Matrix Factorisation for Movie Recommender Systems: Enhancing Collaborative Filtering with Side Information through Evaluating Baseline, Joint, and Collective Matrix Factorisation

Abstract

In an era where online content is constantly increasing, particularly in the movie domain, people get overwhelmed by the overload of the available choices. One of the most prominent tools that have been used to overcome such a challenge are movie recommender systems, as they provide users with personalised suggestions tailored to their preferences. Movie recommender systems follow three main filtering methods: collaborative filtering, content-based filtering and hybrid filtering. This dissertation focuses on investigating matrix factorisation techniques for collabora- tive filtering, comparing three complementary approaches: baseline Matrix Factorisation (MF), Joint Matrix Factorisation (JMF), and Collective Matrix Factorisation (CMF) to evaluate their performances on both warm and cold start scenarios. The core concept of this dissertation is enhancing collaborative filtering through incorporating side information with matrix factorisa- tion and observe the effect on the models’ prediction accuracy as well as the recommendation quality. Using the MovieLens 1M dataset, three movie recommendation models were developed and evaluated in terms of prediction accuracy and handling the cold start problem. Although col- laborative filtering is widely used in movie recommender systems, it presents major challenges which are the high sparsity of the user-item matrices and the cold-start problem. The baseline MF model was applied using Singular Value Decomposition (SVD). The Joint MF model was also applied using SVD, leveraging the demographic side information by combining them into the rating matrix. The last approach was the Collective MF, using the cmfrec package [2] to simultaneously factorise ratings and demographics matrices. The evaluation measures followed include both the rating prediction metrics; mean squared error (MSE) and mean absolute error (MAE), and top-N recommendation metrics including Precision@N and Recall@N. The cold start problem was addressed by varying proportions of observed versus missing values and across multiple values of latent factors. Results revealed that the baseline MF model scored competitive accuracy, however, JMF model outperformed both baseline MF and CMF models and showed improved prediction accuracy in both warm and cold start scenarios, highlighting the importance of integrating side information with the latent factor models. The CMF model, although scored better than the baseline MF, returned mixed results, indicating the complexity of the model and the need for more tuning.