LLM-I2I: Boost Your Small Item2Item Recommendation Model with Large Language Model

Abstract: Item-to-Item (I2I) recommendation models are widely used in real-world systems due to their scalability, real-time capabilities, and high recommendation quality. Research to enhance I2I performance focuses on two directions: 1) model-centric approaches, which adopt deeper architectures but risk increased computational costs and deployment complexity, and 2) data-centric methods, which refine training data without altering models, offering cost-effectiveness but struggling with data sparsity and noise. To address these challenges, we propose LLM-I2I, a data-centric framework leveraging LLMs to mitigate data quality issues. LLM-I2I includes (1) an LLM-based generator that synthesizes user-item interactions for long-tail items, alleviating data sparsity, and (2) an LLM-based discriminator that filters noisy interactions from real and synthetic data. The refined data is then fused to train I2I models. Evaluated on industry (AEDS) and academic (ARD) datasets, LLM-I2I consistently improves recommendation accuracy, particularly for long-tail items. Deployed on a large-scale cross-border e-commerce platform, it boosts recall number (RN) by 6.02% and gross merchandise value (GMV) by 1.22% over existing I2I models. This work highlights the potential of LLMs in enhancing data-centric recommendation systems without modifying model architectures.

• The paper's main contribution is the novel LLM-I2I framework that synthesizes and filters interactions to enhance recommendation performance.
• It employs LLM-based generator and discriminator modules to generate plausible long-tail interactions and remove noisy data without altering underlying models.
• Empirical results demonstrate significant improvements in Recall@K and NDCG@K, highlighting scalability and real-world applicability.

LLM-I2I: Leveraging LLMs for Data-Centric Item-to-Item Recommendation Enhancement

Introduction and Motivation

Item-to-Item (I2I) recommendation models maintain their industrial relevance due to intrinsic scalability and efficiency, but their performance is tightly constrained by data quality and availability. Enhancing I2I systems traditionally follows either model-centric approaches—introducing deeper learning architectures—or data-centric strategies that focus on refining the training corpus. The former incurs substantial computational and operational costs, whereas the latter, being agnostic to model architecture, can be deployed without requiring increased latency or infrastructural redesign (Figure 1).

Figure 1: A comparative illustration of model-centric vs. data-centric approaches in recommendation system optimization.

Within the data-centric paradigm, the dual challenges of data sparsity—crucial in the context of long-tail items—and interaction noise pose significant hurdles. Especially in e-commerce, the majority of items are underrepresented in historical user interactions, impairing their discoverability. Parallelly, noisy interactions (e.g., accidental clicks) skew the learning process. The paper introduces LLM-I2I, a framework that synthesizes new user-item interactions with an LLM-based generator and filters these (and original interactions) via an LLM-based discriminator, thus augmenting I2I models with more informative and less noisy data, especially for the long tail.

LLM-I2I Framework

The LLM-I2I architecture is designed to function as a layer atop existing I2I models, where neither inference nor architecture of underlying recommenders is altered. The pipeline comprises two supervised-finetuned LLM-powered modules:

1. LLM-based Generator: Given users’ profiles and recent interactions, this module synthesizes plausible next-interaction items, using a loss function that upweights long-tail candidates to counteract their underrepresentation.
2. LLM-based Discriminator: Both original and synthetic user-item pairs are evaluated for interaction quality, employing classification and confidence-thresholding to remove noisy or low-confidence data.

   Figure 2: The framework of our proposed LLM-I2I.

Fine-tuning with SFT ensures domain adaptation, a critical requirement given e-commerce’s idiosyncratic item semantics and behavior sequences. The generator’s training objective includes a tunable weighting scheme to emphasize long-tail recovery:

$$\mathcal{L}_g = \sum_{u_{i}}\sum_{t} w_{u_i^{t+1}} \, \mathrm{loss}(\cdot)$$

where $w_{u_i^{t+1}} = \alpha$ for long-tail items, $\beta$ otherwise (with $\alpha > \beta$).

