LANE: Logic Alignment of Non-tuning Large Language Models and Online Recommendation Systems for Explainable Reason Generation

TL;DR

This work tackles the cost and alignment barriers of using large language models to explain recommendations by introducing LANE, a framework that aligns LLM-based explanations with an existing sequential recommender without tuning the LLMs. LANE combines semantic embeddings of item titles, zero-shot multi-preference extraction via prompting, semantic alignment through multi-head attention, and Chain-of-Thought prompting to generate coherent explanations. Key contributions include a text-based semantic embedding module, an integrated SASRec-based model, a multi-preference generation module, a semantic alignment module with residual connections, a prediction module, and a CoT-based explanation generation module; experiments on MovieLens-1M, Amazon-Beauty, and Steam show improved recommendation performance and high-quality explanations. This approach enables leveraging powerful proprietary LLMs for explainability with reduced training costs, potentially broadening the practical deployment of explainable recommendation systems.

Abstract

The explainability of recommendation systems is crucial for enhancing user trust and satisfaction. Leveraging large language models (LLMs) offers new opportunities for comprehensive recommendation logic generation. However, in existing related studies, fine-tuning LLM models for recommendation tasks incurs high computational costs and alignment issues with existing systems, limiting the application potential of proven proprietary/closed-source LLM models, such as GPT-4. In this work, our proposed effective strategy LANE aligns LLMs with online recommendation systems without additional LLMs tuning, reducing costs and improving explainability. This innovative approach addresses key challenges in integrating language models with recommendation systems while fully utilizing the capabilities of powerful proprietary models. Specifically, our strategy operates through several key components: semantic embedding, user multi-preference extraction using zero-shot prompting, semantic alignment, and explainable recommendation generation using Chain of Thought (CoT) prompting. By embedding item titles instead of IDs and utilizing multi-head attention mechanisms, our approach aligns the semantic features of user preferences with those of candidate items, ensuring coherent and user-aligned recommendations. Sufficient experimental results including performance comparison, questionnaire voting, and visualization cases prove that our method can not only ensure recommendation performance, but also provide easy-to-understand and reasonable recommendation logic.

Figures (8)

• Figure 1: The overview of LANE. It consists of six crucial components: (I) semantic embedding module, (II) integrated model module, (III) users' multi-preference generation module, (IV) semantic alignment module, (V) prediction module, and (VI) explainable recommendation generation module.
• Figure 2: The zero-shot prompt template. It consists of five components and fills in a user interaction sequence in the Steam dataset as an example, where $n$ refers to the number of user preferences.
• Figure 3: An example of users' multi-preferences. It was generated by GPT under the guidance of a zero-shot prompt template, where the number of user preferences $n$ is equal to 5.
• Figure 4: The CoT prompt template. It mainly consists of four progressive steps, and needs to fill in the user's interaction sequence, target item, user‘s multi-preferences and attention weight, also taking the data on the Steam dataset as an example.
• Figure 5: The experimental results of the sensitivity analysis on the ML-1M dataset for the four hyperparameters: (a) - (d) hidden size$d_k$, (e) - (h) number of heads $h$, (i) - (l) maximum sequence length $n$, (m) - (p) and number of user preferences $m$. The evaluation metrics used are NDCG@5, NDCG@10, HitRate@5, and HitRate@10, .