second is concerned about effective POI commitment, such as the generating a route. Plus, CM still display the following advantages: (1) act as a clustering 620
r not (2) discover pol associations by anal图 to features related with that POl, and (3)keywords that to special words found within the name and description contained within selected groups of items and (3)search for item, by using text mining techniqu abnormalities between viewed versus used item 2) Likelihood Matrix 6) User Explicit Knowledge Retrieval The likelihood matrix is responsible for linking the user se to the with each one of the categories created in the Pol taxonomy, explicitly provide the system, as well as how should that are concerned. It ranges from-I to 1, where -l means total unlikelihood and 1 represents complete interest. This to the creation of creative ways to ask the user for information. Some approaches include: (I)ask only a certain number of nechanism is the basis for the stereotype module and thus items, and infer the other; (2)ask some information only it is both components work together in order to provide an over- technically needed and (3)shrink user data into shorter confident representation of user interests. Although one versions and then infer the complete data space. Table I shows technique is based on the other, their abstraction level is how, in the devised system, this problem is tackled: different, therefore triggering different results by both sed approach coherently represents TABLE L. USER MODEL COMPONENT ACQUISITION TECHNIQUES both user likes and dislikes, by maintaining a negative and positive action space in which assumptions can diverge within Acquisition technique I Personal, Demographics, Handicaps For Psychology Test,Form Our stereotype system can be easily explained though a set Interests, Current Trip. Past Trips Inferred, Form of development guidelines which originated it. First of all, the Image Association,Form POI taxonomy was re-conceptualized into hierarchical terms er serve a ne basis for the stereotype C. User Customization -Recommender System construction;then, an initial set of stereotypes was created. which will form the comparison basis for that stereotype to be most important step regarding e of the UM process is the each of them being fully described using the previous terms The user customization stage In this inked to a user. Finally, mechanisms were created to stage, the system modifies itself in order to contemplate user's compensate for an eventual insufficiency that might describe objectives within the respective system, therefore acting as the the initial set of stereotypes, as well as the suitability of their ultimate feature in this kind of applications. Up until the terms, namely: (1)propose underused stereotypes for removal (2)propose underused stereotype conditions for removal;(3) component only, which is the RS. Fig. 4 presents al propose overused conditions not included reotypes and techniques contemplated in the devised rS (4)propose new stereotypes based on user profiles(eliminates the grey sheep individual issue) t 4) Psychological Model The user psychological model will be in constant evolution, as the user interacts with the system an it traces of his personality evolution. The features selected to be part of this eu5 odule were based on psychological models devised along the ears, such as [7]. They range from 0 to 1, representing the two extremes of that feature. The user psychological profile Figure 4. Recommender System Techniques evolves by comparing used POI with the cur feeding and evolving the user behavioral model towards the Table I relates the presented techniques with the classical comparison o object(the POI class), adjusting it and changing literature approaches. The system performs a commitment into the input of all components depending on it, such as the RS balancing(1) the acceptance of traditional techniques and(2) introducing innovations into each one of them and also 5) Keywords proposing a new one [12][13[16[17] Tags are extremely user-dependent [10]. If no suitable feedback is shown by users in evolving their use and TABLE IL. COMPARISON BETWEEN LITERATURE AND PROPOSED SYSTEM usefulness, the fact is that tags become less important. With that said, several approaches for initially inserted tags within Literature Technique the items were devised. By automatically gifting items with tags, one of the few downsides of this kind of social media, the cold-start problem, can be diminished. The following types Keywords Content-based filtering to the poi class which classifies the POl; (2)keywords that Stereotypes and Psychologi Psychological Filtering
