Recommender Systems based on biclustering and evolutionary simultaneous learning
Recommender Systems are tools which aid the customized selection of items according to the preferences of users. Movies, music, and e-commerce are examples of well-known applications. Common characteristics of data on these domains are high-dimensionality, due to a large number of users and items, and high sparsity, since each user interacts with only a few items. Obtaining a model that represents the relationship between all users and all items is a challenging task in this setting. The use of co-clustering and learning local models has emerged as a good solution for this problem as proposed in the algorithm SCOAL (Simultaneous Co-Clustering and Learning). However, the number of local models is still a critical hyperparameter that must be set manually. We propose the SCOAL-variant Evolutionary Simultaneous Co-Clustering and Learning (EvoSCOAL) to optimize the search for the best value of this hyperparameter. Our results using sythetic data suggest that EvoSCOAL can surpass the bisection based strategy from literature in both predictive accuracy and hyperparameter estimation capacity. Also, results in real datasets suggests that EvoSCOAL is competitive in predictive accuracy when compared to traditional Recommender Systems approaches.