NCF: Neural Collaborative Filtering(WWW, 2017)

Dataset

I used MovieLens dataset(file size: 100K). 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)

Neural Collaborative Filtering model directory tree

.
├── README.md
├── main.py
├── data_utils.py
├── GMF.py
├── MLP.py
├── NeuMF.py
├── evaluate.py
└── dataset
    ├── ml-latest-small
    │   └── ratings.csv
    └── ml-latest-small.zip

Neural Collaborative Filtering Result

MovieLens 100K	HR	NDCG	Runtime	epoch	learning rate	batchsize	predictive factor	the number of layer
GMF	0.825	0.554	2m	20	0.001	256	8	X
MLP	0.825	0.559	2.8m	20	0.001	256	8	3
NeuMF(without pre-training)	0.848	0.578	3m	20	0.001	256	8	3
NeuMF(with pre-training)	0.875	0.577	2.6m	20	0.001	256	8	X

Development Enviroment

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

Quick Start Example

python main.py -m GMF -nf 8 -b 256 -e 20 -lr 0.001 -tk 10 -pr False -save True

python main.py -m MLP -nf 8 -nl 3 -b 256 -e 20 -lr 0.001 -tk 10 -pr False -save True

python main.py -m NeuMF -nf 8 -nl 3 -b 256 -e 20 -lr 0.001 -tk 10 -pr False

python main.py -m NeuMF -nf 8 -nl 3 -b 256 -e 20 -lr 0.001 -tk 10 -pr True

Reference

paper : Neural Collaborative Filtering

review written in korean : Review

Neural Collaborative Filtering with MovieLens in torch

In progress 'with pre-training'