personalized recommendation

several algorithms for personalized recommendation

Getting Started

These instructions will get you a copy of the project up and running on your local machine for development and testing purpose.

Prerequisites

You need the amazon-book/yelp2018/last-fm dataset for training and testing. To get one copy of that, take a look at dataset. After you get the dataset, unzip the knowledge graph file.

unzip -x kg_final.txt.zip

Please bear in mind that this dataset directory should be put under /Data. Then go to the /Data to prepare the test set, test sample and ground truth will be saved in pickle file.

python dump_v1.py

See configuration in config/config.py under bprmf, you can choose which dataset you want to dump. Once finishing these processes, you are all set for data part.

Installing

Just use the requirement file to prepare the environment, here we use python3.6.

pip install -r requirement.txt

Train model

Currently we support three models, choose one and go to the corresponding directory.

python train.py

Just use this command to train a model, you can specify argument in command line, various parameters can be found at config.py. I test model for every ten epochs. Both the best and the final model will be saved in /output.

Evaluate model

Typically, we saved our model under /output, and if you would like to test your own model, you should put your .ckpt file under that directory or specify the file path in command line.

python test.py --model_path output/best_model.ckpt

Use this command to test your model, all metric will be printed on screen.

Result

dataset: yelp2018

Model	Recall@20	Hit ratio@20	Precision@20	NDCG@20
BPR-MF	0.0488	0.1994	0.0118	0.0607
Random walk	0.0522	0.2134	0.0127	0.0643
DNS	0.0686	0.2601	0.0169	0.0839
IRGAN	0.0556	0.2203	0.0136	0.0814

dataset: amazon-book

Model	Recall@20	Hit ratio@20	Precision@20	NDCG@20
BPR-MF	0.1214	0.2195	0.0128	0.0805
Random walk	0.1172	0.2153	0.0124	0.0772
DNS	0.1530	0.2638	0.0161	0.1047
IRGAN	0.1296	0.2311	0.0137	0.0920

dataset: last-fm

Model	Recall@20	Hit ratio@20	Precision@20	NDCG@20
BPR-MF	0.0660	0.3065	0.0270	0.1056
Random walk	0.0670	0.3053	0.0275	0.1082
DNS	0.0876	0.3694	0.0361	0.1385
IRGAN	0.0615	0.3103	0.0268	0.1238

Note

In this repo, we actually provide five types of sampling strategies for negative item sampling, you can using this by specifying parameters in dataset preparation

Random Sampling
Random walk in knowledge graph
Popularity based sampling
Dynamic sampling strategy
Adversarial sampling strategy

Reference

[1] Rendle, Steffen, et al. "BPR: Bayesian personalized ranking from implicit feedback." Proceedings of the twenty-fifth conference on uncertainty in artificial intelligence. AUAI Press, 2009.

[2] Zhang, Weinan et al. “Optimizing top-n collaborative filtering via dynamic negative item sampling.” SIGIR (2013).

[3] Wang, Jun et al. “IRGAN.” Proceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval - SIGIR ’17 (2017): n. pag. Crossref. Web.

Funding Source Acknowledgement

This research is supported by the National Research Foundation, Singapore under its International Research Centres in Singapore Funding Initiative. Any opinions, findings and conclusions or recommendations expressed in this material are those of the author(s) and do not reflect the views of National Research Foundation, Singapore.

xiangwang1223/AdvRec