《电子商务 E-business》阅读文献：Context-Aware Recommender Systems

2nd ACM international Conference on Recommender Systems( RecSys 2008 Context-Aware Recommender Systems October 23. 2008 Gedas Adomavicius U. of minnesota Alex Tuzhilin, New York University Quick Note on Terminology a Several terms have been used to describe recommender systems that can take advantage of contextual information a Google(as of October 19, 2008) 口 Context- aware context-aware recommender systems"(113 results) m"context-aware recommendations"(151 results contextual recommender systems"(11 results) contextual recommendations"(775 results) 口 Context-dependent context-dependent recommender systems"(5 results) context-dependent recommendations"(17 results) In this tutorial we use"context-aware"and"contextual", but the discussion is welcome regarding the most appropriate term

Context-Aware Recommender Systems Gedas Adomavicius, U. of Minnesota Alex Tuzhilin, New York University 2nd ACM International Conference on Recommender Systems (RecSys 2008) October 23, 2008 Quick Note on Terminology „ Several terms have been used to describe recommender systems that can take advantage of contextual information „ Google (as of October 19, 2008) … Context-aware „ “context-aware recommender systems” (113 results) „ “context-aware recommendations” (151 results) … Contextual „ “contextual recommender systems” (11 results) „ “contextual recommendations” (775 results) … Context-dependent „ “context-dependent recommender systems” (5 results) „ “context-dependent recommendations” (17 results) „ In this tutorial, we use “context-aware” and “contextual”, but the discussion is welcome regarding the most appropriate term

Motivation Motivating EXamples: Context- Dependent Recommendations ■ Recommend a vacation 口 Winter vs. summer a Recommend a purchase(e-retailer) 口 Gift vs. for yourself ■ Recommend a movie O To a student who wants to see it on Saturday night with his girlfriend in a movie theater a Recommendations depend on the context a It is sometimes important to know not only what to recommend to whom but also under what circumstances

Motivation Motivating Examples: Context￾Dependent Recommendations „ Recommend a vacation … Winter vs. summer „ Recommend a purchase (e-retailer) … Gift vs. for yourself „ Recommend a movie … To a student who wants to see it on Saturday night with his girlfriend in a movie theater „ Recommendations depend on the context … It is sometimes important to know not only what to recommend to whom, but also under what circumstances

Rudimentary Contextual Recommendations: Amazon amazonicum Your Account Cat I Your Lists C Heb I ommended for you a Amazon makes sure that DIt is you 口Hasa"git” button a But there is much more to capturing and using contexts in recommendations Question: Does Context Matter? Matters enough for Amazon to add the gift button aC K Prahalad, Beyond CRM, MWorld/AMA, 2004 o C K Prahalad predicts: "Customer Context is the Next"Big Thing o The ability to reach out and touch customers anywhere at anytime ans that companies must deliver not just competitive products but Iso unique, real-time customer experiences shaped by custome ontext a Goal: Demonstrate that certain contextual information does matter in some recommendation applications D E.g., recommending a vacation in the winter or a movie to see on Saturday night with a girlfriend in a movie theate

Rudimentary Contextual Recommendations: Amazon „ Amazon makes sure that …It is you …Has a “gift” button „ But there is much more to capturing and using contexts in recommendations… John Doe’s Question: Does Context Matter? „ Matters enough for Amazon to add the “gift” button „ C.K. Prahalad, Beyond CRM, MWorld/AMA, 2004: … C.K. Prahalad predicts: “Customer Context” is the Next “Big Thing” … “The ability to reach out and touch customers anywhere at anytime means that companies must deliver not just competitive products but also unique, real-time customer experiences shaped by customer context” „ Goal: Demonstrate that certain contextual information does matter in some recommendation applications … E.g., recommending a vacation in the winter or a movie to see on Saturday night with a girlfriend in a movie theater

Outline of the tutorial ■ What is context? a Incorporating context in recommender systems a conceptual framework a Different paradigms for contextual recommender a Additional capabilities for contextual recommender systems ■ Future directions What Is context?

Outline of the Tutorial „ What is context? „ Incorporating context in recommender systems: a conceptual framework … Different paradigms for contextual recommender systems „ Additional capabilities for contextual recommender systems „ Future directions What Is Context?

What Is Context? (Palmisano et al. 2008) Conditions or circumstances affecting some thing(Webster) a 150( other definitions from various communities/disciplines D Presented in(Bazire& Brezillon, CONTEXTO5 O DM/CRM. those events which characterize the life of a customer and can determine a change in his/her preferences and status, and affect the customer's value for a company (Berry Linoff, 1997) D Context-aware systems: the location of the user, the identity of people near the user, the objects around, and the changes in these elements ( Schilt Theimer, 1994) 口 Marketing: the same mer may use different decision-making strategies and prefer nt products or brands under different contexts(Bettman et al. 199 Kotler 1992, Lussier& Olshavsky 1979, Klein Yadav 989, Bettman et al. 1991) a Conclusion: many different approaches and views! Context in Context-Aware Systems a Location of the user, identity of people and objects near the user(schilt Theimer, 1994) Date, season, temperature(Brown, Bovey, Chen 1997) Physical and conceptual statuses of interest to a user(Ryan, Pascoe, Morse 1997 a Any information which can characterize and is relevant to the interaction between a user and an application(Dey, Abowd Salber 2001)

