Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

CoNRec: Context-Discerning Negative Recommendation with LLMs

Xinda Chen, Jiawei Wu, Yishuang Liu, Jialin Zhu, Shuwen Xiao, Junjun Zheng, Xiangheng Kong, Yuning Jiang

TL;DR

CoNRec tackles the underexplored problem of modeling negative user preferences in offline recommendations by combining semantic item representations with large language models. It introduces semantic IDs via an RQ-VAE, an item-level alignment task, and a Progressive GRPO training regime with unbiased rewards that incorporate future negative feedback and collaboration signals. The approach achieves state-of-the-art offline performance on Taobao data, with strong improvements in cold-start scenarios and online-like candidate accuracy, while mitigating noise from positive feedback. This framework enables practical offline negative-filtering for large-scale systems and offers insights for extending negative-feedback modeling to broader recommendation settings.

Abstract

Understanding what users like is relatively straightforward; understanding what users dislike, however, remains a challenging and underexplored problem. Research into users' negative preferences has gained increasing importance in modern recommendation systems. Numerous platforms have introduced explicit negative feedback mechanisms and leverage such signals to refine their recommendation models. Beyond traditional business metrics, user experience-driven metrics, such as negative feedback rates, have become critical indicators for evaluating system performance. However, most existing approaches primarily use negative feedback as an auxiliary signal to enhance positive recommendations, paying little attention to directly modeling negative interests, which can be highly valuable in offline applications. Moreover, due to the inherent sparsity of negative feedback data, models often suffer from context understanding biases induced by positive feedback dominance. To address these challenges, we propose the first large language model framework for negative feedback modeling with special designed context-discerning modules. We use semantic ID Representation to replace text-based item descriptions and introduce an item-level alignment task that enhances the LLM's understanding of the semantic context behind negative feedback. Furthermore, we design a Progressive GRPO training paradigm that enables the model to dynamically balance the positive and negative behavioral context utilization. Besides, our investigation further reveals a fundamental misalignment between the conventional next-negative-item prediction objective and users' true negative preferences, which is heavily influenced by the system's recommendation order. To mitigate this, we propose a novel reward function and evaluation metric grounded in multi-day future negative feedback and their collaborative signals.

CoNRec: Context-Discerning Negative Recommendation with LLMs

TL;DR

CoNRec tackles the underexplored problem of modeling negative user preferences in offline recommendations by combining semantic item representations with large language models. It introduces semantic IDs via an RQ-VAE, an item-level alignment task, and a Progressive GRPO training regime with unbiased rewards that incorporate future negative feedback and collaboration signals. The approach achieves state-of-the-art offline performance on Taobao data, with strong improvements in cold-start scenarios and online-like candidate accuracy, while mitigating noise from positive feedback. This framework enables practical offline negative-filtering for large-scale systems and offers insights for extending negative-feedback modeling to broader recommendation settings.

Abstract

Understanding what users like is relatively straightforward; understanding what users dislike, however, remains a challenging and underexplored problem. Research into users' negative preferences has gained increasing importance in modern recommendation systems. Numerous platforms have introduced explicit negative feedback mechanisms and leverage such signals to refine their recommendation models. Beyond traditional business metrics, user experience-driven metrics, such as negative feedback rates, have become critical indicators for evaluating system performance. However, most existing approaches primarily use negative feedback as an auxiliary signal to enhance positive recommendations, paying little attention to directly modeling negative interests, which can be highly valuable in offline applications. Moreover, due to the inherent sparsity of negative feedback data, models often suffer from context understanding biases induced by positive feedback dominance. To address these challenges, we propose the first large language model framework for negative feedback modeling with special designed context-discerning modules. We use semantic ID Representation to replace text-based item descriptions and introduce an item-level alignment task that enhances the LLM's understanding of the semantic context behind negative feedback. Furthermore, we design a Progressive GRPO training paradigm that enables the model to dynamically balance the positive and negative behavioral context utilization. Besides, our investigation further reveals a fundamental misalignment between the conventional next-negative-item prediction objective and users' true negative preferences, which is heavily influenced by the system's recommendation order. To mitigate this, we propose a novel reward function and evaluation metric grounded in multi-day future negative feedback and their collaborative signals.
Paper Structure (14 sections, 7 equations, 4 figures, 5 tables, 1 algorithm)

This paper contains 14 sections, 7 equations, 4 figures, 5 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Motivational Studies
  3. CoNRec
  4. Experiments
  5. Related Work
  6. Conclusion
  7. RQ-VAE Implementation
  8. RQ-VAE Architecture and Training
  9. Implementation Notes
  10. More Experimental Details
  11. Data Cleaning
  12. Baseline Models
  13. LLM backbone setting
  14. Limitation

Figures (4)

  • Figure 1: User Negative-Interest Modeling (icon generated by Doubao): For a user who dislikes bulky footwear and wired audio (A, C, E in bold), rule-based methods lead to over-suppression (red box represents wrong results) while traditional models perform poorly on cold-start items like bulky slippers (C), which never appear before. CoNRec effectively captures users' negative interests.
  • Figure 2: Illustration of motivational studies. (a) The next negative feedback item covers only 7% of the user's top negative interest (17% for Top-4), causing a lot of noise; while extending to a future 7-day horizon can raise Top-4 coverage to 48%. (b) Adding extra positive interactions (3× longer than negative) unexpectedly causes a large performance drop.
  • Figure 3: Overview of CoNRec framework. CoNRec first compresses the item information into semantic IDs, which are used in both Context Understanding stage and Context Utilization stage. Context Understanding includes the LoRA finetuning of traditional bidirectional translation, as well as proposed item-level alignment. Then, the model is post-trained using GRPO, where we progressively increase the complexity of the context during training, with a novel reward design that utilizes future negative, positive feedback and collaborative items to create an unbiased environment.
  • Figure 4: Illustration of CoNRec's offline industrial application. The target item from the ranking stage and the items predicted by CoNRec are reconstructed into embeddings via the stored codebook. A similarity score is then computed as the maximum embedding similarity between the target and predicted items. Items with scores exceeding the threshold are filtered out.