DeepInterestGR: Mining Deep Multi-Interest Using Multi-Modal LLMs for Generative Recommendation

Yangchen Zeng; Zhenyu Yu; Zhiyuan Hu; Wenxin Zhang; Jinze Wang; Rongfeng Guo

cs.LGFeb 21, 2026

DeepInterestGR: Mining Deep Multi-Interest Using Multi-Modal LLMs for Generative Recommendation

Authors: Yangchen Zeng, Zhenyu Yu, Zhiyuan Hu, Wenxin Zhang, Jinze Wang, Rongfeng Guo

Organizations: Southeast University

Abstract

We introduce DeepInterestGR, a novel framework that integrates deep interest mining into the generative recommendation pipeline. This addresses the "Shallow Interest" problem - existing generative methods rely on surface-level textual features and fail to capture latent user motivations, limiting personalization depth and recommendation interpretability. Our approach leverages Multi-LLM Interest Mining (MLIM) via structured reasoning prompting, Reward-Labeled Deep Interest (RLDI) for quality control, and Interest-Enhanced Item Discretization (IEID) via RQ-VAE, combined with a two-stage SFT-GRPO training pipeline guided by an Interest-Aware Reward. We validate DeepInterestGR on three Amazon Review benchmarks (Beauty, Sports, Instruments), comparing against 14 state-of-the-art baselines including SASRec, BERT4Rec, TIGER, LC-Rec, and S-DPO. Our method achieves 5.8%-8.3% relative improvements on HR@10 and 7.7%-9.9% on NDCG@10 over the strongest baseline, with cross-domain generalization gains of +24.8%. These results provide evidence that incorporating deep semantic interests can effectively improve SID-based generative recommendation.

Figures & tables

Figure 1: Comparison of traditional generative recommendation (top) and DeepInterestGR (bottom). Traditional methods suffer from “Shallow Interest” by only using surface-level features, while DeepInterestGR incorporates deep interest mining and Interest-Aware Reward for reinforcement learning optimization.

Figure 2: DeepInterestGR workflow overview. User interactions are encoded as SID tokens, which are processed through Multi-LLM Interest Mining (GPT, Gemini, Kimi, Grok) to generate deep interest representations. These interests compute the Interest-Aware Reward that guides the Generative LLM for autoregressive next-SID prediction, forming a closed-loop optimization process. See Appendix Figure 4 for detailed architecture including IEID encoding and two-stage training.

Method	Beauty				Sports				Instruments
Method	HR@5	HR@10	N@5	N@10	HR@5	HR@10	N@5	N@10	HR@5	HR@10	N@5	N@10
Traditional Sequential Models
GRU4Rec	0.0312	0.0518	0.0189	0.0256	0.0198	0.0324	0.0118	0.0161	0.0285	0.0467	0.0172	0.0231
Caser	0.0287	0.0483	0.0171	0.0234	0.0175	0.0291	0.0103	0.0142	0.0261	0.0432	0.0156	0.0212
HGN	0.0335	0.0549	0.0201	0.0271	0.0212	0.0348	0.0126	0.0172	0.0302	0.0495	0.0183	0.0246
Transformer-based Models

Table 1: Overall performance comparison on three Amazon Product Reviews datasets. Bold indicates the best performance, and underline indicates the second best.

\Delta

denotes the relative improvement of DeepInterestGR over the best baseline.

Variant	HR@5	HR@10	N@5	N@10
DeepInterestGR (Full)	0.0678	0.1032	0.0436	0.0558
w/o MLIM	0.0598	0.0921	0.0378	0.0489
w/o IEID	0.0621	0.0958	0.0401	0.0517
w/o Interest-Aware Reward	0.0635	0.0973	0.0408	0.0526
w/o RL (SFT only)	0.0567	0.0879	0.0358	0.0463

Table 2: Ablation study on the Beauty dataset. Each row removes one component from the full model.

Figure 3: Component ablation study on the Beauty dataset (HR@10). Blue bar represents the full model, orange bars represent ablation variants. The RL stage yields the largest overall impact (-14.8%), while MLIM is the most critical novel component (-10.8%) as it provides the semantic foundation for Interest-Aware Reward.

Setting	HR@5	HR@10	N@5	N@10
Text-Only	0.0641	0.0983	0.0412	0.0529
+Multi-Modal	0.0678	0.1032	0.0436	0.0558
$\Delta$ Improv.	+5.8%	+5.0%	+5.8%	+5.5%

Table 3: Impact of multi-modal interest mining on the Beauty dataset.

Setting	HR@5	HR@10	N@5	N@10
Reward Strategy Comparison
Rule-based (Binary)	0.0601	0.0928	0.0382	0.0493
Collaborative	0.0623	0.0954	0.0398	0.0513
Prefix-Match	0.0642	0.0981	0.0411	0.0528
RLDI Label Quality
w/o RLDI (uniform)	0.0612	0.0945	0.0392	0.0506

Table 4: Comparison of reward strategies and RLDI settings on Beauty.

Train $\rightarrow$ Test	MiniOneRec	DeepInterestGR
Beauty $\rightarrow$ Sports	0.0412 / 0.0218	0.0523 / 0.0284
Sports $\rightarrow$ Instruments	0.0578 / 0.0312	0.0712 / 0.0398
Instruments $\rightarrow$ Beauty	0.0623 / 0.0341	0.0768 / 0.0425
Avg. Improvement	—	+24.8% / +27.3%

Table 5: Cross-domain generalization performance (HR@10 / N@10).

Appendix figures & tables7 assets

Supplementary material from the paper’s appendix.

