This is our Pytorch implementation for the paper:
Jianxin Chang, Chen Gao, Xiangnan He, Depeng Jin and Yong Li. Bundle Graph Convolutional Network, Paper in ACM DL or Paper in arXiv. In SIGIR'20, Xi'an, China, July 25-30, 2020.
Author: Jianxin Chang ([email protected])
Bundle Graph Convolutional Network (BGCN) is a bundle recommendation solution based on graph neural network, explicitly re-constructing the two kinds of interaction and an affiliation into the graph. With item nodes as the bridge, graph convolutional propagation between user and bundle nodes makes the learned representations capture the item level semantics.
If you want to use our codes and datasets in your research, please cite:
@inproceedings{BGCN20,
author = {Jianxin Chang and
Chen Gao and
Xiangnan He and
Depeng Jin and
Yong Li},
title = {Bundle Recommendation with Graph Convolutional Networks},
booktitle = {Proceedings of the 43nd International {ACM} {SIGIR} Conference on
Research and Development in Information Retrieval, {SIGIR} 2020, Xi'an,
China, July 25-30, 2020.},
year = {2020},
}
The code has been tested running under Python 3.7.0. The required packages are as follows:
- torch == 1.2.0
- numpy == 1.17.4
- scipy == 1.4.1
- temsorboardX == 2.0
The hyperparameter search range and optimal settings have been clearly stated in the codes (see the 'CONFIG' dict in config.py).
- Train
python main.py
- Futher Train
Replace 'sample' from 'simple' to 'hard' in CONFIG and add model file path obtained by Train to 'conti_train', then run
python main.py
- Test
Add model path obtained by Futher Train to 'test' in CONFIG, then run
python eval_main.py
Some important hyperparameters:
-
lrs
- It indicates the learning rates.
- The learning rate is searched in {1e-5, 3e-5, 1e-4, 3e-4, 1e-3, 3e-3}.
-
mess_dropouts
- It indicates the message dropout ratio, which randomly drops out the outgoing messages.
- We search the message dropout within {0, 0.1, 0.3, 0.5}.
-
node_dropouts
- It indicates the node dropout ratio, which randomly blocks a particular node and discard all its outgoing messages.
- We search the node dropout within {0, 0.1, 0.3, 0.5}.
-
decays
- we adopt L2 regularization and use the decays to control the penalty strength.
- L2 regularization term is tuned in {1e-7, 1e-6, 1e-5, 1e-4, 1e-3, 1e-2}.
-
hard_window
- It indicates the difficulty of sampling in the hard-negative sampler.
- We set it to the top thirty percent.
-
hard_prob
- It indicates the probability of using hard-negative samples in the further training stage.
- We set it to 0.8 (0.4 in the item level and 0.4 in the bundle level), so the probability of simple samples is 0.2.
We provide one processed dataset: Netease.
-
user_bundle_train.txt
- Train file.
- Each line is 'userID\t bundleID\n'.
- Every observed interaction means user u once interacted bundle b.
-
user_item.txt
- Train file.
- Each line is 'userID\t itemID\n'.
- Every observed interaction means user u once interacted item i.
-
bundle_item.txt
- Train file.
- Each line is 'bundleID\t itemID\n'.
- Every entry means bundle b contains item i.
-
Netease_data_size.txt
- Assist file.
- The only line is 'userNum\t bundleNum\t itemNum\n'.
- The triplet denotes the number of users, bundles and items, respectively.
-
user_bundle_tune.txt
- Tune file.
- Each line is 'userID\t bundleID\n'.
- Every observed interaction means user u once interacted bundle b.
-
user_bundle_test.txt
- Test file.
- Each line is 'userID\t bundleID\n'.
- Every observed interaction means user u once interacted bundle b.