What Is Context? (Palmisano et al. 2008) „ Conditions or circumstances affecting some thing (Webster) „ 150(!) other definitions from various communities/disciplines … Presented in (Bazire & Brezillon, CONTEXT’05) … DM/CRM: those events which characterize the life of a customer and can determine a change in his/her preferences and status, and affect the customer’s value for a company (Berry & Linoff, 1997) … Context-aware systems: the location of the user, the identity of people near the user, the objects around, and the changes in these elements (Schilit & Theimer, 1994) … Marketing: the same consumer may use different decision-making strategies and prefer different products or brands under different contexts (Bettman et al. 1991, Lilien & Kotler 1992, Lussier & Olshavsky 1979, Klein & Yadav 1989, Bettman et al. 1991) „ Conclusion: many different approaches and views! „ Location of the user, identity of people and objects near the user (Schilit & Theimer, 1994) „ Date, season, temperature (Brown, Bovey, & Chen 1997) „ Physical and conceptual statuses of interest to a user (Ryan, Pascoe, & Morse 1997) „ Any information which can characterize and is relevant to the interaction between a user and an application (Dey, Abowd & Salber 2001) Context in Context-Aware Systems

What Is Context in Recommender Systems? a Additional information besides information on users and Items. that is relevant to recommendations ■Re| evant in D Identifying pertinent subsets of data when computing recommendations a Building richer rating estimation models D Providing various types of constraints on recommendation outcomes ■ Examples D Exclude gift purchases when recommending products to you D Use only winter-based ratings when recommending a vacation in the winter Defining context via contextual Variables Taxono contextual information More vs granular context (levels of context) The contextual information K of a purchase Context K 1 level Personal K=a GitK=阝 n level Work K=al Other K=a2 Friend/Partner K=B12 Parent/other K=B34 Friend K=Bl Partner K=B2 Parent K=B3 Other K=B4 Finer

What Is Context in Recommender Systems? „ Additional information, besides information on Users and Items, that is relevant to recommendations „ Relevant in … Identifying pertinent subsets of data when computing recommendations … Building richer rating estimation models … Providing various types of constraints on recommendation outcomes „ Examples: … Exclude gift purchases when recommending products to you … Use only winter-based ratings when recommending a vacation in the winter „ Taxonomy of contextual information „ More vs. less granular context (levels of context) The contextual information K of a purchase: Defining Context via Contextual Variables Context K Personal K=α Gift K=β Work K= α1 Other K= α2 Friend/Partner K=β12 Parent/Other K=β34 Friend K=β1 Partner K=β2 Parent K=β3 Other K=β4 Context K Personal K=α Gift K=β Work K= α1 Other K= α2 Friend/Partner K=β12 Parent/Other K=β34 Friend K=β1 Partner K=β2 Parent K=β3 Other K=β4 Rough Finer 1st level 2nd level 3rd level

EXample Temporal Contextual Variable a Time-related context can be described using a temporal hierarchy with multiple temporal relationships of varying granularity, e.g Tme(eg,2008.10.1911:59:59pm)Date (2008.10.19)→Year(2008) 口Tme(2008.10.1911:59:59pm)Hour(11pm) TimeofDay(evening) a Date(2008. 10.19)) Month(October), Season(Fall) o Date(2008.10.19)>DayofWeek(Sunday)> Time OfWeek(Weekend) Formalizing Contextual Information via Contextual variables a Formally, contextual information can be defined as a vector of contextual variables c=(C1,., cn) where c∈C OC=C1x.Cn denotes the space of possible values for a gIven context Each component C, is represented as a tree: it is defined as a hierarchical set of nodes(concepts) a If c EC, then c, represents one of the nodes in the hierarchy CI aC=(work, weekend), i.e., purchasing something for work on a weekend

„ Time-related context can be described using a temporal hierarchy with multiple temporal relationships of varying granularity, e.g., … Time (e.g., 2008.10.19 11:59:59pm) Æ Date (2008.10.19) Æ Year (2008) … Time (2008.10.19 11:59:59pm) Æ Hour (11pm) Æ TimeOfDay (evening) … Date (2008.10.19) Æ Month (October) Æ Season (Fall) … Date (2008.10.19) Æ DayOfWeek (Sunday) Æ TimeOfWeek (Weekend) Example: Temporal Contextual Variable Formalizing Contextual Information via Contextual Variables „ Formally, contextual information can be defined as a vector of contextual variables c = (c1,…,cn), where ci ∈Ci … C = C1×…×Cn denotes the space of possible values for a given context „ Each component Ci is represented as a tree: it is defined as a hierarchical set of nodes (concepts) „ If ci ∈Ci , then ci represents one of the nodes in the hierarchy Ci … Example: „ C = PurchaseContext × TemporalContext „ c = (work, weekend), i.e., purchasing something for work on a weekend

