APR enhances the pairwise ranking method BPR by performing adversarial training. To illustrate how it works, APR on MF is implemented here by adding adversarial perturbations on the embedding vectors of users and items.

This is our official implementation for the paper:

Xiangnan He, Zhankui He, Xiaoyu Du & Tat-Seng Chua. 2018. Adversarial Personalized Ranking for Recommendation , In Proceedings of SIGIR'18.

(Corresponding Author: Dr. Xiangnan He)

If you use the codes, please cite our paper . Thanks!

Python 2.7

TensorFlow >= r1.0

Numpy >= 1.12

PS. For your reference, our server environment is Intel Xeon CPU E5-2630 @ 2.20 GHz and 64 GiB memory. We recommend your free memory is more than 16 GiB to reproduce our experiments.

This command shows the effect of APR by adding adversarial perturbation on pretrained MF model (--restore) for dataset yelp in epoch 40 (--adv_epoch). After loading the pretrained model, the first 40 epochs are normal MF-BPR, followed by adversarial training APR.

python AMF.py --dataset yelp --adv_epoch 40 --epochs 1000 --eps 0.5 --reg_adv 1 --verbose 20 --restore 2018_05_06_16_09_24

or use ./demo.sh for short.

To launch the entire training experiment quickly, you can use:

python AMF.py --dataset yelp --adv_epoch 1000 --epochs 2000 --eps 0.5 --reg_adv 1 --ckpt 1 --verbose 20

or ./train.sh yelp for short. This command can generate the above figure (same as the paper). Specifically, it trains 2000 epochs (--epochs) in total, where the first 1000 epochs train MF-BPR (--adv_epoch), and then followed by 1000 epochs of AMF training.

Use python AMF.py -h to get more argument setting details.

-h, --help            show this help message and exit
--path [PATH]         Input data path.
--dataset [DATASET]   Choose a dataset.
--verbose VERBOSE     Evaluate per X epochs.
--epochs EPOCHS       Number of epochs.
--adv_epochs          The epoch # that starts adversarial training (before that are normal BPR training). 
......

We provide three processed datasets: Yelp(yelp), MovieLens 1 Million (ml-1m) and Pinterest (pinterest-20) in Data/

train.rating:

• Train file.

• Each Line is a training instance: userID\t itemID\t rating\t timestamp (if have)

test.rating:

• Test file (positive instances).
• Each Line is a testing instance: userID\t itemID\t rating\t timestamp (if have)

test.negative

• Test file (negative instances).
• Each line corresponds to the line of test.rating, containing 99 negative samples.
• Each line is in the format: (userID,itemID)\t negativeItemID1\t negativeItemID2 ...

PS. In our experiments, we adopt the all ranking evaluation strategy. But we still provide the dataloader function for negative sampling evaluation in Dataset.py for people who are interested in it.

Update: May 7, 2018