Collaborative Filtering with a User-Item Matrix Reduction Technique

Collaborative filtering (CF) is regarded as one of the most popular recommendation methods. However, CF has some significant weaknesses, such as problems of sparsity and scalability. Sparsity causes inaccuracy in the formation of neighbors with similar interests, and scalability prevents CF from scaling up with increases in the number of users and/or items. To mitigate these problems, this study proposes a hybrid CF and genetic algorithm (GA) model. GAs are widely believed to be effective on NP-complete global optimization problems, and they can provide good suboptimal solutions in a reasonable amount of time. In this study, the GA searches for relevant users and items from a user-item matrix not only to condense the matrix but also to improve the prediction accuracy. The reduced user-item matrix may reduce the sparsity problem by increasing the likelihood that different customers rate common items. It also shrinks the search space for CF, which ameliorates the scalability problem. Experimental results show that the proposed model improves performance and speed compared to the typical CF model.