Year of Graduation
Boolen Tensor Decompositions in Recommender Systems
Applied Mathematics and Information Science
The problem of extracting the benefits from negative feedback of users in recommender systems is one of the main questions laying the idea of this work. As it is known, standard methods of collaborative filtering consider the task of recommendation as the task of forming a list of the most relevant elements. However, this formulation of the problem does not take into account the converse, e.g. it avoids irrelevant recommendations. Standard algorithms, as well as commonly used valuation metrics, become insensitive to negative feedback because of this omission. To solve this problem, we propose to consider user feedback as a categorical variable and assign it in the form of a tensor with the help of users and evaluated items. For this, we use the third-order tensor factorization method. This method is equally sensitive to the entire spectrum of user ratings (both positive and negative) and is able to accurately predict the relevant elements, even if we have only negative feedback. We also offer a modification of the standard evaluation metrics that help to identify unwanted omissions and take into account the sensitivity to negative feedback. Our model achieves high quality in standard recommender systems and, moreover, it far exceeds other methods of collaborative filtering in cases with negative user feedback or in cases with a lack of information about users.