Context Matters

The view we have of recommender systems is that of two-dimensional systems (users x items) whose main goal is to `recommend items to users’. However, as well illustrated in this paper, “decision making is contingent upon the context of decision making; the same consumer may [...] prefer different products or brands under different contexts”. For example, I (the user) may want to be recommended different restaurants (the item), depending on when I am going (the context), with whom I am going (the context, again), and for what purpose (the context, yet again).

The authors then suggest to incorporate contextual information into recommender systems, thus transforming them into multi-dimensional (MD) systems. Now the 2 big questions are: how do we distinguish relevant contextual information from irrelevant one, and how do we use relevant info to produce recommendations?

The authors suggest an initial solution to the problem whereby dimensions are predefined (e.g., user, movies, place, time, companion) and hierarchically structured. Moreover, each dimension is described by pre-defined attributes (e.g., for companion, could be alone, friend, partner, family, co-worker, other). Based on these assumptions, feature-selection techniques are used to identify the “relevant” dimensions; a reduction-based approach in then used to map the MD problem into a traditional 2D one. Their initial evaluation, although limited, proves that context does matter!

The paper is particularly interesting as it points to a new direction of research in recommender systems. Besides the open questions that the authors suggest, there are a few others: for example, taxonomies are very difficult to create and maintain; can we extract dimensions/attributes from the reviews written by users? In so doing, we depart from a single multi-dimensional space (N context dimensions for all users) to a highly heterogeneous one (different dimensions for different users): with what impact on sparsity and accuracy? How easy can a mapping, and subsequent reduction, be performed?

Besides difficulties, there are many new possibilities: rather than simply suggesting to a user some top K individual items, recommendations across more than 1 dimension can be produced, and for groups of items at the same time. Evaluating these systems is not at all easy, due to the lack of suitable datasets, but projects like IYOUIT may soon come to the rescue! :)

3 Responses to “Context Matters”

  1. Surely multi-dimensional recommender systems will be useful for “mobile search” which differs from the Web search model – web users do research and are looking for exhaustive answers, whereas mobile users will be looking for quick answers.

  2. Neal Lathia says:

    i think what this paper is saying (from your nice summary) is that recommender system context is not fine grained enough: remember that recommender systems do not work across contexts (music -> movies)!

  3. Licia Capra says:

    uhm, i would agree when you say “rec sys context is not fine grained enough” (i.e., considering only 2 dimensions, users and items, is not enouogh; what they say, and i agree, is that time, space, company, and a number of other dimensions may be very relevant – going back to what daniele said, some of these are particularly important when going mobile). but that’s different from the movies/music distintion, which could be seen as referring to the SAME dimension, that is, the item. note that, in the paper, they say that, within the same dimension (e.g., item), there could be taxonomies (e.g., entertainment –> movies / music / theatre –> horror / comedy …), and that you could do some reasoning about this too, but that’s a separate thing. i find the first aspect (multiple, distinct dimensions) very interesting and worth exploring :)