This repository contains code for paper "On Sampling Strategies for Neural Network-based Collaborative Filtering", which propose (1) a general NNCF framework incorporates both interaction and content information, and (2) sampling strategies for speed up the process.

Loss Functions

• pointwise: loss=skip-gram, mse
• pairwise: loss=log-loss, max-margin

Content Embedding

• CNN: model_choice=cnn_embedding
• RNN: model_choice=rnn_embedding
• Mean of word vectors: model_choice=basic_embedding

Interaction Module

• dot product

Sampling strategies

• IID sampling: train_scheme=presample + shuffle_st=random

• Negative sampling: train_scheme=original

• Stratified sampling: train_scheme=group_sample + shuffle_st=item

• Negative sharing: train_scheme=neg_shared

• Stratified sampling with negative sharing: train_scheme=group_neg_shared

Other parameters explained

• eval_scheme: whole@k for using all test items as candidates, given@-1 for using true items + random sampled 10 items as candidates.

• neg_loss_weight & gamma: adjustment in functions

• chop_size: number of positive links per item in a batch. Only approximately when set group_shuffling_trick=True (which is recommended).

1. unzip data in ./data folder, and go to ./code/sampler, execute ./make.sh
2. run using scripts under ./code/scripts/demos, which are prepared for each of the sampling strategies.
3. after running, the results are stored in ./results folder

0. Unix system with python 2.7, GCC 4.8.x and GSL
1. Keras 1.2.2
2. Tensorflow 1.0

The code may or may not be working properly with other versions.

Tips

• GSL can be installed with sudo apt-get install libgsl-dev
• Keras can be installed by first downloading zip file of version 1.2.2 and then installing with command python setup.py install

@inproceedings{chen2017onsampling,
	title={On Sampling Strategies for Neural Network-based Collaborative Filtering},
	author={Chen, Ting and Sun, Yizhou and Shi, Yue and Hong, Liangjie},
	booktitle={Proceedings of the 23th ACM SIGKDD international conference on Knowledge discovery and data mining},
	year={2017},
	organization={ACM}
}