Numerai Experiments

Folder structure:

• ensemble.py - combines multiple predictions using geometric mean
• fit_tsne.py - uses this t-SNE implementation for 2D embedding (does not work in 3D)
• search_params.py - uses RandomSearchCV for hyperparameter search
• tpot_test.py - runs tpot over the data
• tpot_pipeline.py - best tpot model
• notebooks/ - contains Jupyter notebooks
• bh_tsne/ - is the original C++ t-SNE implementation with scripts for converting the csvs to the format the binary expects
• models/ - various model implementations
  • adverarial/ - generative adversarial model that saves the learned features for each sample
  • autoencoder/ - simple autoencoder with regular and denoising variants (also saves learned features)
  • classifier/ - simple neural network classifier
  • pairwise/ - pairwise model implementation described in the blog post
  • pipeline/ - various scikit-learn models
    • estimators.py - custom wrappers around KernelPCA and Isomap that fit on a small portion of the training samples to avoid memory errors
    • transformers.py - contains ItemSelector which allows for selecting data by a key when building pipelines (source)
    • fm.py - factorization machines
    • lr.py - logistic regression with t-SNE features
    • pairwise.py - sklearn variant of the pairwise model
    • simple.py - simple logistic regression with polynomial features