Semantic Retrieval Augmented Contrastive Learning for Sequential Recommendation

Conference: NeurIPS 2025 arXiv: 2503.04162 Code: GitHub Area: Recommender Systems Keywords: Sequential Recommendation, Contrastive Learning, Large Language Models, Semantic Retrieval, Data Augmentation

TL;DR

This paper proposes SRA-CL, a framework that leverages the semantic understanding capabilities of LLMs to construct high-quality contrastive sample pairs. By combining semantic retrieval with a learnable sample synthesizer, SRA-CL enhances contrastive learning for sequential recommendation and achieves state-of-the-art performance across four datasets in a plug-and-play manner.

Background & Motivation

Contrastive learning has been widely adopted in sequential recommendation to alleviate data sparsity. However, existing methods exhibit two critical limitations when constructing positive sample pairs:

Semantic Bias: Random augmentation strategies (e.g., masking, dropout) may fundamentally distort the user preference semantics encoded in a sequence; clustering-based methods leveraging collaborative signals (e.g., K-means) are constrained by sparse ID signals, resulting in imprecise cluster assignments.

Non-learnability: Existing methods rely on predefined hard rules (same-cluster pairing, shared next-item pairing) and cannot allow the model to autonomously learn optimal contrastive sample construction strategies.

The key insight of this work is that textual semantic information (categories, brands, descriptions) is inherently stable and unaffected by data volume or training dynamics, making it a more reliable source for contrastive samples. Furthermore, LLMs possess strong semantic understanding and reasoning capabilities, making them well-suited for generating semantic embeddings that capture user preferences and item characteristics.

Method

Overall Architecture

SRA-CL is a model-agnostic plugin framework comprising three modules: (1) cross-user contrastive learning via user-level semantic retrieval; (2) intra-user contrastive learning via item-level semantic retrieval; and (3) the main recommendation task. The three components are jointly trained via a weighted loss: \(\mathcal{L} = \mathcal{L}_{\text{Rec}} + \alpha \mathcal{L}_{\text{CS}} + \beta \mathcal{L}_{\text{IS}}\). At inference time, only the recommendation backbone is used, incurring no additional LLM overhead.

Key Designs

1. LLM-based User Preference Semantic Embedding: The user interaction sequence (item attributes and descriptions sorted chronologically) is formatted into a prompt \(\mathcal{P}_u\) and fed into an LLM (DeepSeek-V3) to infer the user preference summary \(\mathcal{A}_u = \text{LLM}(\mathcal{P}_u)\). A pre-trained text embedding model (SimCSE-RoBERTa) then encodes this summary into a fixed semantic vector \(\tilde{\mathbf{h}}_u\). Top-\(k\) semantically similar users are retrieved via cosine similarity to form a candidate pool \(\mathcal{N}_u\). All semantic embeddings are frozen throughout training.

2. Learnable Contrastive Sample Synthesizer: Rather than directly selecting a user from the candidate pool as a positive sample (which yields suboptimal results under hard rules), an attention mechanism computes the suitability score of each candidate: \(p_{u,u'} = \text{softmax}(\text{LeakyReLU}(\mathbf{a}^\top[\mathbf{W}\tilde{\mathbf{h}}_u \| \mathbf{W}\tilde{\mathbf{h}}_{u'}]))\). The positive representation is then obtained by weighted aggregation of candidate representations from the recommendation model: \(\mathbf{h}_u^+ = \sum_{u' \in \mathcal{N}_u} p_{u,u'} \mathbf{h}_{u'}\). The synthesizer parameters are jointly trained with the recommendation model.

3. Intra-user Contrastive Learning via Item Semantic Retrieval: LLMs are similarly used to understand items (given item attributes and contextual sequence information), producing semantic embeddings \(\tilde{\mathbf{e}}_v\). Top-\(k\) similar items are retrieved to form a candidate pool \(\mathcal{N}_v\). Two semantically consistent augmented views \(\mathcal{S}_u', \mathcal{S}_u''\) are generated by randomly replacing 20% of items with semantically similar candidates, serving as positive pairs. A learnable synthesizer is not applied here, as experiments show no additional benefit—item semantics are more directly quantifiable than user preferences.

Loss & Training

• Recommendation loss: standard cross-entropy \(\mathcal{L}_{\text{Rec}} = -\hat{y}_{v^+} + \log(\sum_v \exp(\hat{y}_v))\)
• Cross-user contrastive loss (InfoNCE): \(\mathcal{L}_{\text{CS}} = -\log \frac{\exp(\mathbf{h}_u \cdot \mathbf{h}_u^+)}{\exp(\mathbf{h}_u \cdot \mathbf{h}_u^+) + \sum_{\mathbf{h}_u^-} \exp(\mathbf{h}_u \cdot \mathbf{h}_u^-)}\)
• Intra-user contrastive loss: \(\mathcal{L}_{\text{IS}}\) follows the same formulation, treating the representations of two augmented views as positive pairs and other in-batch samples as negatives.
• All semantic embeddings are pre-computed and frozen prior to training, introducing no inference latency.

