Introduction
In the rapidly evolving landscape of e-commerce and social media, recommendation engines must operate in dynamic environments where user interests shift and new items frequently emerge. Traditional "static" models, trained once on a massive dataset, quickly become obsolete. To maintain relevance, models require Continual Learning (CL)—the ability to update periodically with new data. However, simple fine-tuning on the latest data batch often leads to Catastrophic Forgetting, where the model loses patterns learned from historical interactions. While retraining on all historical data (the "Joint" method) is a potential solution, its computational and storage costs are unsustainable at scale.
The ADER Framework
The authors introduce ADER (Adaptively Distilled Exemplar Replay), a framework designed to balance the acquisition of new knowledge with the preservation of past insights through two primary mechanisms:
1. Smart Exemplar Replay (Herding)
Rather than storing the entire history, ADER maintains a compact, high-quality subset of past sessions known as "exemplars." The selection process is governed by:
- Frequency-based sizing: The number of saved samples for an item is proportional to its popularity.
- Herding Technique: An iterative algorithm selects sessions closest to the "mean feature vector" of an item, ensuring that the saved samples are statistically representative of the item's interaction patterns.
2. Adaptive Distillation Loss
During updates, ADER employs Knowledge Distillation on these exemplars. Instead of merely retraining on the correct labels, the current model is forced to match the output distribution of the previous model version. This preserves the "nuance" of prior knowledge. The weight of this distillation loss ($\lambda_t$) is adaptive, automatically adjusting based on the ratio of new to old items and the size of the incoming data batch.
Experimental Results
ADER was evaluated using two industry-standard datasets: DIGINETICA (5 months of e-commerce data) and YOOCHOOSE (6 months of clickstream data). The performance was tracked across 16 update cycles to simulate real-world streaming environments.
Performance and Stability
As shown in the performance curves, ADER demonstrates remarkable stability and consistently outperforms baselines like simple Fine-tuning, Dropout, and Elastic Weight Consolidation (EWC). While other methods often exhibit "see-saw" performance as they struggle with new data, ADER maintains a steady upward or stable trend across cycles.
The "Joint" Paradox
One of the most significant findings is that ADER can outperform the "Joint" method—traditionally considered the performance upper bound. On the DIGINETICA dataset, ADER achieved a Recall@20 of 50.21%, surpassing the 50.03% achieved by retraining on the entire history from scratch. This suggests that replaying a curated, high-quality set of exemplars can be more effective than training on a massive, potentially noisy historical log.
Conclusion
ADER provides a robust solution for session-based recommendation in streaming environments. By combining representative exemplar replay with adaptive knowledge distillation, it effectively mitigates catastrophic forgetting while remaining computationally efficient. The framework's ability to match or exceed full retraining performance makes it a highly viable candidate for production-scale recommendation systems.
Summary updated