wenjiewwj / clickbait

Clicks can be Cheating: Counterfactual Recommendation for Mitigating Clickbait Issue

Clicks can be Cheating: Counterfactual Recommendation for Mitigating Clickbait Issue

This is the pytorch implementation of our paper at SIGIR 2021:

Clicks can be Cheating: Counterfactual Recommendation for Mitigating Clickbait Issue

Wenjie Wang, Fuli Feng, Xiangnan He, Hanwang Zhang, Tat-Seng Chua.

Environment

• Anaconda 3
• python 3.7.3
• pytorch 1.4.0
• numpy 1.16.4

Usage

Parameters

• model_name: MMGCN.
• l_r: learning rate. Default: 1e-3.
• weight_decay: the hyper-parameter for weight decay. Default: 1e-3.
• gpu_id: the gpu used for training.

Other parameter settings can be found in train.py. We keep the default setings as MMGCN.

Training

python train.py --model_name=$1 --l_r=$2 --weight_decay=$3

or use run.sh

sh run.sh gpu_id model_name l_r weight_decay

The log file will be in the ./log/ folder.

Inference

1. Download the checkpoints released by us from Google drive.
2. Put the '.pth' file into the model_1 folder.
3. Run inference.py or run_inference.sh:

python inference.py --model_name=$2 --l_r=$3 --weight_decay=$4 --log_name="$2_tiktok_$3lr_$4wd_$5"

sh run_inference.sh gpu_id model_name l_r weight_decay log_name

Examples

1. Train MMGCN on Tiktok:

cd ./code/tiktok
CUDA_VISIBLE_DEVICES=0 python main.py --model_name=MMGCN --l_r=1e-3 --weight_decay=1e-3

2. Inference MMGCN on Adressa

cd ./code/adressa
sh run_inference.sh 0 MMGCN 1e-3 1e-3 TIE

Citation

If you use our code, please kindly cite:

@inproceedings{wang2021Clicks,
  title={Clicks can be Cheating: Counterfactual Recommendationfor Mitigating Clickbait Issue},
  author={Wenjie Wang, Fuli Feng, Xiangnan He, Hanwang Zhang, and Tat-Seng Chua},
  booktitle={SIGIR},
  year={2021},
  publisher={ACM}
}

Acknowledgment

Thanks to the MMGCN implementation:

• MMGCN from Yinwei Wei.

License

NUS © NExT++

clickbait's Issues

About torch-sparse version problem

Does torch-geometric specify a version？torch-sparse Installation failed.Ihave tried many times.

a issue about dataset

hi,wenjie
I've been trying to run your code recently,
but I found that the tiktok dataset byte recommendation can't be downloaded from the official website.
So can you give me a dataset or Google drive disk access (my email address is [email protected] ）
Thank you very muchhhh.

.

Could you please provide the hyper-parameters to help the reader get the results published in the paper?

'Net' object has no attribute 'model'

I came across an error message when I run it with the command "python train.py --model_name=$1 --l_r=$2 --weight_decay=$3": AttributeError: 'Net' object has no attribute 'model'.I checked carefully and didn't know the reason for the mistake.Have you ever made a similar mistake？We hope to receive your reply, which will be greatly appreciated.

edge_index is empty

FYI: if you meet the problem above while running train.py, consider checking the version of TorchGeometric

If version <=1.4.3, try change file in TorchGeometric lib: nn\conv\message_passing.py,
Add these lines at line 170:
out['dim_size'] = out['size_i']
out['edge_index'] = edge_index
out['size'] = size
so the edge_index won't be viewed as "empty".

By the way, updating version may help with the bug, but you have to change "Scatter_" to torch version for there's no “Scatter_” in new version of TorchGeometric

Good Luck

数据集

你好，我看到你的代码Clickbait中提供的tiktok数据集中有三个特征文件，feat_a.pt，feat_v.pt，feat_t.pt。其中feat_a.pt和feat_v.pt的维度为[34756, 128]，但是feat_t.pt的维度为[2, 261230]。这个feat_t.pt文件是否有问题呢？如果有问题的话，您能提供一下正确的feat_t.pt文件吗？非常感谢！