[ ] Convert data to latent space and train LSTM for prediction
[ ] Conditional Seq2seq VAE + conditional MLP VAE
Notes
[x] choose different colour scheme,
[ ] explain that if time/day were fully predictive we would expect to see standard Gaussian in latent space fully mixed and not divided by app, but since it's not we still see some delineation/clusters that are app-centric, however it is less than the case where only encode the app-vector, and this indicates that certain app-usage is most common at certain times and less so other times (obvious), i.e there is some temporal features to app-usage MLP encoder cannot capture and thus we expect an LSTM encoder to perform better.
[ ] when using lstm hidden state - we would expect to see some random points littering a well defined cluster of apps. For example, instant message randomly chilling in cluster of software development since now apps are encoded by their temporal representation and not just feature vector base don app usage in that 5 minute block
[x] VAE to visualize data
[x] Conditional VAE
[x] Seq2seq VAE to model the data
[x] Attention seq2seq model?
----- VAE LSTM + MLP -----
Notes