Add NCF_PyTorch models

[hieuddo]

Description

Add the option to use NCF models in PyTorch (implemented models are tensorflow).

Checklist:

• I have added tests.
• I have updated the documentation accordingly.

[hieuddo]

I did a quick exp with movielens-100k dataset, n_factors=8 and layers=[32,16,8]

• MSELoss for GMF, 100 epochs
• BCELoss for MLP and NeuMF, each 10 epochs

and get these results:

	NDCG@50	Recall@50	Train (s)	Test (s)
GMF-PyTorch	0.0344	0.0556	137.7308	0.5606
MLP-PyTorch	0.2127	0.3179	14.0588	0.7142
NeuMF-PyTorch	0.2123	0.3169	14.3680	0.7654

The results for the same dataset from implemented Tensorflow version, same embedding/structure size (which are acquired from cornac/tutorials notebook ) are:

	NDCG@50	Recall@50	Train (s)	Test (s)
GMF	0.1609	0.2778	33.4836	1.2062
MLP	0.1630	0.2807	24.3624	1.2698
NeuMF	0.2412	0.4268	24.6304	1.3844

Those results are with some random hyperparameters without carefully tuning, so I think PyTorch version is comparable with Tensorflow version. @darrylong please help review the code! Thanks.

[tqtg]

@hieuddo any idea why GMF performance is quite low as compared to others?

[hieuddo]

Turned out to be because of seed. I carried out the exp for GMF with the same hyperparameters, except seed, and got these results:

	NDCG@50	Recall@50	Train (s)	Test (s)
GMF-PyTorch	0.2059	0.3083	166.3036	0.5522
GMF-PyTorch	0.1979	0.2867	163.1750	0.5426
GMF-PyTorch	0.0005	0.0009	162.1954	0.5394

This could be due to overfitting because training_loss is very low (0.04).

Anyway, my point still stands: comparable exp results are achievable with this PyTorch implementation.

One discussion though, should we stick to the original paper and hard-coded BCE as loss function or should we offer the option to change loss function by hyperparameter?

[tqtg]

@hieuddo I refactor the code so we can specify backend for the models. They're all runnable though could you help check the logic if I've missed something? Refer to the updated ncf_example.py.

[tqtg]

One discussion though, should we stick to the original paper and hard-coded BCE as loss function or should we offer the option to change loss function by hyperparameter?

BCE is fine. Any reason why we want to add more options?

[hieuddo]

@hieuddo I refactor the code so we can specify backend for the models. They're all runnable though could you help check the logic if I've missed something? Refer to the updated ncf_example.py.

For now, it seems fine to me. Will check again later because I'll be using NCF-PyTorch in the future.

BCE is fine. Any reason why we want to add more options?

No specific reason. Just my random wonder. So let's just stick with BCE.

Btw, your "refactor" commit is wonderful madness. I guess I'll be contributing more to cornac so I can learn from your bag of tricks haha

[tqtg]

Here are results running the example with the two backend:

TensorFlow	NDCG@50	Recall@50	Train (s)	Test (s)
GMF	0.0404	0.1148	69.4215	1.9982
MLP	0.0397	0.1212	68.0473	3.3358
NeuMF	0.0397	0.1207	68.6332	3.5113
NeuMF_pretrained	0.0398	0.1212	67.5522	3.4957

PyTorch	NDCG@50	Recall@50	Train (s)	Test (s)
GMF	0.0407	0.1167	70.2337	1.7052
MLP	0.0391	0.1105	67.4546	2.071
NeuMF	0.0397	0.1155	68.184	2.3072
NeuMF_pretrained	0.0407	0.1167	67.5983	2.2617

They look comparable both in terms of performance and time. Still missing model load and save logic for PyTorch though we can add them later in another PR. This one has become quite complex with many changes.

@hieuddo @darrylong feel free to merge if you're comfortable with this. Thanks Hieu again for the contribution.