Appendix

Notation	Description
$\mathcal{U},\mathcal{I}$	Set of users and items
$N,M$	Number of users and items
$\mathcal{S}_{u}$	Interaction sequence of user $u$
$\mathbf{t}_{i},\mathbf{d}_{i},\mathbf{v}_{i}$	Title, description, and image of item $i$
$\mathbf{e}_{i}$	Embedding vector of item $i$
$\mathbf{z}_{i}$	Deep interest representation of item $i$

Appendix

Table 6: Summary of key notations.

Figure 4: Detailed architecture of DeepInterestGR. Left : Multi-LLM Interest Mining (MLIM) extracts deep interests using frontier LLMs via API with structured reasoning prompting. Item textual metadata is encoded into [TEXT_EMB] via text encoders, while product images yield [IMG_EMB] via vision encoders. These embeddings are concatenated and fed into RQ-VAE for Interest-Enhanced Item Discretization (IEID), producing hierarchical Semantic ID tokens. Right : Two-stage training pipeline—Stage 1 (SFT) trains Qwen2.5-7B-Instruct for autoregressive next-SID generation; Stage 2 (RL) applies GRPO optimization with Qwen2.5-7B-Instruct guided by our Interest-Aware Reward derived from RLDI labels.

LLM for MLIM	IQ	HR@10	N@10
GPT (API)	0.847	0.0978	0.0521
Gemini (API)	0.832	0.0961	0.0508
Kimi (API)	0.819	0.0943	0.0495
Grok (API)	0.825	0.0952	0.0502
Ensemble (All)	0.891	0.1032	0.0558

Appendix

Table 7: Comparison of different LLMs for deep interest mining on the Beauty dataset.

Figure 5: Impact of reinforcement learning with Interest-Aware Reward on three datasets. Orange bars represent SFT-only performance, while blue bars show final performance after RL training with GRPO. All datasets show consistent improvements (9.2%–12.1%).

Dataset	Users	Items	Interact.	Sparsity
Beauty	22,363	12,101	198,360	0.00073
Sports	35,598	18,357	296,175	0.00045
Instruments	24,733	9,923	206,153	0.00083

Appendix

Table 8: Statistics of the Datasets.

Figure 6: Dual-panel analysis of RLDI binary classification performance. (a) Confusion matrix with counts and percentages (human evaluation on 500 samples). (b) Performance metrics including precision, recall, F1-score, accuracy, and Cohen’s kappa (k = 0.720).

SID Length (H)	HR@10
2	0.0956
4	0.1032
6	0.1015
8	0.0987

Appendix

Table 9: SID length sensitivity analysis on Beauty dataset (HR@10).

Recommender Systems

Recommender systems predict and rank items for people using interaction histories, content, context, and feedback.

See all 514 papers

Deep Interest Mining for Intent-Enriched Semantic IDs in Multimodal Generative Recommendation
Mar 3, 2026Yangchen Zeng, Jinze WangRecommender Systems Semantic ID Recommendation
Iterative Semantic Reasoning from Individual to Group Interests for Generative Recommendation with LLMs
Mar 14, 2026Xiaofei Zhu, Jinfei Chen, Feiyang Yuan +1Recommender Systems Generative Recommendation
Chongqing University of Technology, College of Computer Science and Engineering, Chongqing, China · Chongqing Normal University, College of Computer and Information Science, Chongqing, China
REGER: Modeling inter-item Relations with Semantic IDs for Generative Recommendation
Oct 27, 2026Yuxi Huang, Silin Zhou, Chenhao Wang +5Generative Recommendation Semantic ID Recommendation
University of Electronic Science and Technology of China

Mar 3, 2026J/K move · Enter open · S save

Deep Interest Mining for Intent-Enriched Semantic IDs in Multimodal Generative Recommendation

Yangchen Zeng, Jinze Wang

Recommender Systems Semantic ID Recommendation Generative Recommendation Multimodal Recommendation RL for Generative Models

Open paper AskPDF arXiv

Semantic IDs (SIDs) provide the discrete item vocabulary used by generative recommendation, but their quality depends on what item evidence is preserved before quantization. In product recommendation, surface metadata often misses latent usage intent, visual evidence may be only weakly reflected in text, and downstream policy learning provides sparse feedback about whether a generated SID corresponds to a semantically useful item. We introduce DeepInterestGR, an intent-enriched SID framework for generative recommendation. Before SID quantization, CMSA enriches item representations through two complementary evidence paths: recommendation-oriented VLM captions and projected image embeddings. DCIM then uses an LLM to mine item-side intent descriptors -- latent usage motivations implied by product content rather than personalized user states. During policy training over the constructed SIDs, QARM adds a relevance-gated semantic-quality bonus on top of standard SID rewards, applying the bonus only when the generated SID decodes to the target item. Thus, semantic quality cannot reward a fluent but irrelevant item prediction. Experiments on three Amazon Product Review categories (Beauty, Sports, and Instruments) show that DeepInterestGR improves over competitive generative and RL-based baselines, with relative gains of up to 15.1% in NDCG@5 and 13.9% in NDCG@10 over the strongest per-metric baseline. Component ablations, CMSA branch analyses, reward variants, and SID-level case studies support a bounded claim: enriching pre-quantization item evidence with visual cues and item-side intent descriptors, together with relevance-gated semantic rewards, improves SID-based generative recommendation under the evaluated settings.

Abstract

Recommender Systems

Explore similar work

Deep Interest Mining for Intent-Enriched Semantic IDs in Multimodal Generative Recommendation

Iterative Semantic Reasoning from Individual to Group Interests for Generative Recommendation with LLMs

REGER: Modeling inter-item Relations with Semantic IDs for Generative Recommendation