Figure 3: Supervised Fine-Tuning (SFT) of the LLM-based generator and discriminator.

After post-processing by the discriminator, only high-confidence interactions (confidence = 1.0) are retained, forming an augmented dataset for final I2I training.

Data Characteristics and Sparsity Analysis

A core issue addressed is that user-item interaction matrices exhibit pronounced long-tail distributions: as shown in Figure 4, nearly 20% of items in Amazon’s Toys and Games dataset record ≤5 purchases, highlighting severe sparsity for classic similarity-based methods.

Figure 4: Data distribution in ARD Toys And Games dataset.

Traditional I2I models struggle to establish meaningful similarity measures for these rare items, and naïve augmentation or content-based approaches lack sufficient efficacy at scale.

Empirical Results and Numerical Claims

Benchmarks on both public academic datasets (Amazon Reviews—Beauty, Sports & Outdoors, Toys & Games) and a billion-scale industry dataset (AEDS from AliExpress) substantiate multiple claims:

• All backbone models (BM25, BPR, YoutubeDNN, Swing) achieve improved Recall@K and NDCG@K when augmented via LLM-I2I versus both no-augmentation and LLM-CF baselines.
• On the AEDS dataset, the LLM-I2I enhancement on Swing produces an 18.57% uplift in Recall@5 and 16.34% uplift in NDCG@5 over baseline.
• For long-tail items, Recall@10 and NDCG@10 rise by 60.75% and 85.71% respectively on AEDS—strong evidence of LLM-I2I’s long-tail alleviation capability.

Figure 5 demonstrates the effect of varying generator recall numbers on model performance: increasing synthetic interactions initially promotes accuracy, but excess augmentation eventually induces distributional drift and degradation, emphasizing the necessity of discriminator-driven filtering.

Figure 5: The influence of recall number on the LLM-based generator.

Production deployment in AliExpress yields 6.02% improvement in recall number (RN) and 1.22% in GMV—asserted without increased system latency, attesting to scalable practical benefit.

Ablation and Component Contribution

Extensive ablation experiments confirm that both the generator and discriminator (particularly the long-tail loss and high-confidence filtering) are essential for full efficacy. Removal of either module or their specialized objectives significantly impairs performance, and integration with collaborative filtering (Swing) outclasses standalone generative modeling.

Implications and Future Directions

LLM-I2I materializes a pathway to combining foundation LLMs with traditional recommendation indexes at scale without online inference overhead, using data-centric methodologies. Its architecture is modular—amenable for retrofitting existing systems and periodic offline refreshes—as opposed to end-to-end model retraining or heavier hybrid recommenders.

Practically, its strongest implication is the elevation of long-tail coverage without tradeoffs in latency or infrastructure footprint, relevant to large-scale deployments. The formal demonstration of bidirectional synthesis/filtering validates the claim that indiscriminate augmentation can harm performance, and that discriminative refinement is mandatory as dataset dimensionality increases.

Theoretically, this establishes further linkage between LLMs’ generative priors and collaborative signals in systems where explicit semantic representation is sparse.

Looking forward, advances may be possible by migrating to multi-modal generative models, integrating structured knowledge graphs or session-level context, or even training smaller domain-specific LLMs for further cost reduction. An interesting direction is leveraging online learning and active filtering in real time, rather than periodic offline updates.

Conclusion

LLM-I2I sets a technical precedent for data refinement in I2I recommenders via dual LLM modules. The framework yields demonstrable performance improvements over canonical baselines without reengineering recommender architectures, principally via targeted long-tail data synthesis and noise filtration. Both practical business metrics and theoretical recommendation efficacy are elevated, suggesting adoption for robust scalable systems with high item-dimensionality and user heterogeneity. Future explorations may refine its generator/discriminator interplay, hybridize with user-centric modeling, or extend to rich multi-modal behavior domains.