TagRec: Temporal-Aware Graph Contrastive Learning with Theoretical Augmentation for Sequential Recommendation (IEEE KDE 2025)
Overview
TagRec is a sequential recommendation method published at IEEE KDE 2025 (IEEE International Conference on Knowledge Discovery). It integrates temporal-aware graph construction with contrastive learning whose augmentations are justified by theory rather than ad-hoc perturbations, targeting robust user-intent modeling from interaction sequences.
- Source: IEEE Xplore abstract page
- Category: Sequential Recommendation
- Datasets: public sequential recommendation benchmarks;
- Metrics: Recall@k, Hit@k, nDCG@k, MRR;
- Baselines: BERT4Rec, SASRec-style Transformers, and other graph/contrastive sequential recommenders;
- BERT4Rec: Sequential Recommendation with Bidirectional Encoder Representations
- EAGER: Two-Stream Generative Recommender with Behavior-Semantic Collaboration
- Efficient On-Device Session-Based Recommendation (ACM TIST)
- How to Index Item IDs for Recommendation Foundation Models (P5, SIGIR)
- LLMCDSR: Enhancing Cross-Domain Sequential Recommendation with Large Language Models
- Mamba4Rec: Efficient Sequential Recommendation with Selective State Space Models
- Researchers: Reproduce core comparisons; check whether statistical significance and compute cost are reported.
- Engineers: Evaluate how the graph encoder and contrastive modules could be plugged into existing ranking stacks, and measure integration cost.
- Product managers: Focus on user-perceivable gains (latency, recommendation freshness, multi-turn consistency) rather than offline nDCG alone.
Background and Motivation
Large-scale recommendation systems face persistent challenges around efficiency, scalability, and user-intent understanding. Traditional pipelines often separate retrieval, ranking, and generation, which limits adaptation to LLM-era expectations of natural-language interaction and multi-hop reasoning. TagRec is proposed in this context to advance the theory and practice of sequence-aware recommendation.
The paper focuses on the core sequential recommendation scenario: modeling user behavior sequences over large item catalogs, where data sparsity and noisy interactions degrade representation quality. Graph contrastive learning with principled (theoretical) augmentation is used to derive stronger self-supervised signals.
Method Sketch
The general pipeline follows the standard four-step pattern:
1. Input and representation — encode user interaction sequences and item context into dense representations, structured as temporal-aware graphs; 2. Core modules — graph encoders with temporal information, contrastive objectives over augmented views; 3. Learning strategy — contrastive learning with augmentations grounded in theoretical analysis (e.g., preserving temporal order and semantic consistency); 4. Inference — standard top-k recommendation over learned item representations.
Note: exact architectural details, loss formulations, and augmentation operators should be verified against the original PDF, as this entry is based on metadata and abstract-level information.
Evaluation Context
Typical evaluation for works in this category includes:
Specific numerical results are not reproduced here; readers should consult the original paper tables before citing quantitative conclusions.
Key Takeaways
1. Architecture: Graph-based contrastive learning remains a strong paradigm for sequence recommendation; temporal awareness helps distinguish short-term intent from long-term preference. 2. Augmentation theory: Heuristic augmentation (random dropout, crop, mask) may destroy critical signals; theory-guided augmentation is an emerging direction. 3. Deployment: Latency budgets, embedding version management, and offline-vs-online metric gaps remain hard constraints for industrial adoption. 4. Open problems: Cross-domain generalization, cold-start robustness, and fairness/causal constraints in sequential recommendation.