README.md

Useful AI computing procedures

Environment setup

• What files are contained in training and test folders?
• Create symbolic links to access the file from another location without duplicating it
• Suppress warning messages

Dataframe manipulation

• Read data
  • Check column names
• Deep copy features
• Slice and index a dataframe
  • Select target columns
  • Randomly select columns
  • Drop a columm
  • Create balanced folds in cross-validation for continuous target, where each each bin/stratum includes a roughly equal number of samples
• Split data into training and validation sets before any fitting of preprocessing steps (e.g., imputation) to avoid train-test contamination
  • Use train_test_split
  • Randomly select rows

Data exploration

• Explore data with summary statistics
• Generate a bivariate relationship table

Feature engineering

• Encode categorical variables
  • Which variables are categorical?
  • Ordinal encoding
  • One-hot encoding
  • Should the year be encoded as an original integer or treated with one-hot encoding? It depends. If the year has a meaningful numeric relationship with the target (e.g., older years might correlate with higher frequency of a disease), we keep it as an original integer. However, if specific years have unique relationships with the target, the year is better treated with one-hot encoding.
• Manage missing values
  • Impute missing values
  • Drop missing values
• Normalize a feature

Build a model: define -- fit -- predict -- evaluate

• Train a decision tree regressor
• Train a decision tree classifier
• Train a random forest regressor
• Train a gradient boosting regressor
• Save model parameters and full model

Make a pipeline that bundles imputation, one-hot encoding, model training, and model evaluation

• Pipeline

Fine-tune a model for better performance

• Evaluate the impact of different arguments on model performance
• Use cross-validation (CV) to evaluate the impact of different arguments on model performance; CV is suitable for small datasets, as computational burden isn't a big issue
• Define several models with varied arguments and select the model with the best performance
• Add more relevant features
• Use another model algorithm (e.g., graph neural network to process connectome)