TLDR: MEGG is a new incremental learning method for neural recommendation models that uses a novel “GGscore” to identify and replay the most influential historical data samples. This approach effectively prevents models from forgetting old user preferences while continuously learning new ones, leading to consistently high recommendation performance in dynamic real-world environments.
In the fast-paced world of online platforms, where user preferences and trends are constantly shifting, recommendation systems play a crucial role. From suggesting movies and music to products on e-commerce sites, these systems aim to provide personalized experiences. However, a significant challenge arises because the underlying models, often based on Neural Collaborative Filtering (NCF), are traditionally built on static datasets. This means they struggle to adapt when user behaviors and data distributions evolve over time, leading to a problem known as ‘catastrophic forgetting’ – where learning new information causes the model to forget previously learned knowledge.
To tackle this, researchers have been exploring incremental learning, a paradigm designed to allow models to continuously acquire new knowledge without losing old information. While incremental learning has seen success in areas like computer vision and natural language processing, its direct application to recommender systems is hindered by unique challenges, including the distinct nature of recommendation tasks, complex data, and sparsity issues. Existing incremental learning methods for neural recommendation models are also scarce and often lack broad applicability.
A new research paper, MEGG: Replay via Maximally Extreme GGscore in Incremental Learning for Neural Recommendation Models, introduces an innovative solution to this problem. Authored by Yunxiao Shi, Shuo Yang, Haimin Zhang, Li Wang, Yongze Wang, Qiang Wu, and Min Xu, the paper proposes an experience replay-based incremental learning framework called MEGG (Replay Samples with Maximally Extreme GGscore).
The Core Idea: GGscore
At the heart of MEGG is a novel metric called the GGscore. This score quantifies the influence of individual data samples on the model’s training process. By understanding which samples are most impactful, MEGG can strategically select and ‘replay’ these crucial historical data points. This selective replaying is key to mitigating catastrophic forgetting, ensuring that the recommendation model maintains its high predictive performance over extended periods, even as new data continuously streams in.
One of the key advantages of MEGG is its data-centric nature. This means it operates independently of the specific neural network architecture used for recommendations. This flexibility ensures that MEGG can be broadly applied across various neural recommendation models and seamlessly integrated with existing incremental learning frameworks to further boost their performance.
How MEGG Works
The MEGG framework operates by periodically replaying a subset of historical data. When new data arrives, the model is trained on a combination of this new information and carefully selected past data. The selection process is where the GGscore comes into play. The GGscore is derived from a concept called ‘One Step Loss Change,’ which essentially measures how much a single user-item interaction sample affects the model’s training loss during one optimization step. To make this computationally efficient for large-scale recommendation systems, MEGG only calculates gradients for a subset of model parameters, specifically those related to user and item embeddings, and the final prediction layer.
The method then identifies samples with the ‘maximally extreme’ GGscores – meaning those with the highest and lowest influence – and prioritizes them for replay. This ensures that the most critical historical information, whether it strongly supports or challenges current model understanding, is retained and revisited.
Demonstrated Effectiveness
The researchers conducted extensive experiments using three popular neural recommendation models (Wide & Deep, DCN, and NFM) across four benchmark datasets: MovieLens 1M, Douban Movie, Taobao2014, and LastFM-1k. The results were compelling: MEGG consistently outperformed state-of-the-art incremental learning methods. In many cases, MEGG achieved recommendation performance nearly identical to, and sometimes even surpassed, training the model from scratch on the entire dataset (Full-Batch training), but with significantly reduced storage and computational overhead.
Furthermore, MEGG demonstrated its expandability by enhancing the performance of other incremental learning methods when combined with them. The method also proved to be efficient, with its computational overhead for sampling remaining manageable even as model sizes increased, especially for typical embedding dimensions used in industrial applications. Its robustness was also confirmed through various hyper-parameter analyses and different replay ratios.
Also Read:
- Solving Diversity Collapse in LLMs with Diversity-Preserving Hybrid RL
- Enhancing LLM Training Stability with ∆LNormalization for Variable Response Lengths
Looking Ahead
MEGG represents a significant step forward in developing adaptive and robust recommendation systems. By intelligently managing historical data through its novel GGscore and experience replay mechanism, it effectively addresses the critical challenge of catastrophic forgetting. This allows online dynamic recommender systems to continuously learn and adapt to evolving user preferences, ensuring consistently high recommendation performance in real-world scenarios. The implementation of MEGG is planned to be made publicly available, paving the way for its broader adoption and further research in the field.
