vwz / SPE

Subgraph-augmented Path Embedding for Semantic User Search on Heterogeneous Social Network

Reference:
Zemin Liu, Vincent W. Zheng, Zhou Zhao, Hongxia Yang, Kevin Chen-Chuan Chang, Minghui Wu, and Jing Ying. 2018. Subgraph-augmented Path Embedding for Semantic User Search on Heterogeneous Social Network. In Proceedings of The 2018 Web Conference (WWW 2018). ACM. New York, NY. USA, 10 pages. 

This python project implements the SPE model proposed in the above paper. 
Please refer the above paper for all the notations in this readme file. 
If you use it for scientific experiments, please cite this paper:
@inproceedings{LiuZ18,
 author = {Zemin Liu and 
	   Vincent W. Zheng and 
	   Zhou Zhao and 
	   Hongxia Yang and 
	   Kevin Chen-Chuan Chang and
	   Minghui Wu and 
	   Jing Ying},
 title = {Subgraph-augmented Path Embedding for Semantic User Search on Heterogeneous Social Network},
 booktitle = {The 2018 Web Conference (WWW 2018)},
 year = {2018}
} 

****************************

FileList:

	source-code		:	The source code of SPE model
	readMe			:	This file.


=======================================================================
source-code
=======================================================================
FileList in source-code folder:
	dataset preparation-SPE	:	These codes are used to prepare datasets offline for SPE.
	model-autoencoder		:	These codes are used to learn the subgraph embedding from a subgraph structural similarity matrix.
	model-SPE				:	These codes are used to learn the SPE model and evaluate it.
	
1. dataset preparation-SPE
	These codes are written in Java, which are used to prepare datasets offline for SPE.
	First of all, we use codes from [1] to get all the subgraphs in a graph and then get the instances of each subgraph.
	Then we use dataset preparation-SPE to sample o-paths from graph, and then we calculate the number of instances between any node u and node v, and finally generate the m-paths.
	All of these are done offline and then saved in the database(or files), so we could use directly in the period of training and testing.
	
	Main.java is the entry class. And this program reads the parameters from Config.java.
	
2. model-autoencoder
	These codes are written in Python, which are used to learn the subgraph embedding from a subgraph structural similarity matrix.
	Here we use aotuencoder to learn an embedding x for each subgraph m.
	
	Here autoencoderTraining.py is the main method of this autoencoder model. And we set parameters in this file.
	
3. model-SPE
	These codes are written in Python, which are used to train the SPE model and to evaluate it. It takes the output of codes 1 and codes 2 as its input.
	In this folder, experimentForOneFileByParams.py is the main method. This file provides us a method to get parameters from file pythonParamsConfig, and then train the model and test it.
	The pythonParamsConfig is the configuration file of this model, and we could set parameters in this file.
	At last, this model then will output the results of NDCG and MAP.
	
=======================================================================
References
=======================================================================
[1]. Fang, Yuan, et al. "Semantic proximity search on graphs with metagraph-based learning." Data Engineering (ICDE), 2016 IEEE 32nd International Conference on. IEEE, 2016.