Key Experimental Results

Main Results

Dataset	Metric	SRA-CL	MCLRec (2nd-best CL)	ICSRec	DuoRec	Gain
Yelp	HR@20	0.1282	0.1150	0.1165	0.1173	+9.29%
Yelp	NDCG@20	0.0533	0.0486	0.0495	0.0493	+7.68%
Sports	HR@20	0.0823	0.0736	0.0728	0.0706	+11.82%
Sports	NDCG@20	0.0347	0.0318	0.0304	0.0302	+9.12%
Beauty	HR@20	0.1314	0.1239	0.1205	0.1224	+6.05%
Office	HR@20	0.1702	0.1629	0.1643	0.1549	+3.59%

Ablation Study

Configuration	Description
w/o CL	Removing all contrastive learning results in significant performance degradation.
w/o \(\mathcal{L}_{\text{CS}}\)	Removing cross-user contrastive loss leads to a notable performance drop.
w/o \(\mathcal{L}_{\text{IS}}\)	Removing intra-user contrastive loss causes a moderate decline.
w/o learnable synthesizer	Replacing the synthesizer with hard rules degrades performance.
w/o semantic (random augmentation)	Substituting semantic retrieval with random augmentation yields the largest performance regression.
w/o LLM (raw text)	Using raw text embeddings without LLM processing reduces performance.

Key Findings

• The Sports dataset yields the largest gain (+11.82% HR@20), likely due to its high sparsity, where semantic information provides the greatest benefit.
• Model-agnostic validation: SRA-CL consistently improves GRU4Rec (+27.3% HR), SASRec (+15.2%), and DuoRec (+8.3%).
• The learnable synthesizer outperforms hard-rule selection by approximately 2–4%, demonstrating the importance of learning optimal fusion weights.
• No additional inference overhead is incurred, as semantic embeddings are pre-computed and used exclusively during training.

Highlights & Insights

• LLMs are elegantly repurposed as offline semantic encoders rather than online inference components, completely avoiding inference latency.
• The two-level retrieval design (user-level and item-level) covers both inter-user and intra-user contrastive learning paradigms.
• The learnable synthesizer reformulates "positive sample selection" as "weighted aggregation of candidates," offering greater flexibility than hard rules.
• Providing contextual sequence information when encoding items via LLM is a notable contribution—it enables the model to understand the role of items within the recommendation context.

Limitations & Future Work

• The LLM inference cost (DeepSeek-V3 API) is non-trivial; pre-computing semantic embeddings at scale introduces considerable overhead.
• Item context sequences are capped at 10 interactions, potentially missing important usage patterns.
• Sensitivity analysis on the Top-\(k\) hyperparameter is insufficient.
• Semantic embeddings are fully frozen and cannot be updated dynamically during training; lightweight fine-tuning strategies warrant exploration.
• Experiments are conducted on medium-scale datasets; effectiveness in industrial-scale settings remains to be validated.

Related Work & Insights

• Compared to LLM-augmented recommendation methods such as LRD and RLMRec, SRA-CL focuses specifically on improving contrastive learning rather than directly enhancing the recommendation model.
• The semantic retrieval augmentation paradigm is generalizable to other self-supervised learning tasks (e.g., graph representation learning).
• The attention mechanism in the learnable sample synthesizer can be interpreted as a soft sample selection strategy.

Rating

• Novelty: ⭐⭐⭐⭐ The combination of LLM semantics with contrastive learning is novel, and the learnable synthesizer design is elegant.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Four datasets, 13 baselines, model-agnostic validation, and detailed ablation studies.
• Writing Quality: ⭐⭐⭐⭐ Problem motivation is clearly articulated; figures and tables are intuitive.
• Value: ⭐⭐⭐⭐ Establishes a new semantic augmentation paradigm for contrastive learning in sequential recommendation.

Related Papers

• [NeurIPS 2025] TV-Rec: Time-Variant Convolutional Filter for Sequential Recommendation
• [ICLR 2026] C2AL: Cohort-Contrastive Auxiliary Learning for Large-scale Recommendation Systems
• [NeurIPS 2025] Transformer Copilot: Learning from The Mistake Log in LLM Fine-tuning
• [AAAI 2026] Inductive Generative Recommendation via Retrieval-based Speculation
• [AAAI 2026] From IDs to Semantics: A Generative Framework for Cross-Domain Recommendation with Adaptive Semantic Tokenization