Semantic-Retrieval-Reasoning Pipeline: A Hybrid Large Language Model Framework for Intent-Aware and Explainable Bundle Recommendation

Technical Summary: Semantic-Retrieval-Reasoning Pipeline (SRRP)

Problem Statement
Current Bundle Recommender Systems (BRS) face a triad of challenges: lack of explainability, rigid bundle sizing, and the "mono-intent" assumption. Traditional methods, ranging from constraint-based algorithms (Apriori, FP-Growth) to Graph Neural Networks (GNNs like BGCN and BundleNet), often function as "black boxes," failing to provide logical justifications for item groupings. Furthermore, these models frequently assume a single user intent per shopping basket, leading to monotonous recommendations. While recent generative approaches using Large Language Models (LLMs) have introduced explainability, they suffer from computational inefficiencies, combinatorial explosion, and a tendency to hallucinate. To mitigate these issues, existing generative models often enforce static, small bundle sizes (e.g., 2–3 items), which fails to accommodate the dynamic and diverse nature of real-world consumer needs, particularly for bulk or multi-intent purchases.

Methodology: The Semantic-Retrieval-Reasoning Pipeline (SRRP)
The authors propose SRRP, a hybrid framework that integrates LLM prompting and embedding techniques into a three-stage pipeline to generate intent-aware, dynamic, and explainable bundles.

1. Semantic Segmentation Unit (SSU):

   • Function: Acts as a cognitive filter to decompose noisy, multi-intent shopping carts into distinct intent clusters.
   • Mechanism: Utilizes an LLM with In-Context Learning (ICL) and few-shot prompting to perform semantic clustering. Instead of treating the basket as a single vector, the SSU labels items based on usage occasions (e.g., "Breakfast," "Office Supplies").
   • Theoretical Basis: The "Semantic Variance Reduction Theorem" posits that clustering items by functional proximity reduces intra-cluster semantic variance, thereby minimizing retrieval error bounds compared to processing the entire cart as a single entity.

2. Contextual Retrieval Unit (CRU):

   • Function: Retrieves factual candidate bundles from a historical database based on the intent clusters generated by the SSU.
   • Mechanism: Transforms intent clusters into high-dimensional dense vectors using a Sentence-Transformer (e.g., all-MiniLM-L6-v2). These vectors query a vector database (ChromaDB) using Cosine Similarity to find semantically relevant historical bundles.
   • Theoretical Basis: The "Semantic Relevance Bound Theorem" demonstrates that dense vector retrieval can identify semantically relevant items even without keyword overlap, outperforming traditional exact-match or association rule methods in recall.

3. Reasoning Validation Unit (RVU):

   • Function: Validates candidate bundles, dynamically adjusts bundle size, and generates persuasive natural language explanations and names.
   • Mechanism: Employs Chain-of-Thought (CoT) prompting to evaluate the logical complementarity of candidate items against the user's original cart. It filters out hallucinations, discards irrelevant items, and constructs explanatory narratives.
   • Human-in-the-Loop (HITL) Alignment: To ensure cognitive alignment and prevent hallucinations, the RVU's reasoning is validated against ground truth established by nine independent human annotators. The system uses a "Dynamic CoT" approach, injecting curated examples from this human consensus to align machine reasoning with human perception.
   • Theoretical Basis: The "Dynamic Bundle Size Optimality Theorem" argues that conditional pruning based on logical relevance minimizes inclusion errors compared to static sizing systems.

Key Contributions

• Hybrid Framework: The introduction of SRRP, which unifies technical accuracy (via embedding-based retrieval) and cognitive transparency (via LLM reasoning) in a single pipeline.
• Multi-Intent Decomposition: The use of LLMs as a cognitive filter to dynamically decompose noisy user baskets into multi-intent clusters, eliminating reliance on rigid manual feature tuning.
• Dynamic Bundle Sizing: A reasoning validation module that breaks the static constraints of previous generative models, allowing bundle sizes to adapt dynamically (ranging from 3 to 45 items) based on user context.
• Explainability and Validation: The integration of a Human-in-the-Loop methodology to rigorously validate bundle naming and explanations, ensuring alignment with human cognitive consensus and reducing hallucinations.

Experimental Results
Evaluated on the Instacart 2017 dataset using a Leave-One-Out protocol, SRRP was compared against five baseline models (Apriori, FP-Growth, BundleNet, BGCN, CrossCBR).

• Relevance and Ranking: SRRP demonstrated superior performance in Top-20 recommendations. The Random Masking (RM) strategy achieved a Hit Rate of 0.1100 and Recall of 0.1298, while the Sequential Splitting (SS) strategy reached a Recall of 0.1502 and MRR of 0.0683. These metrics significantly outperformed state-of-the-art graph models (e.g., CrossCBR Recall: 0.0942).
• Diversity: SRRP maintained high Intra-List Diversity (ILD > 0.53), achieving an optimal trade-off between accuracy and diversity, whereas graph models showed high diversity but significantly lower accuracy.
• Dynamic Sizing: The system successfully generated bundles with sizes between 3 and 5 items for approximately 79% of users, while robustly accommodating long-tail distributions up to 45 items for bulk buyers. It effectively suppressed the bias toward 2-item pairs common in deep learning baselines.
• Explainability and Alignment: Generated explanatory texts achieved a G-Eval Fluency score of 4.79/5.0. Semantic alignment with human annotators (measured by Cosine Similarity) reached an overall score of 0.5907 (LLaMA-3), indicating strong alignment with human cognitive reasoning regarding bundle validity and naming.

Significance and Claims
The paper claims that SRRP represents a significant advancement in Bundle Recommender Systems by addressing the "black box" nature of previous models and the rigidity of static sizing. By leveraging LLMs not just for generation but as a cognitive filter for intent detection and reasoning validation, the framework delivers recommendations that are not only accurate but also transparent and trustworthy. The authors assert that the system successfully bridges the gap between mathematical retrieval and human logical understanding, providing a rational and credible shopping assistant. However, the authors modestly acknowledge limitations, noting that the system occasionally suffers from "forced rationalization" (over-justifying random item combinations) and that future work is needed to reduce inference latency and further calibrate rejection thresholds to minimize false positives.