Influence of Auxiliary Features in Factorization-based Collaborative Filtering

The tremendous growth of the Internet brings with it a massive amount of data that users are exposed to on a daily basis. Consequently, information filtering techniques like recommender systems have become increasingly important to sift through the data and find what is relevant to a particular user. A recent approach for recommender systems, called Factorization Machines (FM), has been attracting lots of attention for its flexibility and accuracy. In this work we explore the importance of optimizing FMs, the influence of different auxiliary features on the performance of FMs and how to weight these features. The results show that the optimization of hyper-parameters, factorization dimensionality and feature weights is a very crucial part of using FMs for recommendations. Using static un-normalized feature weights is found to outperform using previously proposed normalized weights in terms of rating prediction accuracy. Several new approaches to auxiliary features are proposed. Expanding feature vectors with info based on the ratio of users that found a review of the item helpful or not is found to improve performance slightly. Additionally, constructing feature vectors using the similarity of users (based on either common movies commented on, common friends or common ratings) was also beneficial to the performance. However, the overall conclusion is that there is no clear answer to which auxiliary features are useful, as it depends heavily on the dataset and in many cases auxiliary features seem to introduce noise into the model. A noticeable trend in the results indicates that item attributes are more useful for rating predictions than user attributes, while the opposite applies for top-N ranking predictions and diversity, which benefits more from user attributes.Electrical Engineering, Mathematics and Computer ScienceMultimedia ComputingEmbedded System