Mitigating Consistency Bias in Recommender Systems via Transformer-Based Dynamic Fuzzy Bias Adjustment

Abstract

: Recommender systems provide personalized services by learning from users' historical interactions, but recorded feedback may combine intrinsic preference with external influence and persistent rating habits. This mixture can produce consistency bias, in which users remain close to historical item evaluations or their own previous scoring patterns even when current interests have changed. To address this problem, this paper proposes a consistency-bias mitigation framework that combines Transformer-based sequential modeling with dynamic fuzzy bias adjustment. A Transformer encoder first captures long-term dependencies and short-term changes in users' interaction sequences. The model then measures rating variation between adjacent temporal stages, adaptively refines temporal windows, and converts the variation signal into continuous fuzzy weights. The adjusted historical representation is fused with the sequential interest representation for candidate-item prediction. Experiments on MovieLens100K, Amazon Beauty, and Amazon Toys compare the framework with collaborative filtering, neural recommendation, recurrent sequential recommendation, and Transformer-based baselines. Macro-level comparisons, dataset-wise analysis, ablation tests, robustness experiments, parameter sensitivity, and a case study show that the proposed method improves Recall and NDCG while degrading more slowly when biased interactions are introduced. These results suggest that temporal change detection and adaptive weighting can reduce the excessive influence of repeated historical patterns without discarding useful behavioral evidence.