Semantic Recommender Systems. Analysis of the state of the topic

E. Peis; J. M. Morales-del-Castillo; J. A. Delgado-López

Citación recomendada: E. Peis; J. M. Morales-del-Castillo; J. A. Delgado-López. Analysis of the state of the topic [en linea]. "Hipertext.net", num. 6, 2008. <http://www.hipertext.net>

  1. Introduction
  2. Classification criteria for recommender systems.
  3. Semantic recommender systems: a bibliographic review
    3.1. Ontological and conceptual diagram based systems
      3.1.1. Trust network based systems
      3.1.2. Context-adaptable systems
  4. Conclusions
  5. References

 

1. Introduction

One of the main challenges information systems now confront is the effective management of large volumes of documents they store to easily and swiftly facilitate information consumers´ access to resources that satisfy their needs. This need becomes even more pressing in today's society where the user's demand continues to grow.

Traditionally in the context of libraries, the challenges of information overloads was tackled through different measures, like the creation of selective diffusion of information services (SDI), in which, depending on the profile agreement of the users subscribed to the service, a series of alerts notifying the existence of resources that suit the user's interests is generated periodically (or upon user request) (Aksoy et al., 1998), (Foltz; Dumais, 1992).

The Web, despite having unique characteristics that clearly distinguish it from libraries, suffer from essentially the same problem, and to handle them it has focused on applying similar solutions, such as information filtering systems (also known as recommender systems). In fact, recommender systems could be considered SDI systems applied to the Web, but obviously with infinitely greater and more sophisticated filtering capacities. These systems apply information filtering techniques that enable user access to information they need. In textual domains, filtering systems evaluate and encode the available resources on the Web (usually in HTML or XML formats) basically to assist the user in information retrieval tasks (Resnick; Varian, 1997) (mainly through filtering agents), even though they are also used to predict the user's valuation on items they have not yet evaluated (Szomszor, 2007).

The origin of these types of systems goes back to the early nineties, when news filtering services arose within newsgroups allowing its community of users to access only that which may possibly interest them (Foltz; Dumais, 1992) (Resnick et al., 1994) (Stodolsky, 1990).

Now recommender systems have evolved and it is possible to find them in various fields of applications like e-mails, where it has become a fundamental tool for on-line providers (Schafer; Konstan; Riedl, 2001) or in scientific information services.

However, each domain presents different problems, each requiring different solutions. The necessary ability to evolve has forced recommender systems to diversify.

Within the multiple possible categories of recommender systems existing in this paper, we are going to focus on reviewing the semantic recommender systems, which in the last five years have been the focus of discussion in the expert literature, and because it is characterised as basing its performance on different technology and vocabulary from the Semantic Web.

The article is structured as follows: section 2 lists the different classification criteria that traditionally have been applied to recommender systems. Section 3 reviews the main semantic recommender systems discussed in the literature. Finally, we discuss some conclusions.

 

2. Classification criteria for recommender systems.

Traditionally, filtering and recommender systems were classified into three categories relative to the filtering technique used (Popescul et al., 2001): social-based recommender systems, content-based recommender systems and economic factor-based recommender systems. Let us look at each one in greater detail.

The social filtering systems, also known as collaborative filtering, use the information provided by users with similar characteristics to generate recommendations, removing content from the resources (this is based exclusively on the valuations received from the users). In this type of system users are often grouped into specific categories or stereotypes that characterise them through a series of default preferences values and which represent the group's common information needs and search habits. This type of system tends to offer poor results when little information is available on the users or when they have heterogeneous tastes (Popescule et al, 2001).

The content-based filtering systems generate recommendations comparing user preferences (expressed implicitly or explicitly) with the metadata or characteristics used in representing the resources or products, ignoring the information concerning other users. These systems, just like the social-based systems, are not very reliable when little user information is available.

Economic factor-based recommender systems are those that generate recommendations based on cost (Resnick et al., 1994). For example, the relationship between the service cost and the benefit reported by the client, or the relationship between the bandwidth and size of the file to be downloaded. However, the use of this type of systems today is still uncommon.

