ML & DataKaggle competition · University of Helsinki2025Completed
New Particle Formation Classification
5th out of 149 teams in the University of Helsinki ML course competition.

- final leaderboard rank
- 5th / 149
- final leaderboard rank
- binary accuracy
- 93.1%
- binary accuracy
- features after selection
- 102 → 12
- features after selection
Overview
Classifying days with and without new particle formation events from 102 atmospheric measurements. Feature selection reduced the input to 12 features without losing accuracy.
Approach
- Statistical tests, tree-based importances and L1 regularisation combined to select features.
- Logistic regression and random forests with class weighting, because the classes were imbalanced.
- Cross-validation for model selection. Accuracy on the harder four-class version of the task was 70.2%.