technique, naturally grouping related POI, physically related or not; (2) discover POI associations, by analyzing patterns contained within selected groups of items and (3) search for abnormalities between viewed versus used items. 2) Likelihood Matrix The likelihood matrix is responsible for linking the user with each one of the categories created in the POI taxonomy, being classified as a linear model in what literature techniques are concerned. It ranges from -1 to 1, where -1 means total unlikelihood and 1 represents complete interest. This mechanism is the basis for the stereotype module and thus both components work together in order to provide an overconfident representation of user interests. Although one technique is based on the other, their abstraction level is different, therefore triggering different results by both components. The proposed approach coherently represents both user likes and dislikes, by maintaining a negative and positive action space in which assumptions can diverge within. 3) Stereotypes Our stereotype system can be easily explained though a set of development guidelines which originated it. First of all, the POI taxonomy was re-conceptualized into hierarchical terms which would better serve as the basis for the stereotype construction; then, an initial set of stereotypes was created, each of them being fully described using the previous terms which will form the comparison basis for that stereotype to be linked to a user. Finally, mechanisms were created to compensate for an eventual insufficiency that might describe the initial set of stereotypes, as well as the suitability of their terms, namely: (1) propose underused stereotypes for removal; (2) propose underused stereotype conditions for removal; (3) propose overused conditions not included in stereotypes and (4) propose new stereotypes based on user profiles (eliminates the grey sheep individual issue). 4) Psychological Model The user psychological model will be in constant evolution, as the user interacts with the system and gives it traces of his personality evolution. The features selected to be part of this module were based on psychological models devised along the years, such as [7]. They range from 0 to 1, representing the two extremes of that feature. The user psychological profile evolves by comparing used POI with the current profile, feeding and evolving the user behavioral model towards the comparison o object (the POI class), adjusting it and changing the input of all components depending on it, such as the RS. 5) Keywords Tags are extremely user-dependent [10]. If no suitable feedback is shown by users in evolving their use and usefulness, the fact is that tags become less important. With that said, several approaches for initially inserted tags within the items were devised. By automatically gifting items with tags, one of the few downsides of this kind of social media, the cold-start problem, can be diminished. The following types of tags are automatically set for POI: (1) keywords that relate to the POI class which classifies the POI; (2) keywords that pertain to features related with that POI, and (3) keywords that pertain to special words found within the name and description of the item, by using text mining techniques. 6) User Explicit Knowledge Retrieval This is the UM component most close to the system interface, which deals with what information should the user explicitly provide the system, as well as how should that information retrieval be processed [11]. This issue gives birth to the creation of creative ways to ask the user for information. Some approaches include: (1) ask only a certain number of items, and infer the other; (2) ask some information only it is technically needed and (3) shrink user data into shorter versions and then infer the complete data space. Table 1 shows how, in the devised system, this problem is tackled: TABLE I. USER MODEL COMPONENT ACQUISITION TECHNIQUES User Model Component Acquisition Technique Personal, Demographics, Handicaps Form Psychologics Psychology Test, Form Interests, Current Trip, Past Trips Inferred, Form Stereotypes Image Association, Form C. User Customization - Recommender System The user customization stage of the UM process is the most important step regarding the user point of view. In this stage, the system modifies itself in order to contemplate user’s objectives within the respective system, therefore acting as the ultimate feature in this kind of applications. Up until the current state of the project, this stage encompasses one component only, which is the RS. Fig. 4 presents all techniques contemplated in the devised RS. Figure 4. Recommender System Techniques Table 1 relates the presented techniques with the classical literature approaches. The system performs a commitment into balancing (1) the acceptance of traditional techniques and (2) introducing innovations into each one of them and also proposing a new one [12][13][16][17]. TABLE II. COMPARISON BETWEEN LITERATURE AND PROPOSED SYSTEM System Technique Literature Technique Likelihood Matrix Knowledge-based Filtering Keywords Content-based Filtering Socialization Collaborative Filtering Stereotypes and Psychological Model Psychological Filtering 621
IV. EVALUATION However, a prototype has already been developed in order to In this section, some informal evaluation guidelines will be evaluate the proposed platform, showing already very presented in order to demonstrate the capabilities of the successful results. Some new developments include the tartup quality of response concerning filtering features and and Web 2.0 technol sion or% ir devised system, as well as a survey that targeted the so-far addition of morphological analysis in the text mining sers of the application. algorithm, effective use of demographic Startup quality: the devised prototype delivers an increased evolution and the inclus more averse personalization mechanisms. By making use of a clever and REFERENCE abstracting model for initially asked information(see 3.2.6), the application can automatically fill around 50% of the user [] coelho B. E.: Graduation Project Report-webmeeting.coMputer model information within the initial form. This kind of startup [2] Gay G.