SC-Rec: Enhancing Generative Retrieval with Self-Consistent Reranking for Sequential Recommendation (2408.08686v2)

Abstract: LLMs (LMs) are increasingly employed in recommendation systems due to their advanced language understanding and generation capabilities. Recent recommender systems based on generative retrieval have leveraged the inferential abilities of LMs to directly generate the index tokens of the next item, based on item sequences within the user's interaction history. Previous studies have mostly focused on item indices based solely on textual semantic or collaborative information. However, although the standalone effectiveness of these aspects has been demonstrated, the integration of this information has remained unexplored. Our in-depth analysis finds that there is a significant difference in the knowledge captured by the model from heterogeneous item indices and diverse input prompts, which can have a high potential for complementarity. In this paper, we propose SC-Rec, a unified recommender system that learns diverse preference knowledge from two distinct item indices and multiple prompt templates. Furthermore, SC-Rec adopts a novel reranking strategy that aggregates a set of ranking results, inferred based on different indices and prompts, to achieve the self-consistency of the model. Our empirical evaluation on three real-world datasets demonstrates that SC-Rec considerably outperforms the state-of-the-art methods for sequential recommendation, effectively incorporating complementary knowledge from varied outputs of the model.

• The paper proposes a novel self-consistency reranking strategy that integrates diverse prompt templates and item indices for sequential recommendation.
• The methodology employs hierarchical quantization with collaborative and semantic embeddings to capture complementary knowledge.
• Empirical results on real-world datasets show up to a 22.45% improvement in Hit@10 over state-of-the-art methods.

SC-Rec: Enhancing Generative Retrieval with Self-Consistent Reranking for Sequential Recommendation

The paper "SC-Rec: Enhancing Generative Retrieval with Self-Consistent Reranking for Sequential Recommendation" presents a nuanced approach to improving sequential recommendation systems by leveraging diverse prompt templates and heterogeneous item indices. The core innovation lies in the SC-Rec framework, which aims to integrate complementary knowledge derived from varied outputs of a recommendation model into a coherent, self-consistent system.

Background and Motivation

Sequential recommendations aim to predict a user's next interaction by capturing the user's historical interaction patterns. With the advancements in generative LMs, recommender systems have started to leverage these models to generate item sequences directly. Traditional methods have focused on generating item indices based on either textual semantics or collaborative information. However, the standalone effectiveness of these aspects, while demonstrated, lacks exploration in integrated scenarios.

Contributions

The authors identify a significant knowledge gap in existing literature—how diverse item indices and prompt templates can be integrated into a unified system for sequential recommendation, thus extracting complementary knowledge from these varied inputs. The primary contributions of the paper are:

1. Unified Model for Diverse Indices and Prompts: SC-Rec integrates two distinct item indices and multiple prompt templates to enhance the generative retrieval process.
2. Self-Consistency Reranking Strategy: The paper introduces a novel reranking strategy that aggregates various ranking results based on different indices and prompts, achieving self-consistency in model predictions.
3. Empirical Validation: Extensive experiments on three real-world datasets demonstrate SC-Rec's superior performance compared to state-of-the-art methods in sequential recommendation.

Methodology

Item Multi-Index Generation

The authors generate two types of item indices based on collaborative (CeID) and semantic embeddings (SeID) through a hierarchical quantization process using RQ-VAE. This process ensures rich and diverse item representations, crucial for capturing complementary knowledge.

Multi-Index Recommender Training

The P5 model, fine-tuned for sequential recommendation, serves as the backbone of SC-Rec. This model employs prompt templates to encode user interactions and generates predictions based on the provided item indices. The integration of heterogeneous indices into a single model allows the system to learn and predict based on a richer context.

Reranking Strategy

The reranking strategy is pivotal to SC-Rec's success. It uses self-consistency scores derived from "confidence" and "consistency" metrics. Confidence measures how often an item is ranked near the top, while consistency measures the stability of an item's ranking across different prompts. By combining these metrics, SC-Rec aggregates the most reliable recommendations from diverse outputs, ensuring robust performance.

Empirical Results

The SC-Rec framework is validated against three datasets: Amazon Beauty, Amazon Sports, and Yelp. In these experiments, SC-Rec consistently outperformed state-of-the-art methods, demonstrating its effectiveness in integrating complementary knowledge from diverse sources. Specifically, SC-Rec showed an improvement of up to 22.45% in Hit@10 over the best baseline method.

Analysis and Future Directions

The analysis section underscores the significant impact of using multiple prompt templates and heterogeneous item indices. The results show a strong positive correlation between the number of prompts and recommendation performance, supporting the strategy of leveraging diverse inputs.

The paper opens avenues for future research to explore computational efficiency aspects of SC-Rec, given the higher computational costs associated with its extensive reranking strategy. Moreover, there is potential for SC-Rec to be adapted across various domains beyond recommendation systems, benefiting from the integration of diverse information sources.

Conclusion

SC-Rec represents a significant advancement in the field of sequential recommendation by effectively integrating heterogeneous information into a unified model. The framework's novel use of self-consistency for reranking iteratively refines predictions, ensuring that the final recommendations are both accurate and reliable. This paper provides a meaningful contribution to the literature, setting a foundation for further exploration into multi-source integration in recommendation systems.