Why are community parameters not preserved over multiple iterations of ComE?

[abegehr]

On p.381 in algorithm 1 line 3 of the ComE paper, it states that multiple iterations are run: updating, in turn, community detection/embedding with EM and node embedding with SGD.
In the code repository published on GitHub, the iterating for-loop can be found in main.py L116: https://github.com/andompesta/ComE/blob/master/main.py#L116
The for-loop contains two more for-loops that iterate over the hyper-parameters K, alpha, and beta. Since the model is re-loaded again from pre-training parameters on each for-loop iteration (https://github.com/andompesta/ComE/blob/master/main.py#L121), I do not see the connection between model parameters of the same hyper-parameters between multiple iterations (1..T1).
A new Gaussian Mixture Model is reinitialized on each fit of the com_learner (Community2Vec): https://github.com/andompesta/ComE/blob/master/ADSCModel/community_embeddings.py#L24

@andompesta, I am wondering why community parameters (means and covariances) are not preserved over multiple iterations of ComE?