PyTorch Implementation for Adaptive Hardness Negative Sampling for Collaborative Filtering, AAAI2024
--- Based on "MixGCF: An Improved Training Method for Graph Neural Network-based Recommender Systems", https://github.com/huangtinglin/MixGCF
The code has been tested running under Python 3.7.6. The required packages are as follows:
- pytorch == 1.7.0
- numpy == 1.20.2
- scipy == 1.6.3
- sklearn == 0.24.1
- prettytable == 2.1.0
The training commands are as following:
python main.py --dataset ml --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 0 --ns ahns --alpha 0.1 --beta 0.4 --n_negs 16
python main.py --dataset phone --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 0 --ns ahns --alpha 1.0 --beta 0.1 --n_negs 32
python main.py --dataset sport --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 0 --ns ahns --alpha 0.5 --beta 0.1 --n_negs 32
python main.py --dataset tool --gnn lightgcn --dim 64 --lr 0.001 --batch_size 2048 --gpu_id 0 --context_hops 0 --ns ahns --alpha 1.0 --beta 0.1 --n_negs 32
The ipynb files are also provided.
We use four processed datasets: MovieLens-1M, Amazon-Phones, Amazon-Sports and Amazon-Tools. The processing code is also provided.
| #user | #item | #inter. | avg. inter. | |
|---|---|---|---|---|
| ML-1M | 6.0k | 3.7k | 1000.2k | 165.6 |
| Phones | 27.9k | 10.4k | 194.4k | 7.0 |
| Sports | 35.6k | 18.4k | 296.3k | 8.3 |
| Tools | 16.6k | 10.2k | 134.5k | 8.1 |