An Improved Collaborative Filtering Recommendation Algorithm Based on Retroactive Inhibition Theory

Abstract

Collaborative filtering (CF) is the most classical and widely used recommendation algorithm, which is mainly used to predict user preferences by mining the user’s historical data. CF algorithms can be divided into two main categories: user-based CF and item-based CF, which recommend items based on rating information from similar user profiles (user-based) or recommend items based on the similarity between items (item-based). However, since user’s preferences are not static, it is vital to take into account the changing preferences of users when making recommendations to achieve more accurate recommendations. In recent years, there have been studies using memory as a factor to measure changes in preference and exploring the retention of preference based on the relationship between the forgetting mechanism and time. Nevertheless, according to the theory of memory inhibition, the main factors that cause forgetting are retroactive inhibition and proactive inhibition, not mere evolutions over time. Therefore, our work proposed a method that combines the theory of retroactive inhibition and the traditional item-based CF algorithm (namely, RICF) to accurately explore the evolution of user preferences. Meanwhile, embedding training is introduced to represent the features better and alleviate the problem of data sparsity, and then the item embeddings are clustered to represent the preference points to measure the preference inhibition between different items. Moreover, we conducted experiments on real-world datasets to demonstrate the practicability of the proposed RICF. The experiments show that the RICF algorithm performs better and is more interpretable than the traditional item-based collaborative filtering algorithm, as well as the state-of-art sequential models such as LSTM and GRU.