The current trend is to develop hybrid filtering systems that combine characteristics of the content-based and collaborative-based systems, to minimise the disadvantages of each of them and thus improve the overall efficiency of the system's performance in terms of precision and comprehensiveness (Basu; Hirsh, Cohen, 1998) (Balabanovic; Shoham, 1997).

However, there are other criteria that could be used to classify them. For example, focusing on the way user preferences are obtained, we can distinguish between explicit data collection systems (when the user is asked to voluntarily provide their valuations) and implicit data collection systems (where the system user is monitored).

Depending on the information filtering method, there are passive filtering systems (Rafter; Bradley; Smyth, 1999): when a single recommender is generated for all system users; and active filtering systems (Boutilier; Zemer; Marlin, 2003) (Maltz; Ehrlich, 1995): where the recommendation is generated from the user's recommendation history to generate new, customised recommendations. An example of passive systems are those that recommend the most valued items by the user community (that is, the user does not directly influence the information received). The active systems are more complicated since they generate recommendation from the opinions of similar users. Moreover, within the active systems you can define the different information retrieval models (Gnasa et al, 2005): The pull information model (when a user must execute a query in the system to receive the recommendation) and the push information model (when the query is implicitly made through the preferences defined in the user profile).

There are also distinctions made between user-centred filtering systems (Xin et al., 2005), when the recommendations are made by comparing the similarity between users relative to their saved profile preferences; and item or product-based filtering systems (Resnick et al., 2004), where the relationship between different items is first found, then recommendations are generated from the active user preferences.

Focusing on the way in which the filtering algorithm processes the information, we find systems that load all the stored data memory to generate the recommendations and others that use a limited amount of data from which predictive models are generated, allowing them to process the recommendations more quickly.

Another criteria consists in distinguishing between centralised systems (when the product descriptions and user profiles are stored in a centralised Server) or non-centralised Systems (generally developed on P2P networks).

As we can see, there many different kinds of classification criteria, along with those not listed here. However, here we focus on analysing the state of this subject for a specific category of recommendation system that are widely covered in the scientific literature in recent years: semantic recommender systems.

 

3. Semantic recommender systems: a bibliographic review

Independent of the type of system, we have found that a common characteristic in most semantic recommender systems is the use of profiles to represent the users´ long-term information needs and interests. Therefore user profiles have become a key part of efficient filtering in recommender systems, since an inadequate profile may lead to low quality and irrelevant user recommendations.

As we have seen, the recommender systems normally use software instead of users for the information filtering tasks. However, these models do have some weaknesses: i) the communication processes between agents, and between agents and users are complicated by the various ways in which the information is represented; and ii) the heterogeneity of the information's representations makes it incapable of being reused in other processes and applications.

To mitigate these deficiencies, the representation of the information may be improved and enriched through the application of technologies used in the Semantic Web (Berners-Lee; Hendler; Lassila, 2001).

Here we will consider semantic recommender systems as any system that bases its performance on a knowledge base, normally defined through conceptual maps (like a taxonomy or thesaurus) or an ontology, and that use technologies from the Semantic Web.

The Semantic Web project is presented as an extension of the current Web, hoping to become a universal platform for information exchange. In the Semantic Web's model, information is granted a well-defined meaning that allows for better human / machine collaboration (Berners-Lee, 2000). The project is basically based on two ideas: the resources´ semantic tags (implying a formal separation between document content and structure) and the development of software agents capable of processing and operating these resources on a semantic level (Berners-Lee, 2001), (Hendler, 2001) using ontologies (understood as formalised conceptualisations of a plot of reality that gives full meaning to information in a certain context (Guarino, 1998) (Gruber, 1995).

In recent years many recommender systems have appeared that use Semantic Web technologies and that propose various application solutions in different fields. Basically, they can be classified as: Ontological or conceptual map based systems, along with other systems that add additional information systems like context-adaptable systems and trust network-based systems.

