Closed yonromai closed 11 months ago
Gist: These new features capture a lot of useful signal to classify the labels (albeit significantly overlapping with the content of the text embedding features)
This is really great to know. I suspect cancer, and the associated TA feature, is the biggest contributor to this effect.
I'm somewhat indifferent on whether or not to keep the embedding features now in a final model, cf. https://github.com/related-sciences/nxontology-ml/issues/8#issuecomment-1736098847. I lean towards keeping them though since I know there are a lot of text patterns in many parent vs child relationships that should be important for this classification, and propagating the text terms across the ontology structure may be the best way to exploit that in the future (e.g. via GNNs).
This PR introduces new logic to include Therapeutic Areas as features in the CatBoost model (see issue #33).
ta_mae
(TA + Topological features) andpca64_ta_mae
(PCA 64 + TA + topo features).Gist: These new features capture a lot of useful signal to classify the labels (albeit significantly overlapping with the content of the text embedding features):![image](https://github.com/related-sciences/nxontology-ml/assets/1596570/8ca6fa8e-4d8d-4974-acc0-ff4ed78757f0)
Notes:
cc @eric-czech @dhimmel