Melting-Point Prediction (Multi-level ML)
Effect of stacking order on ensemble performance. Each block = 10 tuned models (RF/XGB/LGBM/MLP)
R² ≈ 0.83
~3.041k samples (Citrination)
Custom decorrelated (1st) & stacking (2nd) layer
What: Predict melting points of organic compounds with a two-level stacking ensemble (RF/XGB/LGBM/MLP).
Why: Reproduced and extended a published approach (DOI) to validate best practices on a different materials system.
How: Parsed a ~3.041k dataset (Citrination); featurized SMILES with RDKit + custom bond-count features; Shap-guided model/feature, ensemble optimization; Custom-stacked tuned Ensemble.
Results: R² ≈ 0.83; custom stacking order improved baseline by ~4% out-of-sample.
