DataHack Datathon, 3rd Edition

The challenge

DataHack 3.0 ran seven data challenges at once, and teams were ranked on their combined results. Among them were Pharma-Semantic, which asked for a clinical alert about two drugs, written from their biological profiles alone with the drug names masked, and Well Log Reconstruction, which asked for missing measurements along oil wells.

My contribution

I built our solutions to both. For Pharma-Semantic, I retrieved the training cases closest to each drug pair and built the alert from them, checking the meaning with BioBERT. For Well Log Reconstruction, I trained two different models, a temporal convolutional network that reads each well as a sequence and a LightGBM model on the surrounding measurements, then blended them, tuning the blend weights for each measurement on out-of-fold predictions. That blend is what moved us up the leaderboard.

The outcome

We placed second on Pharma-Semantic and third on Well Log Reconstruction, and our team, Hybr1ds, finished first in the overall ranking.