This is the implementation of our paper "HINTS: Citation Time Series Prediction for New Publications via Dynamic Heterogeneous Information Network Embedding", published in WWW'21.
The original data used could be access from Aminer and APS.
We also provide our processed data Dropbox or Google Drive to reproduce the results reported in our paper.
All required packages could be found in requirements.txt (generated by pip freeze).
NOTE: our implementation is on Tensorflow 1.14, may not be very friendly if you are familar with dynamic computational graph based frameworks.
-
Step0 (data):
- Download the processed data and
unzip *.zipunder the root folder.
- Download the processed data and
-
Step1 (run):
cd ./src- For Aminer dataset:
python main.py --dataset aminer --epochs 700 --batch_size 3000 - For APS dataset:
python main.py --dataset aps --epochs 500 --batch_size 1200
Arguments interpretation:
-
--dataset: processed dataset, either AMiner or APS. -
--epochs: number of training on the training set. -
--batch_size: batch size of one training.
Note that the batch size in set as the number of papers used in training, i.e., there is only one interation per epoch. I didn't try "mini-batch"-like training.
The prediction files will be stored under the result folder.
Song Jiang songjiang@cs.ucla.edu
@inproceedings{hints,
title={HINTS: Citation Time Series Prediction for New Publications via Dynamic Heterogeneous Information Network Embedding},
author={Song Jiang, Bernard J. Koch, Yizhou Sun},
booktitle={Proceedings of The Web Conference},
year={2021}
}