Code for Multisample-based Contrastive Loss for Top-k Recommendation (IEEE TMM): H. Tang, G. Zhao, Y. Wu and X. Qian, "Multisample-based Contrastive Loss for Top-k Recommendation," in IEEE Transactions on Multimedia, doi: 10.1109/TMM.2021.3126146.
What we propose is a loss function that can be applied to many methods. Our code is based on the LightGCN. Thanks for the previous work and code https://github.com/gusye1234/LightGCN-PyTorch . We followed its setup, which you can refer to.
Use Cpp Extension in code/sources/ for negative sampling. To use the extension, please install pybind11 and cppimport under your environment.
The parameter adjustment is relatively simple, and the following is available for reference.
python main.py --layer 3 --dataset="amazon-book" --temperature 0.1 --info sgk15_alpha0.45
python main.py --layer 1 --dataset="ifashion" --decay 1e-5 --temperature 0.2 --info sgk15_alpha0.60
python main.py --layer 2 --dataset="yelp2018" --temperature 0.2 --info sgk15_alpha0.45