3.1. Ontological and conceptual diagram based systems

The system designed by Wang and Kong (2007) is a personalised recommender system which tries to limit the problems of collaborative recommender systems by ontologically using semantic information from the categorical characteristics of an item. The similarities between user pairs is calculated by a weighted mean method that calculates three similarity measures: the similarity of user evaluation histories (using the Pearson correlation coefficient on usage information of the system in terms of a user-item evaluation data matrix); the similarity of these user's demographic data (calculated with a weighted mean); and the users similarity in interest or preference based on the semantic similarities of the items retrieved and/or evaluated. At the same time, the system incorporates an offline-user cluster mechanism to limit the scalability problem.

Khosravi, Farsani and Nematbakhsh (2006) suggest a methodology for personalised recommendations in the context of e-commerce. This is a procedure to recommend products to potential clients. The proposed algorithm is based on modelling information on products and users with OWL (Ontology Web Language). The process starts by classifying the products and consumers with OWL, which enables the analysis of product-client similarity. In a second phase, active consumers are selected, keeping in mind previous recommendations (the system does not make recommendations to a client if the number of previous recommendation does not pass a threshold). The product and client classification is used to create a matrix of product-client evaluations. The algorithm recommends some products from each class within the classes of products based on the number of evaluations in the matrix.

Another model used in the field of e-commerce is one presented by Ziegler, Lausen and Schmidt-Thieme (2004). The system is based on the collaborative-recommender paradigm through content (Pazzani, 1999) using a product taxonomy from which the user profiles are defined (without users needing to provide explicit valuations). The active user profile is used to discover users with similar interests, whose valuations help the system generate recommendations.

Jung and collaborators (2005) propose a recommender system based on personal information which they claim suits the Semantic Web context.. The model is based on the representation of Web services and user profiles with RDF triples (Resource Description Framework). Each company wanting to provide Web services registers its data in the information repository, where the system converts the data into documents in RDF format. The search module extracts the repository's information and sends it to the document retrieval agent. The agent accesses the space with the corresponding name and retrieves the RDF documents relative to the required Web services. These documents are sent to the information integration agent, where they are merged in a single RDF document containing the relevant information. Finally, the information retrieval agent extracts the most relevant RDF triples in accordance with the user profile and offers the user the stored objects that coincide with those triples.

Other systems are defined with decentralised structures like P2P networks. For example, the model presented by Díaz-Avilés (2005), where the information is not available in a centralised repository, but in each of the network's components. The items or objects are modelled through a common ontology that uses all members of the network. The selection of the network's components is done dynamically and the recommendations are generated using a nearest neighbour-based recommender algorithm that is locally executed in each of the network components.

An original approximation is Cantador´s and Castells´ (2006) proposal to develop a multi-layer semantic social network model that can define the system from different perspectives, all from common interests shared by the network's members. From a series of generated user profiles using ontological concepts, and keeping in mind their common preferences, the system is capable of marking out the domain's different concept groups. From these groups, we can identify a set of users with similar interests that interrelate at different semantic levels (according to their preferences). This method allows us to find implicit social networks that may help to define both content-based and collaborative-based recommender systems.

Besides these generic models, there are many others that are defined for more specific domains like the "Foafing the Music" project (Celma; Ramirez; Herrera, 2005) that recommends music and related news; the system proposed by Middleton and collaborators (2002) to recommend scientific articles is based on an automatically generated ontology from information dynamically extracted from various sources (online information, user monitoring, feedback, etc.); the AVATAR multi-agent system (Blanco Fernández et al., 2004) specialises in television programmes; or SemMF (Oldakowsky; Byzer, 2005) that generates job offer recommendations calculating the semantic similarity of two concepts in accordance with its location within a concept hierarchy.

3.1.1. Trust network based systems

As we have a seen, one of the main concerns researchers have in the field of semantic recommender systems is guaranteeing reliability and precision of the recommendations generated. This is why many offer an additional filtering level based on a trust network (Ziegler, 2004).

