Personalized neural network-based aggregation function in multi-criteria collaborative filtering

Modeling an effective aggregation function to improve the accuracy of recommendations remains an issue in model-based multi-criteria collaborative filtering (MCCF). The total-based aggregation function is efficient, but it lacks personalization. The user-based aggregation function is personal, but it faces computational and scalability issues. We propose a new personalized neural network-based aggregation function in MCCF to answer the challenges. We enhance the total-based aggregation function by considering not only criteria ratings but also dynamic personal features. The features are extracted from the user's rating history, including rating tendencies and experience. The empirical results of four real-world datasets confirmed the superiority of our method compared to baselines. It achieved the best ROC AUC for all data sets, with an average improvement of 12%. When recommending items with moderate likelihood to users, three out of four datasets witnessed an improvement of up to 6% in the macro F1-score. Additionally, for the most likely items, the macro F1-score increased by up to 12% in all datasets. These findings indicate that our approach has the potential to provide more effective and promising recommendations that are applicable in various domains. Furthermore, it can lead to higher user engagement and satisfaction, which also benefits businesses.