Evaluating Vector Representations from User's Reviews in a Recommendation Task

Abstract

The recommendation task is a prominent and challenging area of study in Machine Learning. It aims to recommend items such as products, movies, and services to users according to what they have liked in the past. In general, most of the recommendation systems only consider structured information. For instance, in recommending movies to users they might use features such as genre, actors, and directors. However, unstructured data such as users' reviews may also be considered, since they can aggregate important information to the recommendation process, improving the performance of recommendation systems. Thus, in this work, we evaluate the use of text mining methods to extract and represent relevant information about user reviews, which were used alongside with rating data, as input of a content-based recommendation algorithm. We considered three different strategies for this purpose, which were: Topics, Embeddings and Relevant Embeddings. We hypothesized that using the considered strategies, it is possible to create more meaningful and concise representations than the traditional bag-of-words model, and yet, increase the performance of recommendation systems. In our experimental evaluation, we confirmed such a hypothesis, showing that the considered representations strategies are indeed very promising for representing user reviews in the recommendation process.