Traditional Recommender Systems Hybrid Approaches a Heuristic approaches a Model-based approaches o Combine recommendations D Classifier with multiple m Weighted( Claypo types of features(Basu et al 口 Hierarchica|Bay 2000 (Ansani et al. 2000: Condlff et al. 1999) m Voting(Pazzani 1999) O Latent-class generative 口 Feature augmentation models(Popescul et Schein et al. 2002) 口 Relational learning D Graph-based method(Huang methods(probabilistic shami 1999- Ho Newton Greiner 2004) Using context for Recommendations a Early work: task-focused recommendation(Herlocker Konstan 200 a Knowledge about users task can lead to better recommendations a Operates within the traditional 2D User x Item space m Task specification: a list of sample items related to the task at hand u Context-aware information access/retrieval (Jones 2005) O Assisting search-and querying-based user activities using the knowledge of context information-significant amount of work a Typically no modeling of (long-term)user preferences E.g., "find all files created during a spring meeting on a sunny day outside an Italian restaurant in New york Interactive and proactive retrieval of previously seen information Support of mobile users( time-and location-based capabilities, domain knowledge base, rich user interface, etc. ) e.g., travel applications u Human digital memories
Traditional Recommender Systems: Hybrid Approaches Heuristic approaches Combine recommendations Weighted (Claypool et al. 1999) Mixed (Smyth & Cotter 2000) Switching (Billsus & Pazzani 2000) Voting (Pazzani 1999) Feature augmentation (Melville et al. 2002; Soboroff & Nicholas 1999) Graph-based method (Huang et al. 2004) Model-based approaches Classifier with multiple types of features (Basu et al. 1998) Hierarchical Bayesian (Ansari et al. 2000; Condliff et al. 1999) Latent-class generative models (Popescul et al. 2001; Schein et al. 2002) Relational learning methods (probabilistic relational models) (Getoor & Sahami 1999; Huang et al. 2004; Newton & Greiner 2004) Using Context for Recommendations Early work: task-focused recommendation (Herlocker & Konstan 2001) Knowledge about user’s task can lead to better recommendations Operates within the traditional 2D User × Item space Task specification: a list of sample items related to the task at hand Context-aware information access/retrieval (Jones 2005) Assisting search- and querying-based user activities using the knowledge of context information – significant amount of work Typically no modeling of (long-term) user preferences E.g., “find all files created during a spring meeting on a sunny day outside an Italian restaurant in New York” Key applications Interactive and proactive retrieval of previously seen information Support of mobile users (time- and location-based capabilities, domain knowledge base, rich user interface, etc.), e.g., travel applications Human digital memories
Context in Recommender Systems I Focus of this tutorial: contextual recommender systems D Modeling and predicting(long-term) user preferences( e.g, ratings) Data in traditional recommender systems a Rating information: <user, item, rating> D Also, descriptive information/attributes about items(e.g, movie genre)and users(e.g, demographics) a Data in context-aware recommender systems D Rating information: <user, item, rating, context a In addition to information about items and users, also may have descriptive information/attributes about context E.g, context hierarchies(Saturday Weekend a Fundamental questions a How to model context with respect to user preferences? D Can traditional(non-contextual)recommender systems be used to generate context-aware recommendations? Relevance of contextual lnformation a Not all contextual information is relevant for generating recommendations m E.g., which contextual information is relevant when recommending a book? D For what purpose is the book bought? (Work, leisure,. o Where will the book be read?(At home, at school, on a plane,. D How is the stock market doing at the time of the purcha a Determining relevance of contextual information D Manually, e.g., using domain knowledge of the recommender s designer sing feature selection procedures or statistical tests based on existing ratings data a We assume that only the relevant contextual information is
Context in Recommender Systems Focus of this tutorial: contextual recommender systems Modeling and predicting (long-term) user preferences (e.g., ratings) Data in traditional recommender systems Rating information: <user, item, rating> Also, descriptive information/attributes about items (e.g., movie genre) and users (e.g., demographics) Data in context-aware recommender systems Rating information: <user, item, rating, context> In addition to information about items and users, also may have descriptive information/attributes about context E.g., context hierarchies (Saturday Æ Weekend) Fundamental questions: How to model context with respect to user preferences? Can traditional (non-contextual) recommender systems be used to generate context-aware recommendations? Relevance of Contextual Information Not all contextual information is relevant for generating recommendations E.g., which contextual information is relevant when recommending a book? For what purpose is the book bought? (Work, leisure, …) When will the book be read? (Weekday, weekend, …) Where will the book be read? (At home, at school, on a plane, …) How is the stock market doing at the time of the purchase? Determining relevance of contextual information: Manually, e.g., using domain knowledge of the recommender system’s designer Automatically, e.g., using feature selection procedures or statistical tests based on existing ratings data We assume that only the relevant contextual information is kept
Approaches to Integrating Context and User preferences Loose coupling of context and user preferences D Assumption: user preferences don't depend on the context however, the item consumption may depend on the context D E.g., Rating (Me, " For Whom The Bell Tolls)=9; however, I never read long and serious novels on a weekend o Allows to use traditional non -contextual recommenders a Tight coupling of context and user preferences D Assumption: user preferences directly depend on the context D E.g., Rating(Me, "For Whom The Bell Tolls", Saturday)=9 D Requires more complex rating prediction techniques Which approach to use depends on the application How to Use Context in the Recommendation process Context can be used in the following stages a Contextual pre-filtering O Loose coupling of context and user preferences o Contextual information drives data selection for that context Ratings are predicted using a traditional recommender on the selected data a Contextual post-filtering O Loose coupling of context and user preferences D Ratings predicted on the whole data using traditional recommender The contextual information is used to adjust("contextualize")the resulting set of recommendations ng ■ Contextual modelin a Tight coupling of context and user preferences D Contextual information is used directly in the modeling technique as a part of rating estimation
Approaches to Integrating Context and User Preferences Loose coupling of context and user preferences Assumption: user preferences don’t depend on the context; however, the item consumption may depend on the context E.g., Rating(Me, “For Whom The Bell Tolls”) = 9; however, I never read long and serious novels on a weekend Allows to use traditional non-contextual recommenders Tight coupling of context and user preferences Assumption: user preferences directly depend on the context E.g., Rating(Me, “For Whom The Bell Tolls”, Saturday) = 9 Requires more complex rating prediction techniques Which approach to use depends on the application How to Use Context in the Recommendation Process Context can be used in the following stages: Contextual pre-filtering Loose coupling of context and user preferences Contextual information drives data selection for that context Ratings are predicted using a traditional recommender on the selected data Contextual post-filtering Loose coupling of context and user preferences Ratings predicted on the whole data using traditional recommender The contextual information is used to adjust (“contextualize”) the resulting set of recommendations Contextual modeling Tight coupling of context and user preferences Contextual information is used directly in the modeling technique as a part of rating estimation