Collaborative filtering-based group recommender system using sparse autoencoder

Abstract

The development of technology makes the distribution of information easier and faster, but leads to information overload. A recommender system is one tool to overcome information overload, while the collaborative filtering (CF) paradigm is a widely used approach in recommender systems. The recommender system generally focuses on individual recommendations, but in real conditions, recommendations for a group are often needed, for example, when we want to listen to music with friends, or we plan a vacation with family. Many prior studies have used the CF paradigm with matrix factorization to build group recommender systems. Matrix factorization has been shown to alleviate the sparsity problem; however, it does not fully resolve it. Therefore, we propose an approach that uses a sparse autoencoder to address this sparsity issue. We chose the sparse autoencoder because it can effectively capture latent patterns in sparse data by learning a compressed representation while retaining important features crucial for accurate recommendations. We built a group recommender system with three different group sizes and aggregation approaches. For evaluation, we use the root-mean-square error (RMSE) and the mean absolute error (MAE). Test results indicate that the sparse autoencoder outperforms matrix factorization in terms of RMSE and MAE. This study improves group recommender systems by addressing data sparsity using a sparse autoencoder. The proposed approach enhances recommendation accuracy compared to traditional matrix factorization methods.