Recommendation as classification and recommendation as matching: Two information-centered approaches to recommendation

Recommendation can be described as an “information-intensive” task. Before deciding on a doctor to see, a product to buy, or a place to visit, we may feel compelled to consult a variety of sources, which may impact our decision making process. As information proliferates, it becomes important to find ways of representing this information and designing methods that can utilize it effectively. In this thesis, we propose using the availability of information as a guide to formulating a recommendation problem. For instance, when we have user preference data about a set of items, we can formulate recommendation as the task of classifying new instances of data based on learned models of the problem space. When preferences are not available, we will formulate recommendation as the problem of generating queries that match descriptions of items with user interests. In both of these formulations, representation plays an important role since we can only use information if it is represented in a way that is meaningful to our algorithms. Furthermore, we show that there is a common conceptualization of a representational framework for recommendation based on the objects we would like to represent and the information sources we have available to describe them.