Adrian, Sauermann and Roth-Berghofer (2007) propose ConTag, a system that advises users on the most suitable tags to describe the content of different resources used in a Web 2.0 platform. The documents are transformed to a RDF format to then generate a topic map defined in SKOS Core, from which you can calculate topic similarity with concepts defined in the ontological system. The tags recommended by the system may help the user better describe the resource's content, minimising problems caused by synonyms, homonyms, acronyms and spelling variants that may occur with user defined tags. Jäschke and collaborators (2007) propose a similar tag recommendation mechanism.

Szomszor and collaborators (2007) propose the use of prediction algorithms for the integration of a folksonomy of movies with a semantic knowledge base on user movie rentals. The folksonomy is used to enrich the knowledge base with descriptions and the categorisation of movie titles, along with the representation of user opinions and interests. The cloud of tags generated by the folksonomy is used to construct improved user profiles that reflect the user's level of interest in different types of movies, thus providing a base for predicting scores of unseen movies. For the construction of a movie knowledge base and on how the users rent them, Web 2.0 sources were used (Internet Movie Database and Netflix Prize, respectively). To provide a homogeneous representation of both sets of data, an OWL ontology was created in conjunction with the D2RQ mapping technology that allows one to process the data from a relational database, like a virtual RDF graph.

Along these lines, Massa and Avessani (2004) present a model based on the denominated "Web of Trust." The system takes inputs from a "web of trust" (representing the set of a community's trust declarations -the users make explicit those users they trust) and a matrix (representing all of the user item evaluations) and produces as output a matrix of predictable evaluations that users can assign to the items. This matrix is used by the system to recommend user preferred items. That is, the system selects the highest valued items from a user predictable evaluation column. This way the collaborative recommender system's coverage is expanded while also maintaining the quality of the recommended items, thus mitigating the problems arising form cold starts (for new users) and the lack of trust from malicious user data.

A different approach is provided by the FOAFRealm system (Kruk; Decker, 2005). Built on P2P networks, it has a library that allows users to manage their own defined profiles defined in FOAF (Friend of a Friend vocabulary). At the same time, the descriptions and categories defined by the users are enriched with terms extracted from taxonomic outlines or ontologies. To manage the security problems that often arise in recommender systems based on distributed architectures, the concept "extrapolated profile" is defined where new users receive a profile that is constructed from a user profile that is added as "friend," and that depends directly on the level of trust granted by them (making it easy to locate malicious users).

Filmtrust works in a similar fashion (Golbek, 2005). This website includes a collaborative recommendation system of movies using FOAF vocabulary as a base for creating a social network of trust. Within this network, each individual must value the trust to grant other users added to their friend network (from which the movie recommendations are made). This system also has a mechanism to determine the precision of the recommendations generated.

Bedi and collaborators (2007) define a new model based on the use of ontological series and establishing trust networks between agents. Each agent has a temporary personal ontology (leaving room for any possible changes made in time) from which they can generate domain-independent recommendations. To calculate these recommendation values, the similarity between agents is implicitly assumed in the trust value, assigned by the agents themselves, to each other. Another nuance in these types of systems is the use of intuitive fuzzy sets (Atanassov, 1999) to manage the inherent uncertainty in the recommender process.

The current literature also presents other recommender system models that use Web 2.0 technology like RSS channels. Like the model presented by Kobayashi and Saito (2006) which proposes a recommender system for newspaper articles and news extracted from RSS channels, using a thesaurus as support to represent both news contents and the users information needs.

Another example of this type of vocabulary is in the model presented by Peis, Herrera-Viedma and Morales-del-Castillo (2008), where a series of RSS fulfil the role of "News Bulletins," within a DSI service for specialised digital libraries. The system is built on a multi-agent platform and is capable of generating alerts and recommendations in accordance with the user profile's defined preferences. These profiles are built from a specialised thesaurus and is dynamically updated through an implicit user preference discovery mechanism which uses fuzzy linguistic modelling techniques.

3.1.2. Context-adaptable systems

Another possibility are context-adaptable (or sensitive) recommender systems. These types of systems analyse and take different factors into consideration (time, location, user experience, the device used upon receiving the recommendation, etc.) to infer the user's context and adapt the recommendation to these circumstances.

