RecGOAT: Graph Optimal Adaptive Transport for LLM-Enhanced Multimodal Recommendation with Dual Semantic Alignment

PyTorch implementation for our paper RecGOAT: Graph Optimal Adaptive Transport for LLM-Enhanced Multimodal Recommendation with Dual Semantic Alignment.

Yuecheng Li, Hengwei Ju, Zeyu Song, Wei Yang, Chi Lu, Peng Jiang, and Kun Gai.

Kuaishou Inc, Fudan University, and University of Southern California

We will release the code, data, and LLM-enhanced multimodal data of RecGOAT after the paper is accepted and upon passing company review!

---

We propose RecGOAT, a novel yet simple dual semantic alignment framework for LLM-enhanced multimodal recommendation, which offers theoretically guaranteed alignment capability. RecGOAT first employs graph attention networks to enrich collaborative semantics by modeling item-item, user-item, and user-user relationships, leveraging user/item LM representations and interaction history. Furthermore, we design a dual-granularity progressive multimodality-ID alignment framework, which achieves instance-level and distribution-level semantic alignment via cross-modal contrastive learning (CMCL) and optimal adaptive transport (OAT), respectively.

---

📢 News

• [2026.08.23] 🔥🔥 The codebase of our RecGOAT is available!

• [2026.02.03] 🔥🔥 The full paper of our RecGOAT is available at arXiv.

Dependence

To install the dependencies:

pip install -r requirements.txt

Usage and Hyperparameter

# Baby
CUDA_VISIBLE_DEVICES=0 python -u main.py --dataset baby_raw --model lgn_mm --use_multimodal 1 --text_feat ./data/baby_raw/text_feat.npy --image_feat ./data/baby_raw/image_feat.npy --item_knn_k 20 --item_branch_layers 1 --epochs 1000 --recdim 800 --layer 3 --lr 8e-4 --bpr_batch 4096 --topks "[10]" --fusion oat --contrastive_weight 0.032 > ./baby.log 2>&1 &

# Sports
CUDA_VISIBLE_DEVICES=0 python -u main.py --dataset sports_raw --model lgn_mm --use_multimodal 1 --text_feat ./data/sports_raw/text_feat.npy --image_feat ./data/sports_raw/image_feat.npy --item_knn_k 30 --item_branch_layers 1 --epochs 1000 --recdim 600 --layer 3 --lr 8e-4 --bpr_batch 4096 --topks "[10]" --fusion oat --contrastive_weight 0.005 > ./sports.log 2>&1 &

# Electronics
CUDA_VISIBLE_DEVICES=0 python -u main.py --dataset electronics --model lgn_mm --use_multimodal 1 --text_feat ./data/electronics_raw/text_feat.npy --image_feat ./data/electronics_raw/image_feat.npy --item_knn_k 20 --item_branch_layers 1 --epochs 2000 --recdim 200 --layer 3 --lr 8e-4 --bpr_batch 4096 --topks "[10]" --fusion oat --contrastive_weight 0.05 > ./electronics.log 2>&1 &

👉 TODO

• Release the code of RecGOAT.

• Release the LLM-enhanced multimodal data of RecGOAT.

• ...

Citation

@article{li2026recgoat,
  title={RecGOAT: Graph Optimal Adaptive Transport for LLM-Enhanced Multimodal Recommendation with Dual Semantic Alignment},
  author={Li, Yuecheng and Ju, Hengwei and Song, Zeyu and Yang, Wei and Lu, Chi and Jiang, Peng and Gai, Kun},
  journal={arXiv preprint arXiv:2602.00682},
  year={2026}
}