Lightweight Similarity-Based Approach for Reducing User Interaction Data with Matrix Factorization in a Recommendation System

The accumulation of large-scale data in recommendation systems can significantly increase resource consumption during model training. In particular, when redundant data exists, the model may overfit as it evaluates test data based on already seen, similar data during the training process. To address the issues of resource consumption and overfitting, this study proposes a lightweight similarity-based algorithm to reduce interaction data. The proposed algorithm decomposes interaction data using matrix factorization to derive a user matrix and then calculates the similarity between users using cosine similarity. The data is then reduced by removing the data of users with fewer items among pairs of users whose similarity exceeds a certain threshold. Experiments were conducted to analyze the impact of the proposed algorithm on recommendation performance and training time after data reduction. Furthermore, the algorithm’s effectiveness was evaluated through performance comparisons with existing data reduction methods. The primary contribution of this study is the introduction of a lightweight similarity-based approach that focuses on eliminating redundant user data in recommendation systems, thereby preventing overfitting and minimizing resource consumption.