Guangyue Xu

I am currently a Senior Data Scientist at Search@Target. Previously, I pursued a Ph.D. at Michigan State University, where I focused on Machine Learning (ML) and Natural Language Processing (NLP). My research centers on pre-training large vision-language models and enhancing their generalization capabilities, with applications in e-commerce search.

I received my B.E. in Software Engineering from Jilin University and M.S. in Computer Science from Tsinghua University. I also interned in MSRA's Web Search and Mining Group.

Email CV Google Scholar GitHub
News
Mar 2026 Paper New preprint: Beyond Text: Aligning Vision and Language for Multimodal E-Commerce Retrieval is on arXiv.
Feb 2026 Paper New preprint: Unified Learning-to-Rank for Multi-Channel Retrieval in Large-Scale E-Commerce Search is on arXiv.
Posts
Beyond Text: How We Built Multimodal Retrieval for E-Commerce Search
Mar 10, 2026  ·  multimodal e-commerce
Most e-commerce search systems rely purely on text — but product images carry signals that text often misses. We propose a two-stage vision-language alignment framework achieving +5% Recall@100 over text-only baselines.
Multi-Channel Retrieval Fusion: From Heuristic Weights to Unified Ranking
May 2, 2025  ·  Search Learning-to-Rank
E-commerce search pulls candidates from text, semantic, and behavioral channels — then merges them with hand-tuned weights. We replace that with a unified LTR model, deployed at Target.com with +2.85% conversion lift.
View all posts →
Selected Publications
Multimodal Retrieval
Beyond Text: Aligning Vision and Language for Multimodal E-Commerce Retrieval New
Guangyue Xu*, Qujiaheng Zhang*, Fengjie Li  (* co-first authors)
arXiv preprint arXiv:2603.04836, 2026
arXiv

We study unified text-image fusion for two-tower retrieval models in e-commerce, proposing a novel modality fusion network that captures cross-modal complementary information between product text and images.

Multi-Channel Retrieval
Unified Learning-to-Rank for Multi-Channel Retrieval in Large-Scale E-Commerce Search New
Guangyue Xu*, A. Gaydhani*, D. Kamath, A. Singh, A. Li  (* co-first authors)
arXiv preprint arXiv:2602.23530, 2026
arXiv

A unified learning-to-rank framework that jointly optimizes ranking across multiple retrieval channels in large-scale e-commerce search, improving relevance and efficiency.

GIPCOL
GIPCOL: Graph-Injected Soft Prompting for Compositional Zero-Shot Learning
Guangyue Xu, Parisa Kordjamshid, Joyce Chai
WACV, 2024
project page arXiv

We introduce a GNN into soft-prompting design to improve CLIP's compositional zero-shot learning ability.

MetaReVision
MetaReVision: Meta-Learning with Retrieval for Visually Grounded Compositional Concept Acquisition
Guangyue Xu, Parisa Kordjamshid, Joyce Chai
EMNLP-Findings, 2023
project page arXiv

We meta-train vision-language models using retrieved items to obtain more generalizable token representations and improve compositional ability.

Prompting
Prompting Large Pre-Trained Vision-Language Models for Compositional Concept Learning
Guangyue Xu, Parisa Kordjamshid, Joyce Chai
arXiv, 2022
arXiv

We systematically investigate various prompting techniques for CLIP in compositional zero-shot learning.

Full list on Google Scholar →
Service
Program Committee / Reviewer: EACL, EMNLP, ACM MM, ACL, WACV, LREC, ICLR, NeurIPS, ...