Skip to content

Latest commit

 

History

History
55 lines (44 loc) · 2.24 KB

README.md

File metadata and controls

55 lines (44 loc) · 2.24 KB

MetaHIN

Source code for KDD 2020 paper "Meta-learning on Heterogeneous Information Networks for Cold-start Recommendation"

Requirements

  • Python 3.6.9
  • PyTorch 1.4.0
  • My operating system is Ubuntu 16.04.1 with one GPU (GeForce RTX) and CPU (Intel Xeon W-2133)
  • Detailed requirements

Datasets

We have uploaded the original data of DBook, Movielens and Yelp in the data/ folder.

The processed data of DBook and Movielens can be downloaded from Google Drive and BaiduYun (Extraction code: ened).

The processed data of Yelp can be generate by the code data/yelp/YelpProcessor.ipynb.

Description

MetaHIN/
├── code
│   ├── main.py:the main funtion of model
│   ├── Config.py:configs for model
│   ├── Evaluation.py: evaluate the performance of learned embeddings w.r.t clustering and classification
│   ├── DataHelper.py: load data
│   ├── EmbeddingInitializer.py: map feature and inilitize embedding tables
│   ├── HeteML_new.py: update paramerters in meta-learning paradigm 
│   ├── MetaLeaner_new.py: the base model 
├── data
│   └── dbook
│       ├── original/: the original data without any preprocess
│       ├── DBookProcessor.ipynb: preprocess data 
│   └── movielens
│       ├── original/: the original data without any preprocess
│       ├── MovielensProcessor.ipynb: preprocess data 
│   └── yelp
│       ├── original/: the original data without any preprocess
│       ├── YelpProcessor.ipynb: preprocess data 
├── README.md

Reference

@inproceedings{lu2020meta,
  title={Meta-learning on Heterogeneous Information Networks for Cold-start Recommendation},
  author={Lu, Yuanfu and Fang, Yuan and Shi, Chuan},
  booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  pages={1563--1573},
  year={2020}
}