RevBrowse: Review-Driven Recommendations

• RevBrowse is a review-driven recommendation framework that leverages LLMs and dual-tower retrieval to extract and match user preferences from review data.
• It employs contrastive learning with InfoNCE loss to align dense user queries with key item review features, ensuring efficient and interpretable evidence retrieval.
• The framework integrates preference extraction, contextual review retrieval, and LLM-based reranking, simulating a human browse-then-decide strategy for improved e-commerce recommendations.

RevBrowse is a review-driven recommendation framework designed to harness LLMs and retrieval-augmented methods for next-item recommendation based on user-generated review data. The system is explicitly inspired by the "browse-then-decide" behavioral pattern observed in e-commerce, aiming to combine semantic comprehension, efficient filtering, and interpretability by dynamically retrieving and leveraging the most relevant review content during the recommendation process (Wei et al., 31 Aug 2025).

1. Framework Architecture and Workflow

RevBrowse integrates three core components functioning in a staged pipeline that emulates human browsing and decision-making:

• Preference and Feature Extraction: An LLM (e.g., Qwen2.5-72B-instruct) is prompted with the user's historical reviews to extract structured "Like" and "Dislike" preferences. Item reviews are parsed to extract lists of "Pros" and "Cons" via templated prompts.
• PrefRAG: Retrieval-Augmented Browsing: The PrefRAG module—the "browsing" analog—uses a dual-tower retrieval architecture, disentangling user and item features. User preference signals are mapped to a dense query representation, and item review features are encoded orthogonally. Contrastive learning with InfoNCE loss is applied to align user queries with preference-relevant review content, and a top-K retrieval selects the most contextually germane fragments for further reasoning.
• LLM-Based Recommendation and Reranking: The selected candidate items (from, e.g., a baseline Lightweight Retrieval Unit) are represented as structured prompts containing user preference summaries and retrieved review snippets. A fine-tuned LLM (Llama2-7B, trained with LoRA in the paper) synthesizes this tokenized context and executes the final reranking and decision.

This composition directly addresses the inefficiency and context-window limitations of LLMs when presented with entire review histories, supporting adaptive attention to contextually relevant evidence without superfluous noise.

2. LLMs: Extraction, Summarization, and Reasoning

LLMs are central to RevBrowse at multiple stages:

• User and Item Representation Extraction: Prompt templates are used for instructing LLMs to derive explicit structured preferences from raw review texts, organizing them (“Like”, “Dislike”, “Pros”, “Cons”) for tractable consumption.
• Semantic Reasoning and Integration: The LLM-based reranker combines signals from baseline history, retrieved fragments, and global summaries, providing coherent judgments on which candidate item fits the user's current tastes.
• Interpretability: As the recommender's output includes the exact review fragments and preferences that drive its decisions, the system exposes the rationale for ranking, supporting ex post interpretability.

The LLM effectively serves as both semantic indexer (feature extraction) and synthesis engine (decision making), leveraging its ability to contextualize and integrate information beyond the reach of classical shallow architectures.

3. Granular Incorporation of User Reviews

RevBrowse emphasizes an adaptive, information-reducing retrieval strategy:

• Dual-Tower Architecture: The system trains user-side and item-side encoders to create dense representations for query (user likes/dislikes) and answer (item pros/cons) vectors, respectively.
• Contrastive Learning and InfoNCE Loss: For each batch of positive (query, review fragment) pairs and m negative examples per query,

$$L = -\frac{1}{B} \sum_{i} \log \frac{e^{s(q_i, a_i^+)}}{e^{s(q_i, a_i^+)} + \sum_j e^{s(q_i, a_{i,j}^-)}}$$

where $s(q,a)$ is the cosine similarity between the respective encodings.

• Top-K Retrieval: PrefRAG selects the top-K review features (K empirically set to 2) for each user–candidate pair, thereby reducing irrelevant evidence and constraining input length for the reranking LLM.

This process ensures only the most aligned fragments with user intent are injected into the reranking prompt.

4. Empirical Results and Benchmarks

RevBrowse was evaluated on four Amazon-2014 verticals (Food, Sports, Clothing, Games) using standard sequential recommendation metrics: Recall@5/10, NDCG@5/10, and MRR@5/10.

• Comparison with Baselines: RevBrowse achieved consistent and substantial improvements over both classic collaborative filtering models (LightGCN, BERT4Rec, SASRec) and prior review-based models (RGCL, DeepCoNN, NARRE), as well as LLM-based methods (ZS-Ranker, LlamaRec, Exp3rt).
• Attribution of Gains: Performance gains are attributed to the synergy of (i) explicit global summarization of user preferences, and (ii) selective, top-K retrieval of context-sensitive item evidence, effectively accommodating both general preference trends and specific, instance-level review matches.

A plausible implication is that, in dynamic recommendation scenarios, models that perform both global user modeling and contextualized review evidence retrieval may generalize better than those using a single review summarization approach.

5. Interpretability and Transparency

Interpretability is directly addressed by RevBrowse as follows:

• Evidence Traceability: The system exposes, for each recommended item, exactly which "pro" or "con" review fragments matched the user's "likes" and "dislikes." In the provided case paper, 3 out of 4 retrieved positive aspects supported the system's selection, and the absence of critical negative features further justified the recommendation.
• Structured Outputs: All review evidence injected during reranking is explicitly marked in the output, allowing practitioners and end users to audit decisions.
• Model Transparency: Disentangling the extraction and retrieval process clarifies the flow of information from raw data to final output.

This approach supports explainable AI and may facilitate debugging, user trust calibration, and regulatory compliance in recommendation platforms.

6. Mathematical Formulations and Key Technical Elements

Key mathematical operations formalized in the framework include:

• Preference Extraction: $$t_{\text{like}}^u, t_{\text{dislike}}^u = \varphi_t(R_u, T_{\text{user}})$$
• Item Feature Extraction: $$\text{pros}_i, \text{cons}_i = \varphi_t(r_i, T_{\text{item}})$$
• Cosine Similarity for Retrieval:

$s(q, a) = \frac{e_q^T e_a}{\|e_q\| \cdot \|e_a\|}$

• InfoNCE Loss for Contrastive Learning:

$$L = -\frac{1}{B} \sum_{i} \log \frac{\exp(s(q_i, a^+_i))}{\exp(s(q_i, a^+_i)) + \sum_j \exp(s(q_i, a^-_{i,j}))}$$

• Recommender Loss: $$\mathcal{L}_{rec} = -\frac{1}{N} \sum_i \log P_{\theta}(y_i | x_i)$$ where $x_i$ is the prompt and $y_i$ is the label, for model parameters $\theta$.

These provide the building blocks for robust representation learning, evidence-matching, and final decision prediction in the context of long, noisy, and context-dependent review histories.

7. Conclusion

RevBrowse exemplifies a hybrid LLM + retrieval framework that marries structured information extraction, targeted evidence retrieval, and transparent, explainable output for review-based recommendation. By reflecting the human browse-then-decide logic and integrating a dual process of preference summarization and adaptive, review-level evidence selection, it establishes a template for future explainable, scalable, and context-sensitive recommender systems leveraging large-scale user-generated textual data (Wei et al., 31 Aug 2025).