jameschartouni / numerai

Code from my experiments on Numerai

Home Page:https://medium.com/@jimfleming/notes-on-the-numerai-ml-competition-14e3d42c19f3#.p3swptim4

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

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

About

Code from my experiments on Numerai

https://medium.com/@jimfleming/notes-on-the-numerai-ml-competition-14e3d42c19f3#.p3swptim4

License:MIT License


Languages

Language:Jupyter Notebook 98.8%Language:Python 0.6%Language:C++ 0.5%Language:MATLAB 0.0%