Obtaining context a Explicitly specified by the user D E.g., "I want to watch a movie at home with my parents a Observed or deduced by the system D Time( from system clock) Location(from GPS O Deduced from user's behavior(e.g, shopping for business or a How to obtain the context is a separate problem that lies beyond the scope of this tutorial D Significant research literature on obtaining inferring, and predicting context( e.g. for mobile computing) D We assume that the context is given incorporating Context in Recommender Systems A Conceptual Framework

Obtaining Context „ Explicitly specified by the user … E.g., “I want to watch a movie at home with my parents” „ Observed or deduced by the system … Time (from system clock) … Location (from GPS) … Deduced from user’s behavior (e.g., shopping for business or pleasure) … Etc. „ How to obtain the context is a separate problem that lies beyond the scope of this tutorial … Significant research literature on obtaining, inferring, and predicting context (e.g., for mobile computing) … We assume that the context is given Incorporating Context in Recommender Systems: A Conceptual Framework

Traditional recommendation problem Quick Overview Two types of entities: Users and Items Utility of item i for user u is represented by some rating r where r∈ Rating) Each user typically rates a subset of items a Recommender system then tries to estimate the unknown ratings, i. e, to extrapolate rating function R based on the known ratings 口R: Users x Items→ Rating o I.e. two-dimensional recommendation framework a The recommendations to each user are made by offering his/her highest-rated items Rating estimation problem Multitude of existing traditional 2D recommendation techniques They are often classified by D Recommendation approach a Content-based, collaborative filtering, hybrid O Nature of the prediction technique a Heuristic-based. model-based

Traditional Recommendation Problem: Quick Overview „ Two types of entities: Users and Items „ Utility of item i for user u is represented by some rating r (where r∈Rating) „ Each user typically rates a subset of items „ Recommender system then tries to estimate the unknown ratings, i.e., to extrapolate rating function R based on the known ratings: … R: Users × Items → Rating … I.e., two-dimensional recommendation framework „ The recommendations to each user are made by offering his/her highest-rated items Rating Estimation Problem „ Multitude of existing traditional 2D recommendation techniques „ They are often classified by: …Recommendation approach „ Content-based, collaborative filtering, hybrid …Nature of the prediction technique „ Heuristic-based, model-based

Traditional Recommender Systems Content-Based Approaches a Heuristic approaches a Model-based approaches D Item similarity methods o Classification models ang 1995: Pazzani BiNsus, 1997, (Pazzani& Billsus 1997; Mooney Zhang et al. 2002) D Instance-based learning D One-class Naive Bayes classifier(Schwab et al. 2000) 口 Case-based reasoning D Latent-class generative models(Zhang et al. 2002) Traditional Recommender Systems: Collaborative Filtering Approaches a Heuristic approaches a Model-based approaches D Neighborhood methods User-based algorithms(Breese 2008)et al. 2008: Toscherera o D Matrix reduction method D Latent-class generative larity fusion(Wa D User-profile generative a User-based classifiers Matrix reduction methods 2000) erg et al.2001; Sanwaret al D Association rule mining(Lin 口 tem dependency0 et al. 1998: Heckeman et al

Traditional Recommender Systems: Content-Based Approaches „ Heuristic approaches … Item similarity methods (Lang 1995; Pazzani & Billsus, 1997; Zhang et al. 2002) … Instance-based learning (Schwab et al. 2000) … Case-based reasoning (Smyth 2007) „ Model-based approaches … Classification models (Pazzani & Billsus 1997; Mooney & Roy 1998) … One-class Naïve Bayes classifier (Schwab et al. 2000) … Latent-class generative models (Zhang et al. 2002) Traditional Recommender Systems: Collaborative Filtering Approaches „ Heuristic approaches … Neighborhood methods ƒ User-based algorithms (Breese et al. 1998; Resnick et al. 1994; Sarwar et al. 1998) ƒ Item-based algorithms (Deshpande & Karypis 2004; Linden et al. 2003; Sarwar et al. 2001) ƒ Similarity fusion (Wang et al. 2006) ƒ Weighted-majority (Delgado & Ishii 1999) ƒ Matrix reduction methods (SVD, PCA processing) (Goldberg et al. 2001; Sarwar et al. 2000) … Association rule mining (Lin et al. 2002) … Graph-based methods (Aggarwal et al. 1999; Huang et al. 2004, 2007) „ Model-based approaches … Matrix reduction methods (Takacs et al. 2008; Toscher et al. 2008) … Latent-class generative model (Hofmann 2004; Kumar et al. 2001; Jin et al. 2006) … User-profile generative model (Pennock et al. 2000; Yu et al. 2004) … User-based classifiers (Billsus & Pazzani 1999; Pazzani & Billsus 1997) … Item dependency (Bayesian) networks (Breese et al. 1998; Heckerman et al. 2000)