AN INTERPRETABLE ATTRIBUTE WEIGHT DRIVEN RECOMMENDATION MODEL USING BIJECTION-BASED LEARNING

Abstract

Recommendation systems based on collaborative filtering and matrix factorization have demonstrated strong predictive performance but often struggle with data sparsity, cold-start scenarios, high computational cost, and limited interpretability. While hybrid and deep learning–based approaches address some of these issues, they typically rely on iterative optimization and latent representations, which obscure transparent reasoning behind recommendations. This paper proposes a Bijection-Based Attribute Weight Ranking Model (BAWRM), a lightweight and interpretable recommendation framework that operates directly on explicit user and item attributes. The model leverages observed user–item interactions as bijective pairs and employs a deterministic, non-iterative learning process to compute attribute weights at the rank level. To enhance robustness under sparse data conditions, a smoothing mechanism is incorporated, and vital attributes are selected using rank-based correlation analysis. Unlike latent factor models, BAWRM avoids abstract hidden representations and provides clear attribute-level explanations for predicted rankings.The proposed approach is evaluated on the MovieLens 100k dataset, a widely used benchmark characterized by high sparsity. Experimental results demonstrate that BAWRM achieves competitive rating prediction accuracy, as measured by RMSE and MAE, while delivering strong ranking performance in terms of Precision@10 and NDCG@10. In addition, analytical comparisons show that the model offers lower computational complexity, reduced memory requirements, improved interpretability, and effective handling of cold-start scenarios compared to traditional matrix factorization methods.

Overall, this work highlights that meaningful and efficient recommendation performance can be achieved through explicit attribute-driven, rank-level modeling, offering a practical alternative to computationally intensive latent-factor-based recommender systems.