Qi Xin

Behavior Retrieval plus Response Generation for Interpretable Conversational Personalized Recommendation

Introduction

Behavior retrieval plus response generation for interpretable conversational personalized recommendation. Empirical study on interpretable conversational personalized recommendation using behavior retrieval & response generation. Achieves high accuracy on retail data, outperforming baselines without unconstrained LLM generation.

Abstract

This paper presents an empirical study of interpretable conversational personalized recommendation built from two tightly coupled layers: a behavior retrieval layer for candidate ranking and a response generation layer for grounded recommendation wording. The study is motivated by recent large language model surveys in recommendations, but the implementation remains deterministic so that candidate selection, evidence extraction, and wording can be inspected separately. We used the 2010-2011 partition of Online Retail II for transaction-grounded next-basket recommendation and a retail/service subset of the Bitext customer-support 27K corpus for conversational response generation. The behavior retrieval layer combined popularity, repeat purchase memory, recency-weighted memory, item-to-item collaborative filtering, and user-neighborhood scoring. The response layer compared intent-majority, TF-IDF retrieval, intent-conditioned retrieval, and a constrained fusion verbalizer. On 1,999 test users, the hybrid behavior retriever achieved Hit@10 = 0.7282, Recall@10 = 0.1982, NDCG@10 = 0.2855, and MRR@10 = 0.4467, outperforming the strongest baseline RecencyRepeat by 2.03% on Hit@10 and 6.34% on NDCG@10. On 1,989 Bitext test utterances, intent-conditioned retrieval produced the best standalone verbalization quality with BLEU-4 = 0.1699, ROUGE-L = 0.3431, intent accuracy = 0.9915, and slot F1 = 0.8239. Grounded recommendation responses reached Top1Hit = 0.3252, HistoryGroundRate = 1.0000, and BundleValidity = 0.9980. The results show that retrieval-first conversational recommendation can combine ranking accuracy, grounding faithfulness, and low-overhead response realization without relying on unconstrained generation.

Review

This paper introduces an interpretable conversational personalized recommendation system built upon a two-layer architecture: a behavior retrieval layer for candidate ranking and a response generation layer for grounded recommendation wording. A notable strength of this work lies in its deliberate choice of a deterministic implementation, which directly addresses the critical need for inspectability in candidate selection, evidence extraction, and verbalization, particularly in contrast to unconstrained large language model approaches. The behavior retrieval layer adeptly combines established techniques such as popularity, repeat purchase, recency-weighted memory, item-to-item collaborative filtering, and user-neighborhood scoring. The empirical results demonstrate strong quantitative performance, with the hybrid behavior retriever achieving significant improvements over a strong baseline on standard recommendation metrics, while the intent-conditioned retrieval approach produced high-quality standalone verbalizations and grounded recommendation responses with excellent grounding faithfulness. While the deterministic approach is a core strength for interpretability, it inherently poses limitations on the system's adaptability and naturalness compared to more sophisticated generative models, particularly when encountering novel user queries or complex conversational nuances not explicitly covered by its structured generation framework. The abstract provides comprehensive quantitative metrics, yet a deeper qualitative analysis of *why* certain verbalization strategies excel or a discussion of the specific trade-offs between different retrieval components would further enrich the findings. Furthermore, the reliance on the 2010-2011 Online Retail II dataset, while providing solid transaction data, might limit the generalizability of the findings to contemporary retail environments that often feature richer product metadata and more dynamic user interaction patterns. The absence of user studies also means that the perceived "interpretability" and "conversational" quality from an end-user perspective remains underexplored. Overall, this paper makes a valuable contribution by presenting a robust and inspectable framework for conversational recommendation, offering a compelling alternative to black-box generative models. The modular design, effectively integrating strong retrieval capabilities with grounded response generation, provides a solid foundation for applications where transparency, accountability, and resource efficiency are paramount. Future research could explore hybrid approaches that strategically integrate more flexible, yet controlled, generative components within this deterministic framework to enhance response diversity without compromising interpretability. Conducting comprehensive user studies to validate the perceived interpretability and conversational effectiveness would also be a crucial next step, alongside evaluating the system's performance on more contemporary and diverse real-world datasets to ensure broad applicability.

Full Text

Comments

You need to be logged in to post a comment.