logan-lin / elaine-tensorflow

Tensorflow implementation of ELAINE, test on DiDi public data.

TensorFlow Implementation of ELAINE

The basic idea of ELAINE is to capture multi label features by feeding combination of two nodes' latent vector into a decoder and try to reconstruct label features linking the two nodes. The author also used variational autoencoder instead of normal FC autoencoder.

Data used to construct graph is from DiDi. Get it here

Paper: Goyal P, Hosseinmardi H, Ferrara E, et al. Capturing Edge Attributes via Network Embedding[J]. arXiv preprint arXiv:1805.03280, 2018. PDF

Core network structure:

Because I am not really sure about the exact training method used by the author of the paper, this implementation propably can't represent the real ELAINE performance.