This a Tensorflow implementation of the DSNR algorithm, which learns a low-dimensional representations for each node in a network. DSNR explores the implicit semantic information of nodes and captures the high-order semantic proximity. Specifically, to make the model effective, DSNR consist of three module, i.e., semantic aware module which models the implicit semantic information of nodes and gets the enhanced semantic feature representation by a deep autoencoder, embedding module which fuses the enhanced semantic feature in node identity sequence generated by the truncated random walk to generate node semantic sequence, and high-order structure aware module which translate the network from the node semantic sequence to the node identity sequence by the attention-enhanced seq2seq framework.
An example of node with implicit semantic information. All Bob’close friends are explorer. Thus, Bob should also like expedition.The same goes for Tom. Furthermore, Bob and Tom have the similar implicit semantic information, but they are far apart in the network. However, their node representations in the latent space should be similar.
The framework of DSNR consists of three major modules. For the semantic aware module, we propose a heuristic method to get the implicit semantic of nodes by integrating node neighbors’ semantic information. We then use a deep autoencoder to encode and decode the inherent semantic information of nodes. Finally, we reconstruct the implicit semantic feature of nodes instead of its inherent semantic feature. For the embedding module, we construct the node semantic sequence by fusing the enhanced semantic feature vector of nodes and the node identity sequence generated by a truncated random walk. For the high-order structure aware module, we translate network from the node semantic sequence to the node identity sequence by an attention-enhanced seq2seq framework.
python 3.6.4
Tensorflow 1.0
Numpy
networkx
Sklearn
The develop tool is PyCharm Community Edition 2019.3.1
Cora, Citeseer and Wiki datasets are given.
1.cora.edgelist
2.cora.features (node original semantic feature vector representations: attributes of nodes have been processed as TFIDF vector)
3.cora.svd (node original semantic feature vector representations: TFIDF matrix on node attributes is decomposed by SVD, and then the left singular matrix obtained is taken as the input feature)
4.cora.sequence (it has been processed by truncated random walk)
In this paper, we select node TFIDF feature vector as the original semantic feature of nodes
For node classification, each dataset contain 4 files: edgelist,feature, labels and node sequence
node_1 node_2
node_2 node_2
...
the n lines are as follows:(each node per line ordered by node id)
(for node_1) feature_1 feature_2...feature_n
(for node_1) feature_1 feature_2...feature_n
...
node_1 label_1
node_1 label_1
...
node_1 node_2 node_2...node_10
python main.py
