Abstract

The world of video games has changed considerably over the recent years. Its diversification has dramatically increased the number of users engaged in online communities of this entertainment area, and consequently, the number and types of games available. This context of information overload underpins the development of recommender systems that could leverage the information that the video game platforms collect, hence following the trend of new games coming out every year. In this work we test the potential of state-of-the-art recommender models based respectively on Factorization Machines (FM), deep neural networks (DeepNN) and one derived from the mixture of both (DeepFM), chosen for their potential of receiving multiple inputs as well as different types of input variables. We evaluate our results measuring the ranking accuracy of the recommendation and the diversity/novelty of a recommendation list. All the algorithms achieve better results than a baseline based on implicit feedback (Alternating Least Squares model). The best performing algorithm is DeepNN, the high order interactions are more important than the low order ones for this recommendation task. We also analyze the effect of the sentiment extracted directly from game reviews, and find that it is not as relevant for recommendation as one might expect. We are the first in studying the aforementioned recommender systems over the context of online video game platforms, reporting novel results which could be used as baseline in future works.