A Scalable Recommendation System with Hybrid Similarity Matrix Using Accelerated Particle Swarm Optimization

Collaborative filtering (CF) is one of the most popular recommendation approaches that recommends relevant items to users according to their past interactions to help users with the information overload from the large volume of online multimedia content. However, while the traditional similarity measures using solely rating information suffer from data sparsity and cold-start problems, the features of items are biased and do not reflect the true quality of the items. Existing hybrid neighborhood-based models have mitigated these problems by adopting both the transaction history data and profiles of users and items. Previous works on aggregating rating-based and content-based data into an unified hybrid similarity matrix in neighborhood-based models have shown noticeable success. Correspondingly, this paper proposes adopting a variant of particle swarm optimization to automatically adjust the influence of each data type on the combined similarity matrix, which can boost the performance of neighborhood-based systems significantly by at least 1.64% in terms of predicting accuracy and 0.77% in the ranking task compared to the earlier related hybrid methods. Additionally, to deal with the huge timing disadvantage of this optimization algorithm, we propose a tweak to the particle swarm optimization to train the model only on a small batch of the training set, utilizing the existing information included in distinct similarity matrices.