wehlutyk
commented
6 years ago Currently training runs, using the projects/scale/blogcatalog.py script, but the resulting 2D embeddings are awful: they all align on a single line which is more or less y = - x. Accordingly, adjacency reconstruction is awful too. This can be seen in the current projects/scale/blogcatalog-results.ipynb notebook.
Ideas:
- the embedding dimension is too low -> train with a higher dimension, downscale the embeddings to see what they look like, and see if adjacency reconstruction is any better. -> #31
- run training animations with increasingly more nodes to see what goes on during the training -> #30
Work-in-Progress for applying the method to the BlogCatalog data set.