This repository is the official implementation of Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering (Accepted by NeurIPS 2020)

Jingtao Ding, Yuhan Quan, Quanming Yao, Yong Li,Depeng Jin. Simplify and Robustify Negative Sampling for Implicit Collaborative Filtering.

We conducted experiments under：

• python 3.7
• tensorflow 1.14

To tune the hyper-parameter on synthetic dataset（Ecom-toy） with uniform negative sampling method:

$ cd ./SRNS_syn
$ ./run_uniform.sh

To run the synthetic experiments on synthetic dataset:

$ cd ./SRNS_syn
$ ./run_srns.sh

For SRNS experiment, we constantly feed a false negative into each user’s memory $M$, you can tune the hyper-parameter $\sigma \in [1 ,3, 5, 7, 9]$ to vary the size of available false negative instance. The data of different size of available false negative instance are named fn_sigma_*.pkl, which means $\sigma \in [0.1, 0.3, 0.5, 0.7, 1.0]$ actually. They are under the ./SRNS_syn/toy and ./SRNS_syn/toy_tuning folder.

Notice: hyper-parameter tuning and formal experiments use the different default dataset, they are all generated from the Ecom-toy dataset. You can also run the synthetic experiments on the hyper-parameter tuning dataset.

To run SRNS on Ml-1m

$ cd ./SRNS_real
$ ./run.sh

You can get pre-trained models in ./SRNS_real/model, it includes SRNS and all other baselines in our paper.

To evaluate the pre-trained models

$ cd ./SRNS_real
$ ./run_predict_model.sh

The performance of the pretraind model on ML-1M: