What's the purpose of Injecting noise to the hidden layer? - Githubissues

TrustAGI-Lab / ARGA

This is a TensorFlow implementation of the Adversarially Regularized Graph Autoencoder(ARGA) model as described in our paper: Pan, S., Hu, R., Long, G., Jiang, J., Yao, L., & Zhang, C. (2018). Adversarially Regularized Graph Autoencoder for Graph Embedding, [https://www.ijcai.org/proceedings/2018/0362.pdf].

MIT License

185 stars 58 forks source link

What's the purpose of Injecting noise to the hidden layer? #7

Open cshjin opened 5 years ago

cshjin commented 5 years ago

Hi @Ruiqi-Hu, I see here https://github.com/Ruiqi-Hu/ARGA/blob/a970fa583d8c474b18f950da06bf91da03a647db/ARGA/arga/model.py#L67 you injected a Gaussian noise to the first layer.

What's the purpose of doing so?

Based on the original implementation of GAE, https://github.com/tkipf/gae/blob/a1aecb00de894859f509f2d392135086a86946a0/gae/model.py#L55-L69 there is no noise, and

have you compared the results with and without the noise of ARGA with GAE?
Actually adding noise will affect the Gaussian prior p(z), right?