Kai Widdeson

FFM-SVD: A Novel Approach for Personality-aware Recommender Systems

Abstract

This paper addresses and evaluates various approaches to incorporating personality data into a recommender system to improve the accuracy of recommendations. Automatic personality recognition is enabled by the LIWC dictionary and appropriate data balancing and pre-processing has been conducted. Personality-aware pre-filtering techniques are discussed and evaluated using the RPE metric developed in this project, with non-targeted stratified personality sampling performing best. The effectiveness of introducing personality to recommenders in these domains is evaluated against the MAP correlation score which is introduced as a predictor. This suggested that media domains are most susceptible to improvement with personality data, however, experimental results contend this. For example, the Movies domain was predicted to be the most influenced domain, when in actuality the Sports domain showed the greatest improvement. A novel personality-specific model, FFM-SVD, is proposed and shown to outperform alternative models in prediction accuracy. LightGBM, a model currently unseen in personality-aware recommenders, and a neural network recommender are also created and assessed. Real-world implementation details, ethical considerations, and future developments are also explored.