This is a reference implementation for Spatio-Temporal Attentive RNN for Node Classification in Temporal Attributed Graphs (IJCAI'19).
Please feel free to contact Dongkuan Xu (dux19@psu.edu) if you have any question.
-
The datasets we use are included in the folder 'Data'
-
The program that can run on the DBLP3 dataset is included in the folder 'Code'
-
When running the main_DBLP3.py, you should 3.1 change the data path (Lines 366 & 367) 3.2 change 'num_classes' (Line 383) according to the dataset 3.3 change the saving path (Line 541)
-
In Line 541, the results include 4.1 Alpha: the temporal attention values of the test set 4.2 Beta: the patial attention values of neighbors of the test set 4.3 X_test_idx: the idx of the test set 4.4 y_test: the label of the test set 4.5 samples_idx: the idx of neighbors of the test set (k-hop neighbors)
-
In order to get a good results, you should well tune 5.1 lr: learning rate 5.2 training_iters: training epochs 5.3 lambda_l2_reg: penalty for parameters 5.4 lambda_reg_att: penalty for multiple temporal attention units
Please consider citing the following paper if you find STAR useful for your research:
@inproceedings{xu2019sp,
title={Spatio-Temporal Attentive RNN for Node Classification in Temporal Attributed Graphs},
author={Xu, Dongkuan and Cheng, Wei and Luo, Dongsheng and Liu, Xiao and Zhang, Xiang},
booktitle={Proceedings of the 29th International Joint Conference on Artificial Intelligence},
year={2019} }