Reddit post blueprint for time series:
- from a stack classifier by extracting just one model output
- How about a sentiment analysis ?
There are 2 metodologies for training:
-
Individual Model Deals with a twit as input and outputs a probability (confidence) for the for the user to be depressed or in control.
-
Time Series Model Deals with the output probability of the the individual model for a set of twits. The twits are selected using a rolling window of a fixed size and a gap as a stride.
Using the label of the eRisk dataset can be dummy because we don't have a timestamp (or a twit index) for the decising message that can provide a diagnostic, along with a window_size historic.