The availability of additional contextual information can have a significant impact on recommendations. In this chapters, we discuss algorithms to provide context-dependent recommendations to users. We describe some example scenarios in the following.
• Related-item recommendation: Recommending items (such as news articles) that are related to the one that the user is interacting with is useful in many applications. In this case, the item being interacted with provides the context. For instance, when a user is reading a news article or viewing a product on an e-commerce site, it is useful to recommend other news articles or products related to the one that the user is currently reading or viewing.
• Multicategory recommendation: Many websites organize their items using human-understandable categories and recommend top items for each category. Here the categories provide context. An item may be classified into multiple categories, and it is desirable for recommendations within a category to be both semantically relevant and personalized.
• Location-dependent recommendation: In several applications, geographical location is an important context, and it is desirable to provide recommendations that are germane to a user's current location.
• Multiapplication recommendation: A recommendation system may serve multiple applications, for example, modules on a website and apps for different devices. Because different applications have varying screen sizes, layouts, and different ways to present items, it is important to tailor recommendations to incorporate application specific user behavior. Here each application provides context.
If personalized recommendation is not required in a given context, the models presented in Chapters 7 and 8 can be easily modified to provide context-dependent recommendation. For example, the RLFM model described in Chapter 8 can be modified for related-item recommendation, as follows. Let yjk denote the response that a user would give to item j in context k (e.g., reading context article k in related-news recommendation). Then, we predict yjk by
• b is a regression function based on feature vector xjk that characterizes (item j, context k) pair (e.g., the similarity between the two bags of words for articles j and k in related-news recommendation)
• αk is the bias of context k
• βj is the popularity of item j
• uk and vj are the two latent factor vectors for context k and item j Note that,