This is our Pytorch implementation for the paper: Hierarchical User Intent Graph Network for Multimedia Recommendation

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

The instruction of commands has been clearly stated in the codes.

• 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

Some important arguments:

has_ind: It indicates the optional independence loss function.

has_cro: It indicates the optional cross_entropy loss function.

has_v, has_a, and has_t: They are used to indicate which modalities are included in this work.

--num_links: It indicates the number of co-occurrence.

--cluster_list: It describes the structure of hierarchical user intents.

We provide three processed datasets: Movielnes, Tiktok, and Kwai.

• You can find the full version of recommendation datasets via Kwai, Tiktok, and Movielens.

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