Towards Interpretable Collaborative Filtering Based on Factorization Machines Using the Shapley Value

This master thesis presents the ways of factorization machines usability for the collaborative filtering task to build recommendations of potentially attractive content to users. The factorization machine interpretability is achieved by constructing the Shapley vectors. This approach is widely used in cooperative game theory. It explains how fairly divide the payout between attributes if you consider each attribute as a player in the game and the result of the prediction as a payment. This approach to interpretation allows to track the contributions of attributes to the result and find the interdependence in the data. Furthermore, this paper considers the reduction of feature space. The weights and trained parameters from factorization machines are used as new features in other models. The space reduction allows using computing resources more efficiently by factorizing of the available information. Asymptotic quality estimates provide a comparative analysis of the models. It is shown that the use of additional hidden information from factorization machines leads to an improvement in the quality of recommendations.