Title: Mining Deep Interests for Intent-Rich Semantic IDs in Multimodal Generative Recommendation

Abstract: The efficacy of Semantic IDs (SIDs), which serve as the discrete item vocabulary in generative recommendation systems, is contingent upon the preservation of item evidence prior to quantization. In the context of product recommendation, surface-level metadata frequently overlooks latent usage intent; visual information is often only faintly captured by text; and feedback from downstream policy learning remains sparse regarding whether a generated SID aligns with a semantically meaningful item. To address these challenges, we propose DeepInterestGR, a framework designed to generate intent-enriched SIDs for generative recommendation.

Prior to SID quantization, our CMSA module enhances item representations by leveraging two complementary evidence streams: recommendation-focused captions generated by Vision-Language Models (VLMs) and projected image embeddings. Subsequently, DCIM employs a Large Language Model (LLM) to extract intent descriptors from the item side, identifying latent usage motivations inherent in the product content rather than relying on personalized user states. During the policy training phase for the constructed SIDs, QARM introduces a relevance-gated semantic-quality bonus alongside standard SID rewards. This bonus is applied exclusively when the generated SID successfully decodes to the target item, thereby preventing the reward mechanism from incentivizing fluent but irrelevant item predictions.

We evaluated our approach on three Amazon Product Review datasets: Beauty, Sports, and Instruments. The results demonstrate that DeepInterestGR outperforms competitive generative and reinforcement learning-based baselines, achieving relative improvements of up to 15.1% in NDCG@5 and 13.9% in NDCG@10 compared to the strongest baseline for each respective metric. Through component ablations, analyses of CMSA branches, reward variants, and case studies at the SID level, we substantiate the claim that incorporating visual cues and item-side intent descriptors into pre-quantization evidence, combined with relevance-gated semantic rewards, significantly enhances SID-based generative recommendation within the tested parameters.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC