BiCoRec: Bias-Mitigated Context-Aware Sequential Recommendation Model

Sequential recommendation models aim to learn from users evolving preferences. However, current state-of-the-art models suffer from an inherent popularity bias. This study developed a novel framework, BiCoRec, that adaptively accommodates users changing preferences for popular and niche items. Our approach leverages a co-attention mechanism to obtain a popularity-weighted user sequence representation, facilitating more accurate predictions. We then present a new training scheme that learns from future preferences using a consistency loss function. BiCoRec aimed to improve the recommendation performance of users who preferred niche items. For these users, BiCoRec achieves a 26.00% average improvement in NDCG@10 over state-of-the-art baselines. When ranking the relevant item against the entire collection, BiCoRec achieves NDCG@10 scores of 0.0102, 0.0047, 0.0021, and 0.0005 for the Movies, Fashion, Games and Music datasets.

💡 Research Summary

This paper introduces BiCoRec, a novel Bias-Mitigated Context-Aware Sequential Recommendation model designed to address the pervasive problem of popularity bias in sequential recommendation systems. Popularity bias arises from the power-law distribution of item interactions, where a small number of popular items dominate the dataset, causing models to over-recommend them and under-represent niche items. The authors first formalize and analyze this bias, demonstrating that popular items have substantially more training data (historical sequences) than niche items, leading to poor representation learning for the latter. Furthermore, through empirical analysis, they reveal a critical user behavior pattern: as users’ interaction sequences grow longer, their preference shifts increasingly towards niche items over time. This evolution exacerbates the recommendation challenge, as the model must predict niche items that have sparse historical data.

To tackle these dual challenges, the BiCoRec framework proposes three key innovations. First, it introduces a Popularity-Aware Embedding that explicitly encodes the popularity bias inherent within a user’s interaction sequence. Second, it employs a Co-Attention-Based Weighting mechanism. Unlike standard self-attention that focuses on inter-item relationships, this mechanism performs co-attention between the item embeddings and the popularity embeddings. This allows the model to dynamically re-weight the importance of items in a user’s history based on their popularity context, adaptively amplifying signals from niche items when appropriate for a given user’s evolving taste.

Third, BiCoRec adopts a novel semi-supervised training paradigm called Future Preference Anticipation using Cross-Pseudo Supervision. Instead of merely predicting the next single item, this method trains the model to predict a sequence of future items. It utilizes two identical model branches with different dropout masks. Each branch generates predictions for the sequence, and these predictions are used as pseudo-labels to supervise the other branch via a consistency loss function. This enables the model to learn not only from observed, labeled items but also from unlabeled, padded future positions in the sequence, effectively anticipating and learning from the user’s preference evolution trend identified in the analysis.

The model architecture integrates these components: input embeddings (item, positional, and auxiliary information), stacked layers of SASRec-based self-attention modules enhanced with the co-attention mechanism, and a matrix factorization-based prediction layer. Experiments conducted on four real-world datasets (Movies, Fashion, Games, Music) demonstrate BiCoRec’s superiority over state-of-the-art sequential recommendation baselines. Its most significant achievement is a 26.00% average improvement in NDCG@10 for users who prefer niche items. It also shows an average 3.14% NDCG improvement on datasets with long user sequences. The results validate that BiCoRec successfully mitigates popularity bias and effectively captures the dynamic evolution of user preferences, leading to more accurate and personalized recommendations, especially for niche item consumers.