SA-CAISR: Stage-Adaptive and Conflict-Aware Incremental Sequential Recommendation

TL;DR

SA-CAISR tackles non-stationary user behavior in sequential Recommendation by proposing a buffer-free incremental learning framework that combines Fisher-guided conflict screening with an InfoNCE-based consistency loss. The Fisher mechanism identifies and suppresses conflicting parameters from the historical model, while the InfoNCE term preserves compatible historical signals through contrastive alignment with a filtered reference. The approach updates a Transformer-based SR backbone (SASRec) via a joint objective, enabling rapid adaptation with minimal memory and computation compared to replay-based baselines. Empirical results across four public datasets show state-of-the-art accuracy and substantial efficiency gains, underscoring the method’s practicality for real-time, large-scale dynamic recommendation systems.

Abstract

Sequential recommendation (SR) aims to predict a user's next action by learning from their historical interaction sequences. In real-world applications, these models require periodic updates to adapt to new interactions and evolving user preferences. While incremental learning methods facilitate these updates, they face significant challenges. Replay-based approaches incur high memory and computational costs, and regularization-based methods often struggle to discard outdated or conflicting knowledge. To overcome these challenges, we propose SA-CAISR, a Stage-Adaptive and Conflict-Aware Incremental Sequential Recommendation framework. As a buffer-free framework, SA-CAISR operates using only the old model and new data, directly addressing the high costs of replay-based techniques. SA-CAISR introduces a novel Fisher-weighted knowledge-screening mechanism that dynamically identifies outdated knowledge by estimating parameter-level conflicts between the old model and new data, allowing our approach to selectively remove obsolete knowledge while preserving compatible historical patterns. This dynamic balance between stability and adaptability allows our method to achieve a new state-of-the-art performance in incremental SR. Specifically, SA-CAISR improves Recall@20 by 2.0%, MRR@20 by 1.2%, and NDCG@20 by 1.4% on average across datasets, while reducing memory usage by 97.5% and training time by 46.9% compared to the best baselines. This efficiency allows real-world systems to rapidly update user profiles with minimal computational overhead, ensuring more timely and accurate recommendations.

Figures (4)

• Figure 1: Overall illustration of three representative frameworks for incremental sequential recommendation. (a) Replay-based methods require storing and retrieving historical interaction sequences, leading to additional memory and retrieval overhead. (b) Regularization-based methods constrain parameter updates based on fixed importance from the old model, lacking the ability to adaptively handle outdated or conflicting knowledge. (c) The proposed SA-CAISR framework introduces a Fisher-guided knowledge screening mechanism that selectively filters conflicting parameters, enabling fine-grained knowledge updating while preserving compatible historical information.
• Figure 2: Overall architecture of SA-CAISR. The Current-stage Interactions are processed through two parallel branches: (1) the training branch, which learns new knowledge through cross-entropy loss and (2) the reference branch, which estimates Fisher information from the reference model to compute parameter conflicts and obtain a Compatible Knowledge representation via selective masking. The two branches are aligned using an InfoNCE-based consistency loss, forming a joint objective that balances new knowledge acquisition and preservation of compatible historical knowledge. Fisher-guided selective masking and contrastive regularization together enable stable and efficient incremental adaptation.
• Figure 3: Effect of varying the Top-$K$ similarity threshold in the InfoNCE loss across four datasets. Each subfigure presents six normalized performance curves (MRR@10/20, Recall@10/20, NDCG@10/20) as $K$ increases. All metrics are individually normalized to emphasize their relative variation patterns rather than absolute values.
• Figure 4: Illustrative example of model behavior under knowledge conflict. When new data contradicts historical patterns, ADER and EWC tend to overfit outdated knowledge, leading to biased predictions. In contrast, SA-CAISR effectively identifies and filters conflicting information, enabling adaptive knowledge updating and more accurate recommendations in the new stage.