MechanismOfActionPredictions

The notebooks from the Kaggle competition Mechanisms of Action (MoA) Prediction from Andy and Andre. The notebook TabNet+ANN|0.01612LB was our best submission, scoring a top 4% in the leaderboard. A brief overview of our solution:

Preprocessing/Feature Engineering

• Removed control group
• Deleting the features with an information gain of 0.06 or less, acheieved using the mutual_info_classif function from sklearn.
• Perform PCA, selecting the top 20, 60 features from cells, genes, respectively, then adding them back to the original data.
• Perform normalization to the data using the GaussRankScaler.
• Calculating the statistics of the features, performed on genes and cells seperately.
• Added the cp_dose_time feature by combining cp_time and cp_dose.
• Label encoded cp_time and cp_dose_time, then one hot encoded the rest of the categorical features.

Models and Training

• We trained two TabNets on different seeds and one two layer Neural Network, then blended the predictions with weighted average.
• Optimizers: TabNet, AdaBelief. NN, AdamW.
• LrScheduler: ReduceLROnPlateau
• Validation Scheme: MultiStratifiedKFold
• Folds: 10