Hierarchical User Intent Graph Network for Multimedia Recommendation

A novel framework that learns multi-level user intents from co-interacted patterns of items, exhibiting user intents in a hierarchical graph structure to enhance multimedia recommendation.

Authors

Yinwei Wei¹, Xiang Wang², Xiangnan He³, Liqiang Nie¹*, Yong Rui⁴, Tat-Seng Chua²

¹ Shandong University, China
² National University of Singapore, Singapore
³ University of Science and Technology of China, China
⁴ Lenovo Research, China
* Corresponding author

---

Links

• Paper: TMM'21
• Code Repository: GitHub

---

Updates

• [10/2021] Paper accepted to IEEE Transactions on Multimedia (TMM).

---

Introduction

This is the official PyTorch implementation for the paper Hierarchical User Intent Graph Network for Multimedia Recommendation.

Understanding user preference on item context is the key to acquiring high-quality multimedia recommendations. However, present works largely leave user intents untouched. In this work, we aim to learn multi-level user intents from the co-interacted patterns of items to obtain high-quality representations of users and items.

We develop a novel framework, Hierarchical User Intent Graph Network (HUIGN), which exhibits user intents in a hierarchical graph structure, from fine-grained to coarse-grained intents. It utilizes intra-level and inter-level aggregations to model user behaviors and refine representations as a distribution over discovered intents, achieving significant improvements over state-of-the-art methods like MMGCN and DisenGCN.

---

Highlights

• Proposes a novel Hierarchical User Intent Graph Network to explicitly model multi-level user intents.
• Employs intra-level aggregation to distill intent signals and inter-level aggregation to model coarser-grained supernodes.
• Refines user and item representations as a distribution over discovered intents rather than relying solely on pre-existing features.
• Provides semantic visualization of user intents through item representations.

---

Method / Framework

---

Installation

The code has been tested running under Python 3.5.2. The required packages are as follows:

• Pytorch == 1.1.0
• torch-cluster == 1.4.2
• torch-geometric == 1.2.1
• torch-scatter == 1.2.0
• torch-sparse == 0.4.0
• numpy == 1.16.0

---

Dataset / Benchmark

We provide and test our model on three processed datasets: Movielens, Tiktok, and Kwai.

• You can find the full version of recommendation datasets via their official platforms (Kwai, Tiktok, and Movielens).

File Descriptions

• train.npy: Train file. Each line is a pair of one user and one of her/his positive items: (userID, micro-video ID).
• val_full.npy: Validation file. Each line is a user with her/his positive interactions with items: (userID, micro-video ID).
• test_full.npy: Test file. Each line is a user with her/his positive interactions with items: (userID, micro-video ID).

---

Usage

The instruction of commands has been clearly stated in the codes. Run the following examples to start training the model on different datasets:

Movielens Dataset

python main.py --data_path 'Movielens' --l_r 0.0001 --weight_decay 0.0001 --batch_size 1024 --dim_x 64 --num_workers 30 --topK 10 --cluster_list 32 8 4

Tiktok Dataset

python train.py --data_path 'Tiktok' --l_r 0.0005 --weight_decay 0.1 --batch_size 1024 --dim_latent 64 --num_workers 30 --topK 10 --cluster_list 32 8 4

Kwai Dataset

python train.py --data_path 'Kwai' --l_r 0.0005 --weight_decay 0.1 --batch_size 1024 --dim_latent 64 --num_workers 30 --topK 10 --cluster_list 32 8 4

Citation If you use this code or our proposed model in your research, please consider citing:

@article{wei2021hierarchical,
  title={Hierarchical User Intent Graph Network for Multimedia Recommendation},
  author={Wei, Yinwei and Wang, Xiang   and He, Xiangnan and Nie, Liqiang and Rui, Yong and Chua, Tat-Seng},
  journal={IEEE Transactions on Multimedia},
  year={2021},
  publisher={IEEE}
}