Within this group we find, for example, the content-based model proposed by Kim and Kwon (2007). The system's performance is based on the definition of the "use contexts" that correspond to the different levels of specificity to an ontology of concepts. This system generates a recommendation from the set of items most valued by the user and then is capable of adapting it to the level of specificity of the information presented to the user, depending on the use context in which the concepts of interested are located. To determine these contexts, the system uses four ontologies: an ontology of products, another defining use contexts, a third on the historical user system activity registry, and finally a user ontology.

Loizou and Dasmahapatra (2006) propose a system based on an ontology that includes contextual information both of the recommendation process and of the items to recommend. This contextual information (for example, the recommendation time or the recommended item's usefulness to the user) processed with a mathematical device with heuristic rules applied on vectorial spaces allows the system to dynamically evaluate the suitability of a specific recommendation.

Yu´s and collaborators (2007) proposal is also interesting, applying this type of system to the field of e-learning. The model attempts to facilitate the necessary resources to students in order to carry out course work. It is based on the use of ontologies to efficiently represent the system's knowledge of users, resource content and the system's specific domain speciality to generate recommendations based on use context (that is, the recommendation is different depending on the different factors like experience level, course progress level at the time the user recommendation is made). The system is not limited to recommending context suited resources; it can also suggest other related resources for further knowledge on the subject and thus define a complete study programme.

Other models applying this type of filtering were proposed by Laliwala, Sorathia and Chaudhary (2006) where a semantic recommendation system is developed for agricultural information services based on events, or that presented by Woerndl, Schueller and Wojtech (2007) where the ontologies are used to improve the Web service descriptions offered by third parties, and to develop a specialised collaborative recommender service for tourism information through a mobile device that can handle both static information(defined in the user profiles) and dynamic information (contextual).

 

4. Conclusions

The semantic recommender systems are those whose performance are based on a knowledge base usually defined as a concept diagram (like a taxonomy or thesaurus) or an ontology.

The use of ontologies in these types of systems limits specific problems, including the following:

  1. To guarantee the inter-operability of system resources and the homogeneity of the representation of information.

  2. To allow for the dynamic contextualisation of user preferences in specific domains.

  3. To facilitate performance in social networks and collaborative filtering.

  4. To improve communication processes between agents and between agents and users.

  5. To limit the "cold start" problem by completing the incomplete information through inferences.

  6. The ability to semantically extend descriptions of user contextual factors.

  7. To improve the representation and description of different system elements.

  8. Improve the description of system's logic by admitting the inclusion of a set of rules.

  9. Provide the necessary means to generate descriptions enriched by web services and facilitate their discovery by software agents.

Of all the semantic recommender systems, those using Semantic Web technology to define the knowledge base are the most promising in terms of short and mid-term results.

However, the most solid future line of research focuses on the development of mixed systems that use tools involved in developing the Semantic Web project, along with additional filters, like those based on a trust network (ensuring the processes´ results reliability) and those using contextual information (allowing to increase filtering precision.)

 

5. References

Aksoy, D. et al. (1998). "Research in Data Broadcast and Dissemination". Proc. of the 1st International Conference on Advanced Multimedia Content Processing, pp.194-207.

Balabanovic, M.S.Y. (1997). "Fab: content-based, collaborative recommendation." Communications of the ACM, v. 40, pp. 66-72.

Atanassov, K.T. (1999). " Intuitionistic Fuzzy Sets: Theory and Applications ". Heidelberg: Physica-Verlag.

Basu, C.; Hirsh, H. y Cohen, W. (1998). "Recommendation as classification: Using social and content-based information in recommendation." Proc. of the 15th National Conference on Artificial Intelligence, pp. 714-720.

Bedi, P.; Kaur, H. y Marwaha, S. (2007). "Trust based Recommender System for the Semantic Web". Proc. of the IJCAI07, pp. 2677-2682.

Berners-Lee, T., Hendler, J. y Lassila, O. (2001). "The Semantic Web: A new form of Web content that is meaningful to computers will unleash a revolution of new possibilities". Scientific American, May.

