SpecGR

This repository provides the official PyTorch implementation for the AAAI 2026 Oral paper "Inductive Generative Recommendation via Retrieval-based Speculation".

Introduction

Speculative Generative Recommendation (SpecGR) is a plug-and-play framework designed to enable any Generative Recommendation (GR) model to recommend new, unseen items in an inductive setting. While GR models excel at predicting and recommending items by generating token sequences, they struggle to recommend new items not encountered during training. SpecGR addresses this limitation with a framework consisting of two key components:

• Drafter Model: A model with strong inductive recommendation capability that proposes 'candidates' (potentially new items).

• Verifier: The GR model then evaluates these candidates, determining which to accept or reject based on their likelihood of being generated.

Building upon this framework, we propose two drafting strategies:

1. SpecGR_Aux: Drafting with an auxiliary model. This strategy uses an external drafter model for enhanced flexibility.
2. SpecGR++: Self-drafting with the GR model's encoder. This approach leverages the GR model's encoder for a parameter-efficient, integrated method. It involves a two-stage training process and requires no modifications to the GR architecture, leading to more efficient recommendations.

Requirements

To install the required dependicies, simply run:

pip install -r requirements.txt

Model Checkpoints and Artifacts

We conducted extensive experiments across three datasets: Video Games, Office Products, and Cell Phones & Accessories. To ensure complete reproducibility and facilitate validation of our experimental results, we provide all necessary artifacts within this repository, enabling reproduction of all experimental results with minimal effort. Due to the 50MB size limit imposed on supplementary material submissions and the prohibition of external links, we include the complete artifact set exclusively for the Video Games dataset as a demonstration. We commit to releasing comprehensive artifacts for all datasets to enable full reproduction of this work upon publication.

The complete experimental artifacts include:

1. Preprocessed datasets: available in the dataset/ directory
2. Trained semantic IDs: located in the semantic_ids/ directory
3. Pre-trained model checkpoints: stored in the results/SpecGR/ directory

Additionally, we provide comprehensive source code implementations with instructive comments for generating these artifacts from scratch, enabling researchers to reproduce our complete training and evaluation pipeline, semantic ID generation process, and data preprocessing steps.

Quick Start

Data Preprocessing

To download and process datasets from scratch, run:

python -m dataset.process_datasets --config CONFIG_PATH --device GPU_ID

Run SpecGR for Recommendation Using Preprocessed Data and Checkpoints

bash quick_start.sh

Training

End-to-End DDP Training for SpecGR++ with PyTorch Lightning

To start end-to-end training, run:

python -m SpecGR.train --config CONFIG_PATH --devices GPU_IDS

Training SpecGR with an Auxiliary Draft Model

Draft Model Options:

• UniSRec: Uses PLM (Pretrained Language Model) text embeddings as universal item representations, achieving strong inductive recommendation performance. Paper

Generative Recommendation Model Options:

• TIGER: Encodes item metadata into semantic IDs and predicts the next item by generating semantic IDs. Paper

For single-GPU end-to-end training of both the draft and generative models, run:

python -m SpecGR_Aux.train --config CONFIG_PATH --device GPU_ID

Inference

Run SpecGR++ for Recommendation

To run SpecGR++ with DDP acceleration for recommendation, run:

python -m SpecGR.run --config CONFIG_PATH --eval_mode EVAL_MODEL --draft_size DRAFT_SIZE --num_beams NUM_BEAMS --threshold THRESHOLD --max_eval_steps MAX_EVAL_STEPS --devices GPUS_IDS

Run SpecGR with an Auxiliary Draft Model for Recommendation

To run SpecGR with an auxiliary draft model for recommendation, run:

python -m SpecGR_Aux.run --config CONFIG_PATH --eval_mode EVAL_MODEL --draft_size DRAFT_SIZE --num_beams NUM_BEAMS --threshold THRESHOLD --max_eval_steps MAX_EVAL_STEPS --device GPU_ID

Citation

If you find our code or paper helpful, please cite our work:

@article{ding2024inductive,
  title={Inductive generative recommendation via retrieval-based speculation},
  author={Ding, Yijie and Li, Jiacheng and McAuley, Julian and Hou, Yupeng},
  journal={arXiv preprint arXiv:2410.02939},
  year={2024}
}