Goal-driven collaborative filtering

Recommender systems aim to identify interesting items (e.g. movies, books, websites) for a given user, based on their previously expressed preferences. As recommender systems grow in popularity, a notable divergence emerges between research practices and the reality of deployed systems: when recommendation algorithms are designed, they are evaluated in a relatively static context, mainly concerned about a predefined error measure. This approach disregards the fact that a recommender system exists in an environment where there are a number of factors that the system needs to satisfy, some of these factors are dynamic and can only be tackled over time. Thus, this thesis intends to study recommender systems from a goal-oriented point of view, where we define the recommendation goals, their associated measures and build the system accordingly. We first start with the argument that a single fixed measure, which is used to evaluate the system’s performance, might not be able to capture the multidimensional quality of a recommender system. Different contexts require different performance measures. We propose a unified error minimisation framework that flexibly covers various (directional) risk preferences. We then extend this by simultaneously optimising multiple goals, i.e., not only considering the predicted preference scores (e.g. ratings) but also dealing with additional operational or resource related requirements such as the availability, profitability or usefulness of a recommended item. We demonstrate multiple objectives through another example where a number of requirements, namely, diversity, novelty and serendipity are optimised simultaneously. At the end of the thesis, we deal with time-dependent goals. To achieve complex goals such as keeping the recommender model up-to-date over time, we consider a number of external requirements. Generally, these requirements arise from the physical nature of the system, such as available computational resources or available storage space. Modelling such a system over time requires describing the system dynamics as a combination of the underlying recommender model and its users’ behaviour. We propose to solve this problem by applying the principles of Modern Control Theory to construct and maintain a stable and robust recommender system for dynamically evolving environments. The conducted experiments on real datasets demonstrate that all the proposed approaches are able to cope with multiple objectives in various settings. These approaches offer solutions to a variety of scenarios that recommender systems might face.