& Hembrooke, H: Activity-Centered Design: an Ecological quality is not performed by other systems [18] pproach to Designing Smart Tools and Usable Systems. MIT Press Transparency: besides profiting from an automatic UM platform which does everything in an automatic fashion, the [3] Fink J.& Kobsa A: User Modeling for Personalized City Tours.In user can also be invited into viewing, in a transparent manner, Intelligence Review. Volume 18 Number 1. Kluwer Academic everything that the system believes about him []. By making (41 (2002) ly int Generic User Modeling Systems. In: User Modeling and Use raction. Volume 11. Issue 1-2. Kluwer Academic such critical information and enhance RS results immediately very critical state regarding innovation. The UM platform here ( sT Recommender System: the RS state of the art has reached a User Modeling: Recent Work, Prospects and Hazards. In: User Interfaces: Principles and Practice. North Holland presented forms a very diverse basis for RS computation and 993) introduces a new way of filtering complex-domain items: [6 Martins A. C, Faria L, Carvalho C.V.&Carrapatoso E:User behavior-based. This technique merging causes RS's results to Educational Technology Society. Volume 11, 194-207(2008) output items with diverse sources, increasing likelihood for (7 Jung, C G. Psychological Types. Princeton University Press(1971) item commitment. In the process, problems that certain [8 Jennings A.& Higuchi H: A Personal News Service Based on a User techniques might have, such as under-confidence or the cold model Neural Network. In: User Modeling and User-Adapted start problem, are overcome. Interaction. Volume 3, Issue 1. Springer Netherlands(1992 Survey: the following conclusions were taken:(1)users [9] Rich E: User Modeling via Stereotypes. In: Cognitive Science: A were pleased by the short initial form;(2) users enjoyed the 1979) transparent spirit of the system and show interest in using the User Area and (3)users [10] Mathes A: Folksonomies have found the rs's results Computer Mediated satisfactory. Communication- LIS590CMC(2004 [11] Fleming M.& Choen R: Reasoning about Interaction in Mixed- V. CONCLUSIONS Initiative Al Systems. In: IJCAI-03- Workshop on Mixed-Initiative Intelligent Systems(2003) Machine learning reasoning hasn't been fully explored the tourism domain, particularly in what the tourist model is [12] Berka T. M: Designing Recommender Systems fo Tourism. In: ENTER 2004(2003) concerned [20]. User information has been misused and [13] Felfernig A, Gordea S, Jannach D, Teppan E& Zanker M: A Short falsely utilized. On one hand, current systems ask more formation than they really use, which is a downside both OGAI Journal. Volume 25, Issue 2(2006) usability-wise and resource-wise [19]. On the other hand, UM [14] Cramer H, Evers V, Ramlal S, Someren M, Rutledge L, Stash N. supported manner, based on very weak assumptions [18]. In Content-Based Art Recommender. In: User Modelin and User-Adapted Interaction. Volume 18, Issue 5, 455-496(2008) the proposed system, a modeling platform that requires only [15] Coelho B, Martins C.& Almeida A: ATOMICO- Arquitectura para minimal amount of user effort and still manages to create a Turistas com vista a Organizacao, Modelacao e very capable user image is deployed current projects, dictate the innovative nature of this prdi e Next is a list of some advantages that, when compared wit [16] Pazzani M. J& Billsus D: Content-Based Recommendation Systems. (1) innovative UM; (2)on-the-fly user profile update; (3)hot- Heidelberg(2007) start results quality;(4)behavioral-filtering introduction and [17] Porter J: Watch and Learn: How Recommendation Systems ar (5)multi-technique and heterogeneous RS From our point of Redefining the Web(2006 viewtheusermodelpresentinthispaperalongwiththemore[18]TripadvIsor,http://www.tripadvisor.com/(2010) omplexknowledgeinferencemechanismsthatconstitutethe[9]Wayn,http://www.wayn.com(2010) system, are two excellent basis for any tourism-focused [20] Zukerman I. Albrecht D. w: Predictive Statistical Models for User system. In fact, many of the second-stage co Modeling In: User Modeling and User-Adapted Interaction volume 1l actually be applied to completely different scenarios and Numbers 1-2, Pages 5-18(2000) domains, given some content minor content adjustments
IV. EVALUATION In this section, some informal evaluation guidelines will be presented in order to demonstrate the capabilities of the devised system, as well as a survey that targeted the so-far users of the application. Startup quality: the devised prototype delivers an increased startup quality of response concerning filtering features and personalization mechanisms. By making use of a clever and abstracting model for initially asked information (see 3.2.6), the application can automatically fill around 50% of the user model information within the initial form. This kind of startup quality is not performed by other systems [18]. Transparency: besides profiting from an automatic UM platform which does everything in an automatic fashion, the user can also be invited into viewing, in a transparent manner, everything that the system believes about him [14]. By making use of a friendly interface, users can increase