This repository is pytorch reimplementation of author's code which is implemented using keras and theano.
You can check the paper through this link.
I used author's preprocessed dataset.
I only checked the performance of NeuMF model on ml-1m dataset without pre-training. The hyperparmeters were set as the following.
| hyperparameter | value | |
|---|---|---|
| number of epochs | 9 | --epochs |
| batch size | 256 | --batch_size |
| dimmension of gmf vector | 32 | --num_factors |
| dimmensions of mlp hidden layers | [64,64,32,16] | --layers |
| number of negative instances to pair with a positive instance | 4 | --num_neg |
| fixed learning rate | 5e-05 | --lr |
The best performance was as below. These are almost same with what the paper showed.
| Movielens HR | Movielens NDCG |
|---|---|
| 0.6997 | 0.4245 |
- numpy (1.19.0)
- scipy (1.5.1)
- torch (1.5.1)
- implement MLP and GMF models respectively
- add more experiments in various settings. (different datasets, different hyperparmeters)