Blanco Fernández, Y. et al. (2004). "AVATAR: An Advanced Multi-Agent Recommender System of Personalized TV Contents by Semantic Reasoning". En: X. Zhong et al., (Ed.) "Web Information System Engineering." Lecture Notes in Computer Science, LNCS. Berlin, Heidelberg: Springer Verlag, pp. 415-421.

Boutilier, C.; Zemel, R.S. y Marlin, B. (2003). "Active collaborative filtering". Proc. Of the 19th Annual Conference on Uncertainty in Artificial Intelligence, pp. 98-106.

Cantador, I. y Castells, P. (2006). "Multilayered Semantic Social Network Modeling by Ontology-Based User Profiles Clustering: Application to Collaborative Filtering". Managing Knowledge in a World of Networks, pp. 334-349.

Celma, O; Ramírez, M. y Herrera, P. (2005). "Foafing the music: A music recommendation system based on RSS feeds and user preferences". Proc. Of the 6th Int. Conference on Music Information Retrieval (ISMIR), pp. 464-457.

Díaz-Avilés, V.E. (2005). Semantic Peer-To-Peer recommender Systems (Tesis doctoral), < http://citeseer.ist.psu.edu/semantic-peer-to-peer1.pdf > [Retrieved 23/02/2008].

Farsani, H.K. y Nematbakhsh, M.A. (2006). "A Semantic Recommendation Procedure for Electronic Product Catalog". International Journal of Applied Mathematics and Computer Sciences, v. 3, pp. 86-91.

Foltz, P.W. y Dumais, S.T. (1992). "Personalized information delivery: an analysis of information filtering methods". Communications of the ACM, v. 35, pp. 51-60.

Gnasa, M. et al. (2005). "Personalized Peer Filtering for a Dynamic Information Push". En: "Anonymous Foundations of Intelligent Systems." Berlin, Heidelberg: Springer Verlag, pp. 650-659.

Golbeck, J. (2005). "Semantic Web Interaction through Trust Network Recommender Systems". Proc. of the ISWC05 Workshop on End User Semantic Web Interaction, pp. 2005.

Gruber, T.R. (1995). "Toward principles for the design of ontologies used for knowledge sharing". International Journal of Human-Computer Studies, v. 43, pp. 907-928.

Guarino, N. (1998). "Formal ontology and information systems." En N. Guarino (ed.). "Formal ontology and information systems: Proc. of FOIS'98." Amsterdam: IOS Press, pp. 3-17.

Hendler, J. (2001). "Agents and the Semantic Web." IEEE Intelligent Systems, n. March/April, pp. 30-37.

Jäschke, R. et al. (2007). "Tag Recommendations in Folksonomies." Knowledge Discovery in Databases: PKDD07, pp. 506-514.

Jung, K. et al. (2005). "RDF Triple Processing Methodology for the Recommendation System Using Personal Information." Proc. Of the Int. Conference on Next Generation Web Services Practices, pp. 241-247.

Kim, S. y Kwon, J. (2007). "Effective Context-aware Recommendation on the Semantic Web." International Journal of Computer Science and Network Security, v.7, pp. 154-159.

Kobayashi, I. y Saito, M. (2006). "A study on information recommendation system that provides information related to user's inquiry for information retrieval." Proc. Oh the 2006 IEEE/WIC/ACM Int. Conference on Web Intelligence and Intelligent Agent Technology, pp. 385-388.

Kruk, S.R. y Decker, S. (2005). "Semantic Social Collaborative filtering with FOAFRealm." Proc. Of the Semantic Desktop ISWIC.

Laliwala, Z.; Sorathia, V. y Chaudhary, S. (2006). "Semantic and Rule Based Event-driven Services-Oriented Agricultural Recommendation System." Proc. Of the 26th IEEE Conference on distributed Computing Systems Workshop (ICDSW'06), pp. 24-30.

Loizou, A. y Dasmahapatra, S. (2006). "Recommender Systems for the Semantic Web." Proc. of the ECAI 2006, pp. 76-81.

Maltz, D. y Elrich, K. (1995). "Pointing the way: active collaborative filtering." Proc. of the SIGCHI Conference on human Factors in Computing Systems, pp. 202-209.

