Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation

Zhiqiang Wang, Jiayi Pan, Xingwang Zhao, Jianqing Liang, Chenjiao Feng, Kaixuan Yao

Abstract

Sequential recommendation systems aim to predict future user interactions based on historical data. However, these systems frequently encounter the "cold-start" problem, where new users with minimal interaction history receive poor recommendations. This paper presents Counterfactual Task-augmented Meta-learning (CTM), a novel framework designed to bridge the gap between regular and cold-start users. CTM utilizes Counterfactual Sequence Augmentation (CSA) to simulate diverse user behaviors and Meta-path Neighbor Aggregation (MNA) to enrich item representations via graph-based structural knowledge. Experimental results across three benchmark datasets demonstrate that CTM consistently outperforms state-of-the-art baselines, particularly in high-density scenarios.

Introduction: The Optimization Divergence Problem

Sequential recommendation models typically rely on deep interaction histories to learn user preferences. While meta-learning has been proposed as a solution—learning a global initialization that can be quickly adapted to new tasks—it often suffers from a fundamental flaw: optimization divergence.

Optimization divergence between regular and cold-start users.

As illustrated in Figure 1, the gradient update directions for regular users (with abundant data) often differ significantly from the ideal update directions for cold-start users. This bias toward active users prevents the model from generalizing effectively to newcomers, leading to suboptimal performance for the very users who need the most assistance.

The CTM Framework

To address these challenges, the CTM framework introduces three core components that work in tandem to improve model robustness and representation quality.

Overview of the CTM framework, including meta-training with CSA and MNA, and meta-testing.

1. Counterfactual Sequence Augmentation (CSA)

Standard models only learn from observed sequences. CSA introduces a "what if" reasoning capability by intervening in real interaction sequences to generate counterfactual versions—simulated paths a user might have taken. By expanding the hypothesis space, the model practices on a wider variety of potential behaviors, making it more resilient to the unpredictable nature of cold-start sequences.

2. Meta-path Neighbor Aggregation (MNA)

When a user has only a few interactions, sequential patterns are insufficient. MNA compensates for this sparsity by leveraging the global user-item graph to find "meta-path neighbors." For instance, if two users share a common interaction, their other interactions become neighbors. A self-attention mechanism aggregates these neighbors to enrich item representations, uncovering latent preferences that are not explicitly present in the short local sequence.

3. Dual-Loop Meta-learning Optimization

CTM employs a dual-loop strategy. The inner loop (local update) adapts the model to specific user tasks using a combination of real and counterfactual data. The outer loop (global update) then optimizes the meta-parameters to minimize the total loss across all tasks. This ensures the initialization point is well-suited for rapid adaptation to diverse user profiles.

Experimental Evaluation

The authors evaluated CTM on three diverse datasets: Electronics (Amazon), MovieLens 100k, and KuaiRec (a high-density video platform dataset).

Key Results

  • Superior Performance: CTM consistently outperformed baselines like Meta TL and MeLU. On the high-density KuaiRec dataset, CTM achieved an MRR of 0.520, a significant jump from the best baseline at 0.371.
  • Ablation Success: Removing either the CSA or MNA modules led to a significant drop in performance, confirming that both behavioral simulation and structural knowledge are critical to the model's success.
  • Optimal Intervention: Parameter analysis revealed that replacing approximately 50% of a sequence with counterfactual items provided the best balance between maintaining sequence integrity and introducing enough variety for effective learning.

Conclusion and Future Work

CTM effectively mitigates the cold-start problem by addressing optimization divergence through counterfactual reasoning and graph-based neighbor aggregation. While the framework shows great promise, future work will focus on more intelligent counterfactual generation methods and reducing the computational overhead associated with meta-path aggregation in industrial-scale systems.

Summary updated

Abstract

Cold-start sequential recommendation, where user interaction histories are sparse or minimal, remains a significant challenge in recommendation systems. Current meta-learning-based approaches rely heavily on the interaction histories of regular users to construct meta-tasks, aiming to acquire prior knowledge for cold-start adaptation. However, these methods often fail to account for preference discrepancies between regular and cold-start users, leading to biased preference modeling and suboptimal recommendations. To address this issue, we propose a novel counterfactual task-augmented meta-learning method for cold-start sequential recommendations. Our approach intervenes in user interaction histories to create counterfactual sequences that simulate potential but unrealized user behaviors, establishing counterfactual tasks within a meta-learning framework. Additionally, we aggregate meta-path neighbors to uncover latent relationships between items, enabling more detailed and accurate modeling of user preferences. Moreover, by integrating real and counterfactual task losses, we jointly optimize the model through a combination of global and local updates, enhancing its adaptability to cold-start scenarios. Extensive experiments demonstrate that our method significantly outperforms existing state-of-the-art techniques, achieving superior results in cold-start sequential recommendation tasks.
BibTeX
@inproceedings{wang2025counterfactual,
  author = {Zhiqiang Wang and Jiayi Pan and Xingwang Zhao and Jianqing Liang and Chenjiao Feng and Kaixuan Yao},
  title = {{Counterfactual Task-augmented Meta-learning for Cold-start Sequential Recommendation}},
  booktitle = {AAAI Conference on Artificial Intelligence},
  year = {2025},
  doi = {10.1609/aaai.v39i12.33396},
}
Figure