Inductive Generative Recommendation via Retrieval-based Speculation (2410.02939v1)

Abstract: Generative recommendation (GR) is an emerging paradigm that tokenizes items into discrete tokens and learns to autoregressively generate the next tokens as predictions. Although effective, GR models operate in a transductive setting, meaning they can only generate items seen during training without applying heuristic re-ranking strategies. In this paper, we propose SpecGR, a plug-and-play framework that enables GR models to recommend new items in an inductive setting. SpecGR uses a drafter model with inductive capability to propose candidate items, which may include both existing items and new items. The GR model then acts as a verifier, accepting or rejecting candidates while retaining its strong ranking capabilities. We further introduce the guided re-drafting technique to make the proposed candidates more aligned with the outputs of generative recommendation models, improving the verification efficiency. We consider two variants for drafting: (1) using an auxiliary drafter model for better flexibility, or (2) leveraging the GR model's own encoder for parameter-efficient self-drafting. Extensive experiments on three real-world datasets demonstrate that SpecGR exhibits both strong inductive recommendation ability and the best overall performance among the compared methods. Our code is available at: https://github.com/Jamesding000/SpecGR.

• The paper presents SpecGR, a draft-then-verify framework that significantly enhances generative recommendation by enabling inductive item generation.
• It employs an inductive drafting model and target-aware verification using query-likelihood scoring to propose and evaluate candidate items.
• Experimental results on Amazon datasets confirm improved unseen item recommendation and reduced inference time compared to existing models.

Inductive Generative Recommendation via Retrieval-based Speculation

In the paper titled "Inductive Generative Recommendation via Retrieval-based Speculation," the authors propose SpecGR, a framework designed to address the limitations of traditional Generative Recommendation (GR) models, which struggle to recommend new, unseen items without retraining. This paper analyzes the shortcomings of current GR models that operate within a transductive setting, making it difficult for them to generate items that were not included in the training data. Particularly in dynamic fields like e-commerce, where new items frequently emerge, the inefficiency of constantly retraining GR models is evident.

Methodology

SpecGR introduces an innovative draft-then-verify framework. It comprises two main components: a drafter model with inductive capabilities that proposes candidate items, and a GR model that verifies these suggestions using sophisticated ranking techniques.

1. Inductive Drafting: SpecGR employs an inductive drafter model to suggest candidate items—this includes both existing and novel items. This component can utilize either an auxiliary model for flexibility or the GR model’s encoder for parameter-efficient operations.
2. Target-aware Verifying: Once the drafter proposes candidates, the GR model evaluates these using a verification score based on a query-likelihood approach, accepting or rejecting them based on likelihood thresholds optimized during training.
3. Guided Re-drafting: If initial drafts do not fulfill the goal of generating high-quality recommendations, SpecGR can guide the drafter to create new candidates that align more closely with GR model outputs, using semantic ID prefixes from beam search processes.
4. Adaptive Exiting: This function allows SpecGR to conclude its drafting process once the number of suitable recommendations meets a predefined target, thereby enhancing inference efficiency.

Experimental Results

Extensive evaluations on datasets from Amazon's Video Games, Office Products, and Cell Phones categories are presented. These experiments confirm that SpecGR significantly enhances GR models’ ability to recommend new items. Compared to methods like TIGER and various feature-based models, SpecGR demonstrates superior performance, particularly on unseen items, while maintaining robust in-sample capabilities.

In terms of subset ranking—an important metric for real-world deployment—the results showcase SpecGR’s considerable efficiency gains over constrained beam search approaches. The framework manages to significantly reduce inference times without compromising on performance quality, especially as retrieval sizes increase.

Implications and Future Directions

The development of SpecGR marks a significant step towards facilitating more flexible and efficient GR systems. By integrating components with varied paradigms and capabilities, it effectively bridges the capability gap in previous models regarding the recommendation of unseen items. This integration has practical implications for real-time recommendation systems, especially in markets with varying inventory.

Future research may explore scaling SpecGR’s methodology to larger models, potentially unlocking further emergent capabilities in generative recommendation. Additionally, considering the dynamic trade-off between in-sample accuracy and inductive ability, further studies could enhance the balance mechanisms, as real-world applications require adaptable solutions to fluctuating demands.

In conclusion, this framework provides a scalable, efficient, and versatile solution for inductive recommendations, presenting a meaningful advance in generative model-based recommendation systems.