Auto-encoder experiment

[aquemy]

Right now we have two sets of features:

• descriptive, available before the judgement
• textual, available after the judgement

We showed that textual, even with the conclusion removed, is much better to make prediction on the outcome. However, in practice, we can use only descriptive feature to make a prediction.

The idea would be to create a predictive model on both types of features as it performs best but also to create a mapping Phi via an auto-encoder between the space of descriptive features to the space of textual features. That way, to predict the outcome only with the the descriptive features desc, we could enhanced the case with Phi(desc). In theory we should obtain improvement compared to learning only over the descriptive features.

The main problem I see is the difference dimensions between the two spaces, and the potential lack of data for such a complex encoder.