Massa, P. y Avessani, P. (2004). "Trust-Aware Collaborative Filtering for Recommender Systems." En: "On the Move to Meaningful Internet Systems 2004: CoopIS, DOA, and ODBASE." Berlin, Heidelberg: Springer, pp. 3-17.

Middleton, S.E. et al. (2002). "Exploiting synergy between ontologies and recommender systems." Proc. of the Semantic Web Workshop, v. 55, pp. 41-50.

Oldakowsky, R. y Byzer, C. (2005). SemMF: A framework for calculating semantic similarity of objects represented as RDF graphs (Poster of the 4th Semantic Web Conference), < http://ww.corporate-semantic-web.de/pub/SemMF_ISWC2005.pdf > [Retrieved: 07/06/2007].

Pazzani, M.J. (1999). "A framework for collaborative, content-based and demographic filtering." Artif. Intell. Rev., v. 13, pp. 393-408.

Peis, E., Herrera-Viedma, E. y Morales-del-Castillo, J.M. (2008). "Servicio semántico de difusión selectiva de información (DSI) para bibliotecas digitales." El profesional de la información[in press],

Popescul, A. et al. (2001). "Probabilistic models for unified-collaborative and content-based recommendation in sparse-data environments." Proc. of the 17th Conference on Unceertainty in Artificial Intelligence (UAI), pp. 467-444.

Rafter, R.; Bradley, K. y Smyth, B. (1999). "Passive Profiling and Collaborative Recommendation." Proc. of the 10th. Irish Conference on Artificial Intelligence and Cognitive Science.

Resnick , M.L. et al. (2004). "Persuasive Design Through Intelligent Recommendation Systems." Proc. of the Human Factors and Ergonomics Society Annual Meeting, pp. 1503-1507.

Resnick, P. et al. (1994). "GroupLens: An open architecture for collaborative filtering of netnews." Proc. of the ACM Conference on Computer Supported Cooperative Work, pp. 175-186.

Resnick, P. y Varian, H.R. (1997). "Recommender Systems." Special issue of Communications of the ACM, v. 40.

Schafer, J.B.; Konstan, J.A. y Riedl, J. (2001). "E-commerce recommendation applications." Data Mining and Knowledge Discovery , v. 5, pp. 115-153.

Stodolsky, D.S. (1990). "Invitational Journals Based Upon Peer Consensus." Psycoloquy, v. 1 (15).

Szomszor, M. et al. (2007). "Folksonomies, the Semantic Web, and Movie Recommendation." Proc. of the ESWC'07, pp. 71-84.

Wang, R.Q. y Kong, F.S. (2007). "Semantic-Enhanced Personalized Recommender System." Machine Learning and Cybernetics: Proc. of the Int. Conference on Machine Learning and Cybernetics , v. 7 (19-22), pp. 4069-4074.

Woerndl, W.; Schueller, C. y Wojtech, R. (2007). "A Hybrid Recommender System for Context-aware Recommendations of Mobile Applications." Proc. of the IEEE 3rd International Workshop on Web Personalisation, Recommender Systems and Intelligent User Interfaces (WPRSIUI'07).

Xin, Z. et al. (2005). "Information Push-Delivery for User-Centered and Personalized Service." En: "Anonymous Fuzzy Systems and Knowledge Discovery." Berlin, Heidelberg: Springer, pp. 594-602.

Yu, Z. et al. (2007). "Ontology-Based Semantic Recommendation for Context-Aware E-Learning." Proc. of the 4th Conference on Ubiquitous Intelligence and Computing, v.4611, Berlin, Heidelberg: Springer, pp. 898-907.

Ziegler, N. (2004). "Semantic Web Recommender Systems." Proc. of the joint ICDE/EDBT Workshop.

Ziegler, N.; Lausen, G. y Schmidt-Thieme, L. (2004). "Taxonomy driven computation of product recommendations." Proc. of the 30th ACM Conference on Information and Knowledge Management, pp. 406-415.



Creative Commons License
Last updated 05-06-2012
© Universitat Pompeu Fabra, Barcelona