confidence of such critical information and enhance RS results immediately. Recommender System: the RS state of the art has reached a very critical state regarding innovation. The UM platform here presented forms a very diverse basis for RS computation and introduces a new way of filtering complex-domain items: behavior-based. This technique merging causes RS’s results to output items with diverse sources, increasing likelihood for item commitment. In the process, problems that certain techniques might have, such as under-confidence or the cold start problem, are overcome. Survey: the following conclusions were taken: (1) users were pleased by the short initial form; (2) users enjoyed the transparent spirit of the system and show interest in using the User Area and (3) users have found the RS’s results satisfactory. V. CONCLUSIONS Machine learning reasoning hasn’t been fully explored in the tourism domain, particularly in what the tourist model is concerned [20]. User information has been misused and falsely utilized. On one hand, current systems ask more information than they really use, which is a downside both usability-wise and resource-wise [19]. On the other hand, UM and user adaptive-related mechanisms are used in a lowsupported manner, based on very weak assumptions [18]. In the proposed system, a modeling platform that requires only a minimal amount of user effort and still manages to create a very capable user image is deployed. Next is a list of some advantages that, when compared with current projects, dictate the innovative nature of this project: (1) innovative UM; (2) on-the-fly user profile update; (3) hotstart results quality; (4) behavioral-filtering introduction and (5) multi-technique and heterogeneous RS. From our point of view, the user model present in this paper, along with the more complex knowledge inference mechanisms that constitute the system, are two excellent basis for any tourism-focused system. In fact, many of the second-stage components can actually be applied to completely different scenarios and domains, given some content minor content adjustments. However, a prototype has already been developed in order to evaluate the proposed platform, showing already very successful results. Some new developments include the addition of morphological analysis in the text mining algorithm, effective use of demographic data, ontology evolution and the inclusion of more diverse forms of social and Web 2.0 technology. REFERENCES [1] Coelho B. E.: Graduation Project Report - WebMeeting. Computer Science Engineering. Engineering Superior Institute of Porto (2007) [2] Gay G. & Hembrooke, H.: Activity-Centered Design: an Ecological Approach to Designing Smart Tools and Usable Systems. MIT Press (2004) [3] Fink J. & Kobsa A.: User Modeling for Personalized City Tours. In: Artificial Intelligence Review. Volume 18, Number 1. Kluwer Academic Publishers (2002) [4] Kobsa A. Generic User Modeling Systems. In: User Modeling and UserAdapted Interaction. Volume 11, Issue 1-2. Kluwer Academic Publishers (2001) [5] Kobsa A: User Modeling: Recent Work, Prospects and Hazards. In: Adaptive User Interfaces: Principles and Practice. North Holland Elsevier (1993) [6] Martins A. C., Faria L., Carvalho C. V. & Carrapatoso E.: User Modeling in Adaptive Hypermedia Educational Systems. In: Educational Technology & Society. Volume 11, 194-207 (2008) [7] Jung, C. G.: Psychological Types. Princeton University Press (1971) [8] Jennings A. & Higuchi H.: A Personal News Service Based on a User Model Neural Network. In: User Modeling and User-Adapted Interaction. Volume 3, Issue 1. Springer Netherlands (1992) [9] Rich E.: User Modeling Via Stereotypes. In: Cognitive Science: A Multidisciplinary Journal. Volume 3. Morgan Kaufmann Publishers Inc. (1979) [10] Mathes A.: Folksonomies - Cooperative Classification and Communication through Shared Metadata. In: Computer Mediated Communication - LIS590CMC (2004) [11] Fleming M. & Choen R.: Reasoning about Interaction in MixedInitiative AI Systems. In: IJCAI-03 - Workshop on Mixed-Initiative Intelligent Systems (2003) [12] Berka T. & Plöenig M.: Designing Recommender Systems for Tourism. In: ENTER 2004 (2003) [13] Felfernig A., Gordea S., Jannach D., Teppan E. & Zanker M.: A Short Survey of Recommendation Technologies in Travel and Tourism. In: ÖGAI Journal. Volume 25, Issue 2 (2006) [14] Cramer H., Evers V., Ramlal S., Someren M., Rutledge L., Stash N., Aroyo L. & Wielinga B.: The Effects of Transparency on Trust in and Acceptance of a Content-Based Art Recommender. In: User Modeling and User-Adapted Interaction. Volume 18, Issue 5, 455-496 (2008) [15] Coelho B., Martins C. & Almeida A.: ATOMICO - Arquitectura para Turistas com vista à Organização, Modelação e Inferência de COnhecimento - Representação do Utilizador. In: IADIS IberoAmericana - WWW / Internet 2008 (2008) [16] Pazzani M. J. & Billsus D.: Content-Based Recommendation Systems. In: Lecture Notes in Computer Science. Volume 4321. Springer Berlin / Heidelberg (2007) [17] Porter J.: Watch and Learn: How Recommendation Systems are Redefining the Web (2006) [18] TripAdvisor, http://www.tripadvisor.com/ (2010) [19] WAYN, http://www.wayn.com (2010) [20] Zukerman I. & Albrecht D. W.: Predictive Statistical Models for User Modeling. In: User Modeling and User-Adapted Interaction. Volume 11, Numbers 1-2, Pages 5-18 (2000) 622
