Overview
This research introduces BARGE, a system designed to adapt autoregressive generation for sequential recommendation. The primary objective of BARGE is to bridge structural gaps identified in current generative recommendation (GR) paradigms, specifically those related to the flattening of multi-token item IDs and the inconsistency between training and inference in hierarchical codebooks. The system incorporates Item Context-Aware Attention (ICA), Hierarchical Path Reranking (HPR), and Dual-Path Decoding (DPD) to address these issues.
Research Context
Generative Recommendation (GR) represents a paradigm for sequential recommendation. Within this paradigm, a common approach involves encoding items into hierarchical semantic IDs through residual quantization. These IDs are then predicted token by token. However, the existing generative formulation presents two structural challenges with respect to the recommendation task. First, the process of flattening multi-token IDs into a single sequence can disrupt or destroy the inherent item-level structure. Second, an inconsistency between the training and inference phases, particularly concerning a hierarchical codebook, can lead to semantic drift. BARGE was developed to specifically address and mitigate these two aforementioned issues.
Approach
BARGE employs a multi-component strategy to bridge the identified structural gaps in generative recommendation. To restore item-level structure during the encoding phase, BARGE utilizes Item Context-Aware Attention (ICA). This mechanism is designed to incorporate contextual information relevant to individual items, preserving their structural relationships that might otherwise be lost when multi-token IDs are linearized. During the decoding phase, BARGE deploys two complementary mechanisms to suppress semantic drift: Hierarchical Path Reranking (HPR) and Dual-Path Decoding (DPD). HPR aims to re-evaluate and optimize paths within the hierarchical codebook, while DPD provides an alternative decoding pathway. These two components work in conjunction to reduce discrepancies between training and inference with the hierarchical codebook.
Findings
- Extensive experiments conducted on public benchmarks and a large-scale offline test indicated that BARGE achieved superior recommendation performance.
- An online A/B test implemented on a Tencent platform resulted in a 0.60% improvement in click-through rate.
- The same online A/B test demonstrated a 1.34% improvement in click unique visitors.
- The online A/B test also recorded a 1.70% improvement in total reading time.
Why This Matters
The observed improvements in click-through rate, click unique visitors, and total reading time from the online A/B test on a Tencent platform confirm the practical usefulness of BARGE in an industrial-scale recommendation context. These results suggest that the proposed methods for bridging structural gaps in generative recommendation can lead to tangible benefits in user engagement metrics for real-world applications.
The paper was announced via arXiv:2607.21028v2, indicating a replacement update to a previous version.