Multi-scale broad collaborative filtering for personalized recommendation

Recently, neighborhood-based collaborative filtering has been increasingly used in personalized recommender systems. However, inevitably, the neighborhood selection is based on a single scale, i.e. selecting a fixed number of the nearest users/items. To solve this problem, we propose a new recommender system called Multi-scale Broad Collaborative Filtering (MBCF). The main contribution lies in designing a multi-scale collaborative vector for capturing the rich information from different numbers of the nearest users/items. However, it is undesirable to input the low-dimensional multiscale collaborative vector directly into the Deep Neural Networks (DNNs), which can easily lead to overfitting. For this reason, instead of DNNs, the Broad Learning System (BLS) is adopted as the mapping function to learn the complex nonlinear relationships between users and items, which can avoid the above problems while obtaining very satisfactory recommendation performance. Extensive experiments on eight benchmark datasets demonstrate the effectiveness of the proposed MBCF algorithm.

Automated summary

Recently, neighborhood-based collaborative filtering has been used more and more in personalized recommender systems. However, the traditional approach of selecting a fixed number of nearest users/items as neighbors has limitations. To address this issue, a new recommender system called Multi-scale Broad Collaborative Filtering (MBCF) is proposed, which captures rich information from different numbers of nearest users/items. Instead of using deep neural networks (DNNs), the Broad Learning System (BLS) is adopted to learn the complex nonlinear relationships between users and items, achieving satisfactory recommendation performance while avoiding overfitting. Extensive experiments on eight benchmark datasets demonstrate the effectiveness of the proposed MBCF algorithm.