HSTU-BLaIR: Lightweight Contrastive Text Embedding for Generative Recommender

🏷️ Accepted at the Workshop on Large Language Models for E-Commerce, KDD 2025

🌱 Introduction

Recent advances in recommender systems have underscored the complementary strengths of generative modeling and pretrained language models. We propose HSTU-BLaIR, a hybrid framework that augments the Hierarchical Sequential Transduction Unit (HSTU)-based generative recommender with BLaIR, a lightweight contrastive text embedding model. This integration enriches item representations with semantic signals from textual metadata while preserving HSTU’s powerful sequence modeling capabilities.

We evaluate HSTU-BLaIR on two e-commerce datasets: three subsets from the Amazon Reviews 2023 dataset and the Steam dataset. We compare its performance against both the original HSTU-based recommender and a variant augmented with embeddings from OpenAI’s state-of-the-art text-embedding-3-large model. Despite the latter being trained on a substantially larger corpus with significantly more parameters, our lightweight BLaIR-enhanced approach—pretrained on domain-specific data—achieves better performance in nearly all cases. Specifically, HSTU-BLaIR outperforms the OpenAI embedding-based variant on all but one metric, where it is marginally lower, and matches it on another. These findings highlight the effectiveness of contrastive text embeddings in compute-efficient recommendation settings.

📄 The full technical report can be found here.

---

🚀 Getting Started

Install the required Python packages with pip3 install -r requirements.txt.

HSTU-BLaIR is built on top of the HSTU-based generative recommender (commit ece916f) and extends it with additional modules for integrating textual embeddings. It has been tested on Ubuntu 22.04 with Python 3.9, CUDA 12.6, and a single NVIDIA RTX 4090 GPU.

---

🧪 Experiments

To reproduce the experiments from our technical report, follow these steps:

1. Download and preprocess the data

mkdir -p tmp/ && python3 preprocess_public_data.py

To use OpenAI text-embedding-3-large embeddings instead of BLaIR, update the corresponding line in preprocess_public_data.py as follows:

text_embedding_model = "blair"

to

text_embedding_model = "openai"

Make sure you have correctly set up your OpenAI API credentials if you choose this option.

2. Run the model

CUDA_VISIBLE_DEVICES=0 python3 main.py --gin_config_file=configs/amzn23_game/hstu-sampled-softmax-n512-blair.gin --master_port=12345

You can find other configuration files in the configs/ directory for different settings and domains.

---

📊 Results

Evaluation metrics on Video Games, Office Products, and Musical Instruments of Amazon Reviews 2023 dataset and Steam dataset

With the provided .gin files, you should be able to reproduce the following evaluation results within a small margin of variability. Minor differences may arise due to inherent non-determinism in model initialization, data loading, and/or hardware-level operations.

Each cell shows absolute values (top row) and percentage improvements over HSTU / SASRec (bottom row). Best values per column are bolded.

HSTU-OpenAI (TE3L) denotes the HSTU model augmented with text embeddings generated via OpenAI’s text-embedding-3-large API.

Video Games

Model	HR@10	HR@50	HR@200	NDCG@10	NDCG@200	MRR
SASRec	.1028	.2317	.3941	.0573	.1097	.0518
	—	—	—	—	—	—
HSTU	.1315	.2765	.4565	.0741	.1327	.0658
	(+28%)	(+19.3%)	(+15.8%)	(+29.3%)	(+21.0%)	(+27.1%)
HSTU-OpenAI (TE3L)	.1328	.2821	.4645	.0742	.1341	.0658
	(+1.0% / +29.2%)	(+2.0% / +21.8%)	(+1.8% / +17.9%)	(+0.1% / +29.5%)	(+1.1% / +22.2%)	(0.0% / +27.0%)
HSTU-BLaIR	.1353	.2852	.4684	.0760	.1361	.0674
	(+2.9% / +31.6%)	(+3.1% / +23.1%)	(+2.6% / +18.9%)	(+2.6% / +32.6%)	(+2.6% / +24.1%)	(+2.4% / +30.1%)

Office Products

Model	HR@10	HR@50	HR@200	NDCG@10	NDCG@200	MRR
SASRec	.0281	.0668	.1331	.0153	.0335	.0143
	—	—	—	—	—	—
HSTU	.0395	.0880	.1649	.0223	.0443	.0207
	(+40.6%)	(+31.7%)	(+23.9%)	(+45.8%)	(+32.2%)	(+44.8%)
HSTU-OpenAI (TE3L)	.0477	.1050	.1940	.0269	.0526	.0247
	(+20.8% / +69.8%)	(+19.3% / +57.2%)	(+17.6% / +45.8%)	(+20.6% / +75.8%)	(+18.7% / +57.0%)	(+19.3% / +72.7%)
HSTU-BLaIR	.0484	.1068	.1946	.0271	.0529	.0248
	(+22.5% / +72.2%)	(+21.4% / +59.9%)	(+18.0% / +46.2%)	(+21.5% / +77.1%)	(+19.4% / +57.9%)	(+19.8% / +73.4%)

Musical Instruments

Model	HR@10	HR@50	HR@200	NDCG@10	NDCG@200	MRR
SASRec	.0643	.1488	.2784	.0356	.0731	.0326
	—	—	—	—	—	—
HSTU	.0700	.1599	.2910	.0392	.0783	.0359
	(+8.9%)	(+7.5%)	(+4.5%)	(+10.1%)	(+7.1%)	(+10.1%)
HSTU-OpenAI (TE3L)	.0708	.1635	.3005	.0393	.0798	.0360
	(+1.1% / +10.1%)	(+2.3% / +9.9%)	(+3.3% / +7.9%)	(+0.3% / +10.4%)	(+1.9% / +9.2%)	(+0.3% / +10.4%)
HSTU-BLaIR	.0733	.1681	.3066	.0406	.0818	.0371
	(+4.7% / +14.0%)	(+5.1% / +13.0%)	(+5.4% / +10.1%)	(+3.6% / +14.0%)	(+4.5% / +11.9%)	(+3.3% / +13.8%)

Steam

Model	HR@10	HR@50	HR@200	NDCG@10	NDCG@200	MRR
SASRec	.0881	.2277	.4312	.0462	.1068	.0426
	—	—	—	—	—	—
HSTU	.1038	.2575	.4704	.0544	.1195	.0492
	(+17.8%)	(+13.1%)	(+9.1%)	(+17.8%)	(+11.9%)	(+15.5%)
HSTU-OpenAI (TE3L)	.1089	.2657	.4806	.0579	.1241	.0525
	(+4.9% / +23.6%)	(+3.2% / +16.7%)	(+2.2% / +11.5%)	(+6.4% / +25.3%)	(+3.8% / +16.2%)	(+6.7% / +23.2%)
HSTU-BLaIR	.1100	.2668	.4812	.0581	.1241	.0523
	(+6.0% / +24.9%)	(+3.6% / +17.2%)	(+2.3% / +11.6%)	(+6.8% / +25.8%)	(+3.8% / +16.2%)	(+6.3% / +22.8%)

📚 Citation

If you use HSTU-BLaIR in your work, please cite:

@article{hstublair2025,
  title     = {Integrating Textual Embeddings from Contrastive Learning with Generative Recommender for Enhanced Personalization},
  author    = {Yijun Liu},
  journal   = {arXiv preprint arXiv:2504.10545},
  year      = {2025},
  url       = {https://arxiv.org/abs/2504.10545}
}

📬 Contact

If you encounter any problems or have questions, feel free to open an issue or email yijunl [at] usc [dot] edu.

License

This codebase is Apache 2.0 licensed, as found in the LICENSE file.