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

Tongyoung Kim, Soojin Yoon, Seongku Kang, Jinyoung Yeo, Dongha LeearXiv 2024 · arXiv · OpenReview

Abstract

Language Models (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 (implemented as the MVIGER framework), 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.

Introduction

Generative retrieval has emerged as a powerful paradigm for sequential recommendation, where Large Language Models (LLMs) are trained to generate item identifiers directly from a user's interaction history. However, these systems often suffer from a "consistency crisis": minor variations in prompt phrasing or item indexing (e.g., using semantic descriptions vs. collaborative purchase patterns) can lead to significantly different recommendations.

Rather than viewing this inconsistency as a flaw, the MVIGER (Multi-View Variational Integration of GEnerative Recommender) framework treats these diverse outputs as "views" containing complementary knowledge. For instance, one view might excel at capturing brand loyalty through semantic metadata, while another identifies niche interests through collaborative filtering. MVIGER systematically integrates these perspectives to produce more robust and accurate suggestions.

Methodology: The MVIGER Framework

MVIGER introduces a unified variational framework to manage multi-view information through a three-step process: index generation, multi-view training, and adaptive inference.

1. Heterogeneous Indexing

To capture diverse item characteristics, the framework employs a Residual Quantized Variational Autoencoder (RQ-VAE) to create two distinct types of item IDs:

  • Collaborative Index (CeID): Derived from user-item interaction graphs using LightGCN, capturing behavioral similarities.
  • Semantic Index (SeID): Derived from item metadata (titles, descriptions) using Sentence-T5, capturing content-based similarities.

2. Variational Integration

The framework treats the selection of a "view" (a specific combination of prompt template and index type) as a categorical latent variable ($z$).

  • Prior Network: Learns to predict which view is most relevant for a given user's history.
  • Likelihood Network: The underlying LLM that generates the item ID conditioned on the selected view.
Overview of the MVIGER framework, detailing the three-step process of index generation, training, and multi-view inference.

3. Flexible Inference

At test time, the model supports two modes:

  • Single View ($z^*$): Efficiently selects the most probable view according to the prior network.
  • Multi-View Aggregation: Generates predictions across all views and weights them by the prior to achieve self-consistency, maximizing accuracy at the cost of higher latency.

Experimental Evaluation

MVIGER was evaluated on three standard benchmarks—Amazon Beauty, Amazon Sports, and Yelp—using a leave-one-out protocol. The model's performance was compared against a wide array of baselines, including traditional collaborative filtering (LightGCN), sequential models (SASRec), and state-of-the-art generative models (P5, TIGER, EAGER).

Key Results

As shown in the performance table below, MVIGER consistently outperforms existing generative baselines. On the Amazon Beauty dataset, it achieved a Hit@10 of 0.0878, surpassing the previous state-of-the-art, EAGER (0.0836). Notably, even the "Single View" inference mode ($z^*$) often outperformed fixed-index baselines, demonstrating that the multi-view training process itself enhances model robustness.

Performance comparison of MVIGER against state-of-the-art recommendation methods across three real-world datasets.

Discussion: The Value of Complementarity

A critical insight from the study is the high degree of complementarity between different views. Analysis revealed that less than 50% of correctly predicted items were shared between the collaborative (CeID) and semantic (SeID) index types. This confirms that single-index models are fundamentally limited in the types of user preferences they can capture. Furthermore, the model showed steady performance gains as the number of prompt templates increased, proving its ability to extract value from diverse linguistic contexts.

Conclusion

MVIGER addresses the sensitivity of generative recommenders by embracing multi-view diversity. By adaptively weighting different prompt-index combinations through a variational framework, it effectively bridges the gap between semantic and collaborative knowledge. Future work will explore expanding this framework to include visual and temporal indices to further enrich the recommendation views.

Summary updated

Abstract

Language Models (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.
BibTeX
@misc{kim2024sc,
  author = {Tongyoung Kim and Soojin Yoon and Seongku Kang and Jinyoung Yeo and Dongha Lee},
  title = {{SC-Rec: Enhancing Generative Retrieval with Self-Consistent Reranking for Sequential Recommendation}},
  year = {2024},
  eprint = {2408.08686},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2408.08686},
}
Figure