[Paper Review] Long Short-Term Temporal Meta-learning in Online Recommendation

Ruobing Xie, Yalong Wang+5 more|arXiv (Cornell University)|May 8, 2021

Recommender Systems and Techniques4 citations

TL;DR

This paper proposes Long Short-Term Temporal Meta-learning (LSTTM), a novel framework for online recommendation that jointly models users' long-term preferences via a global multi-source behavior graph and short-term preferences via an internal temporal graph using GAT-based aggregators. It employs a temporal MAML strategy with asynchronous optimization to enable fast adaptation to real-time trends, achieving significant gains in both offline and online evaluations on WeChat Top Stories.

ABSTRACT

An effective online recommendation system should jointly capture users' long-term and short-term preferences in both users' internal behaviors (from the target recommendation task) and external behaviors (from other tasks). However, it is extremely challenging to conduct fast adaptations to real-time new trends while making full use of all historical behaviors in large-scale systems, due to the real-world limitations in real-time training efficiency and external behavior acquisition. To address these practical challenges, we propose a novel Long Short-Term Temporal Meta-learning framework (LSTTM) for online recommendation. It arranges user multi-source behaviors in a global long-term graph and an internal short-term graph, and conducts different GAT-based aggregators and training strategies to learn user short-term and long-term preferences separately. To timely capture users' real-time interests, we propose a temporal meta-learning method based on MAML under an asynchronous optimization strategy for fast adaptation, which regards recommendations at different time periods as different tasks. In experiments, LSTTM achieves significant improvements on both offline and online evaluations. It has been deployed on a widely-used online recommendation system named WeChat Top Stories, affecting millions of users.

Motivation & Objective

• Address the challenge of jointly modeling users' long-term and short-term preferences in large-scale online recommendation systems.
• Overcome the inefficiency and asynchrony of internal and external user behaviors across multiple applications.
• Enable fast online adaptation to real-time trends without costly retraining or fine-tuning.
• Balance effectiveness and efficiency in real-world deployment by minimizing online computation costs.

Proposed method

• Construct a global long-term graph integrating multi-source internal and external user behaviors to model stable, long-term user preferences.
• Build an internal short-term graph focused on recent, task-specific behaviors to capture dynamic, short-term interests.
• Apply Graph Attention Networks (GATs) with distinct neighbor sampling strategies to learn representations from both graphs.
• Use a gating mechanism to fuse short-term and long-term representations adaptively based on context.
• Design a temporal meta-learning strategy based on MAML, treating recommendations across different time periods as separate tasks for fast adaptation.
• Implement an asynchronous optimization scheme to decouple training and inference, reducing online inference latency.

Experimental results

Research questions

• RQ1 How can a recommendation system effectively model both long-term and short-term user preferences from heterogeneous internal and external behaviors?
• RQ2 Can temporal meta-learning with MAML enable fast adaptation to new trends in online recommendation with minimal computational cost?
• RQ3 How do external behaviors from other applications contribute to improving recommendation performance, especially in cold-start scenarios?
• RQ4 What is the impact of architectural components such as GAT-based aggregation and multi-hop neighbor-similarity loss on model performance?

Key findings

• LSTTM achieved a 9.60% relative improvement in Click-Through Rate (CTR) and a 9.93% improvement in Average Click Number (ACN) in online A/B tests on WeChat Top Stories.
• The model improved HCR by 3.42% and average dwell time by 2.52%, indicating enhanced user engagement and satisfaction.
• The temporal MAML component significantly improved performance during periods with high new-topic volatility, such as periods 2 and 3.
• Removing the global long-term graph (external behaviors) led to a measurable drop in performance, confirming the value of multi-source behavior integration.
• The gating-based fusion mechanism outperformed simple concatenation, enabling more personalized and effective representation combination.
• The ablation study confirmed that GAT-based aggregation and the multi-hop neighbor-similarity loss were critical for capturing diverse and evolving user preferences.