NCFML (Neural Collaborative Filtering with MovieLens in torch)

Dataset

This repository is about Neural Collaborative Filtering with MovieLens in torch. Dataset is Implict Feedback, If there is interaction between user and item, then target value will be 1.So if there is rating value between user and movie, then target value is 1, otherwise 0. For negative sampling, ratio between positive feedback and negative feedback is 1:4 in trainset, and 1:99 in testset. (these ratios are same as author's code @hexiangnan)

You can use 100k, 1m, 10m, 20m dataset by using parser parameter --file_size.

Dependency

pytorch >= 1.12.0
python >= 3.8
scipy >= 1.7.1
numpy >= 1.20.3

Quick start

python main.py --epoch 30 --batch 256 --factor 8 --model NeuMF --topk 10 --file_size 100k --layer 64 32 16 --download True --use_pretrain False

Development enviroment

• OS: Max OS X
• IDE: pycharm
• GPU: NVIDIA RTX A6000

Result

movielens 100K	Best HR@10	NDCG@10	Runtime	epoch	preditivie factor	batch_size	layer for MLP
GMF	0.815	0.552	2m 40sec	20	8	256	X
MLP	0.803	0.567	17m 42sec	20	8	256	[64,32,16]
NeuMF (without pre-training)	0.828	0.574	21m 19sec	20	8	256	[64,32,16]
NeuMF (pretrained)	0.980	0.702	4m 42sec	20	8	256	X

movielens-1M	Best HR@10	NDCG@10	Runtime	epoch	preditivie factor	batch size	layer for MLP
GMF	0.761	0.493	26m 45sec	20	8	256	X
MLP	0.785	0.526	1h 47m 24sec	20	8	256	[64,32,16]
NeuMF (without pre-training)	0.796	0.538	1h 46m 20sec	20	8	256	[64,32,16]
NeuMF (pretrained)	0.851	0.854	41m 22sec	20	8	256	X

Example of command line

• save GMF

  python main.py --epoch 30 --batch 256 --factor 8 --model GMF --topk 10
  --file_size 100k --layer 64 32 16 --download True --save True

• save MLP

  python main.py --epoch 30 --batch 256 --factor 8 --model MLP --topk 10
  --file_size 100k --layer 64 32 16 --download False --save True

• use pre-trained model

  python main.py --epoch 30 --batch 256 --factor 8 --model NeuMF  --topk 10
  --file_size 100k --layer 64 32 16 --download False --use_pretrain True

Reference

1. Neural Collaborative Filtering
2. Official code from author