Negative Examples Sampling Based on Factorization Machines for OCCF

Abstract

Some applications of recommender system such as news recommendation and music recommendation, are considered as One-Class Collaborative Filtering (OCCF). In OCCF, dataset for training consists only of binary data that reflects whether the user has behavior on the item. In other words, we can only observe user's positive examples through implicit feedback in the system(such as click, collect). As a result of that, ambiguity arises when interpreting non-positive examples. There are two main solutions to solve this problem: one is to find a algorithm suitable for OCCF, and the other is to perform negative examples sampling algorithm to balance the positive and negative samples and convert OCCF into traditional collaborative filtering problem. In this paper, we apply the factorization machines algorithm to negative examples sampling. Experimental results show that the method we proposed effectively solves the problem that OCCF has no negative examples, and this method is better than other basic negative examples sampling methods.