This is our Pytorch implementation for the paper:

Yinwei Wei, Xiang Wang, Liqiang Nie, Xiangnan He, Richang Hong, and Tat-Seng Chua(2019). MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video. In ACM MM`19, NICE, France,Oct. 21-25, 2019

Multi-modal Graph Convolution Network is a novel multi-modal recommendation framework based on graph convolutional networks, explicitly modeling modal-specific user preferences to enhance micro-video 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.

• Kwai dataset
  python train.py --model_name='MMGCN' --l_r=0.0005 --weight_decay=0.1 --batch_size=1024 --dim_latent=64 --num_workers=30 --aggr_mode='mean' --num_layer=2 --concat=False
• Tiktok dataset
  python train.py --model_name='MMGCN' --l_r=0.0005 --weight_decay=0.1 --batch_size=1024 --dim_latent=64 --num_workers=30 --aggr_mode='mean' --num_layer=2 --concat=False
• Movielens dataset
  python train.py --model_name='MMGCN' --l_r=0.0001 --weight_decay=0.0001 --batch_size=1024 --dim_latent=64 --num_workers=30 --aggr_mode='mean' --num_layer=2 --concat=False

Some important arguments:

• model_name: It specifies the type of model. Here we provide three options:

  1. MMGCN (by default) proposed in MMGCN: Multi-modal Graph Convolution Network for Personalized Recommendation of Micro-video, ACM MM2019. Usage: --model_name='MMGCN'
  2. VBPR proposed in VBPR: Visual Bayesian Personalized Ranking from Implicit Feedback, AAAI2016. Usage: --model_name 'VBPR'
  3. ACF proposed in Attentive Collaborative Filtering: Multimedia Recommendation with Item- and Component-Level Attention , SIGIR2017. Usage: --model_name 'ACF'
  4. GraphSAGE proposed in Inductive Representation Learning on Large Graphs, NIPS2017. Usage: --model_name 'GraphSAGE'
  5. NGCF proposed in Neural Graph Collaborative Filtering, SIGIR2019. Usage: --model_name 'NGCF'

• aggr_mode It specifics the type of aggregation layer. Here we provide three options:

  1. mean (by default) implements the mean aggregation in aggregation layer. Usage --aggr_mode 'mean'
  2. max implements the max aggregation in aggregation layer. Usage --aggr_mode 'max'
  3. add implements the sum aggregation in aggregation layer. Usage --aggr_mode 'add'

• concat: It indicates the type of combination layer. Here we provide two options:

  1. concat(by default) implements the concatenation combination in combination layer. Usage --concat 'True'
  2. ele implements the element-wise combination in combination layer. Usage --concat 'False'

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

• You can find the full version of recommendation datasets via Kwai, Tiktok, and Movielens.
• We select some users and micro-videos in Kwai and Tiktok datasets accoding to the timestamp.
• We extract the visual, acoustic, and textual features of all trailers in Movielens dataset.

-train.npy Train file. Each line is a user with her/his positive interactions with items: (userID and micro-video ID)
-val.npy Validation file. Each line is a user with her/his 1,000 negative and several positive interactions with items: (userID and micro-video ID)
-test.npy Test file. Each line is a user with her/his 1,000 negative and several positive interactions with items: (userID and micro-video ID)