RecHub

A library for GNN-based recommendation system.

Models

Model	Full name	Type	Paper
NCF	Neural Collaborative Filtering	Non-graph	https://arxiv.org/abs/1708.05031
GCN	Graph Convolutional Networks	Homogeneous graph	https://arxiv.org/abs/1609.02907
LightGCN	Light GCN	Homogeneous graph	https://arxiv.org/abs/2002.02126
GAT	Graph Attention Networks	Homogeneous graph	https://arxiv.org/abs/1710.10903
NGCF	Neural Graph Collaborative Filtering	Homogeneous graph	https://arxiv.org/abs/1905.08108
HET-GCN	/	Heterogeneous graph	/
HET-LightGCN	/	Heterogeneous graph	/
HET-GAT	/	Heterogeneous graph	/
HET-NGCF	/	Heterogeneous graph	/

Note: we define the heterogeneous graph as a graph with different types of edges instead of a graph with different types of edges or nodes. Thus, for a common user-item bipartite graph, although more than one types of node exist, we still think it as a homogeneous graph.

WIP

• DeepFM
• DSSM
• DiffNet
• DiffNet++
• DANSER
• GraphRec

Requirements

• Linux-based OS
• Python 3.6+

Get started

Install RecHub

Install from https://pypi.org/:

pip install rechub

Or install manually:

git clone https://github.com/yusanshi/RecHub.git
cd RecHub
pip install .

Install DGL

Note one of the most important dependencies for RecHub, DGL, will not be automatically installed while installing RecHub. You should manually install CPU or CUDA build of DGL.

# This is for CPU version. For CUDA version, use dgl-cu[xxx]
pip install dgl # or dgl-cu92, dgl-cu101, dgl-cu102, dgl-cu110 for CUDA 9.2, 10.1, 10.2, 11.0, respectively.

Check out the instructions on https://www.dgl.ai/pages/start.html for more details.

If there are any problems with later commands, try to install this specific version:

pip install dgl==0.5.3 # or CUDA version: dgl-cu[xxx]==0.5.3

Download the dataset

Here we use the LSEC-Small dataset used in our work LSEC-GNN. It is a dataset featuring live stream E-commerce.

Create an empty directory as our ROOT_DIRECTORY. Then:

# In ROOT_DIRECTORY
mkdir data && cd data
wget https://github.com/yusanshi/LSEC-GNN/files/6520753/LSEC-Small-aa.dummy.gz \
 https://github.com/yusanshi/LSEC-GNN/files/6520754/LSEC-Small-ab.dummy.gz \
 https://github.com/yusanshi/LSEC-GNN/files/6520757/LSEC-Small-ac.dummy.gz \
 https://github.com/yusanshi/LSEC-GNN/files/6520760/LSEC-Small-ad.dummy.gz
cat LSEC-Small-* | tar -xzvf -

Write the metadata file

We use a metadata file to define the nodes, edges for the graph and the tasks. For LSEC-Small dataset, create ROOT_DIRECTORY/metadata/LSEC.json as follows:

{
  "graph": {
    "node": [
      {
        "filename": "item.tsv",
        "attribute": []
      },
      {
        "filename": "user.tsv",
        "attribute": []
      },
      {
        "filename": "streamer.tsv",
        "attribute": []
      }
    ],
    "edge": [
      {
        "filename": "user-item-buy.tsv",
        "weighted": false
      },
      {
        "filename": "user-streamer-follow.tsv",
        "weighted": false
      },
      {
        "filename": "streamer-item-sell.tsv",
        "weighted": false
      }
    ]
  },
  "task": [
    {
      "filename": "user-item-buy.tsv",
      "type": "top-k-recommendation",
      "loss": "binary-cross-entropy",
      "weight": 1
    }
  ]
}

Run

# In ROOT_DIRECTORY

# Train
python -m rechub.train \
    --dataset_path ./data/LSEC-Small/ \
    --metadata_path ./metadata/LSEC.json \
    --model_name HET-GCN \
    --embedding_aggregator concat \
    --predictor mlp \
    --edge_choice 0 1 2 \
    --save_checkpoint True \
    --checkpoint_path ./checkpoint/

# Load latest checkpoint and evaluate on the test set
python -m rechub.test \
    --dataset_path ./data/LSEC-Small/ \
    --metadata_path ./metadata/LSEC.json \
    --model_name HET-GCN \
    --embedding_aggregator concat \
    --predictor mlp \
    --edge_choice 0 1 2 \
    --checkpoint_path ./checkpoint/

You can visualize the metrics with TensorBoard.

tensorboard --logdir runs

Tip: by adding REMARK environment variable, you can make the runs name in TensorBoard and log file name more meaningful. For example, REMARK=lr-0.001_attention-head-10 python -m rechub.train ....

Development

Use your own dataset

Using LSEC-Small dataset as the example, here we demonstrate the dataset format. After this section, you can convert your own dataset into this format.

The LSEC-Small dataset captures the tripartite interaction information in live stream E-commerce scenario. We have three types of nodes: items, users and streamers, and three types of edges: user-item-buy, user-streamer-follow and streamer-item-sell. The structure of the dataset is as follows:

JD-small
├── train
│   ├── user.tsv
│   ├── item.tsv
│   ├── streamer.tsv
│   ├── user-item-buy.tsv
│   ├── user-streamer-follow.tsv
│   └── streamer-item-sell.tsv
├── val
│   └── user-item-buy.tsv
└── test
    └── user-item-buy.tsv

In train, the first three files are node description files and the last three are edge description files.

In node description files are the indexs and other attributes for nodes. In LSEC-Small dataset, there are no other attributes for nodes, but only the basic index information. So the contents of user.tsv, item.tsv and streamer.tsv are:

user
0
1
2
3
4
...

item
0
1
2
3
4
...

streamer
0
1
2
3
4
...

In the edge description files, each line represents an edge. Take user-item-buy.tsv for example, its content is:

user    item
0    9349
0    10535
0    19326
1    555
1    2154
...

In val and test directory, there are edge description files for model evaluation. Different from those in train, they have additional column value indicating the existence of the edge. For example, in val the content of user-item-buy.tsv is:

user    item    value
1    11301    1
1    13353    1
1    15315    1
1    11318    1
1    18206    1
...

TODO

1. Support more models.
2. Support node attributes.
3. Support multiple tasks (e.g., interaction-attribute-regression).

Tricks

• Use this to automatically select the GPU with most free memory.

  alias python='CUDA_VISIBLE_DEVICES=$(nvidia-smi --query-gpu=memory.free --format=csv,nounits,noheader | nl -v 0 | sort -nrk 2 | cut -f 1 | head -n 1 | xargs) python'