DSTK: The Data Science Toolkit

TravisCI build status:

This repository contains advanced tools for Data Scientists to use them as off-the-shelf solution. Feel free to use and contribute.

Setup

To setup, clone the repository and run

python setup.py install

Content

• Autoencoder:

  The AutoEncoder package contains a simple implementation using TensorFlow under the hood. It provides support for arbitrarily deep networks and tied as well as untied versions of the encoder. To create an AutoEncoder you need to provide a network architecture dictionary

  network_architecture ={
      'n_input': 100,
      'n_compressed': 2,
      'encoder': [100, 10],
      'decoder': [10, 100]
  }

  where n_compressed is the dimension of the compressed representation. The encoder is set up by calling

  sae = SimpleAutoencoder(network_architecture,
                          weight_regularization=1.0,
                          bias_regularization=0.1,
                          batch_size=25,
                          transfer_fct=tf.nn.softsign,
                          tied=False)

  To train the encoder you can provide a learning rate schedule

  schedule = {
      0: 0.001,
      250: 0.00025,
      2500: 0.0001,
      4000: 0.00005,
  }

  The training is then started with

  sae.train(data, 5000, learning_rate=schedule, display_step=100)

  The training is incremental, so you can continue training for more epochs after it finished, by just calling train() again. It will continue with the state it is in. To encode/decode a new data point you simply can call

  sae.encode(data)
  sae.decode(compressed_data)

  To monitor the progress during training, there is an internal recording dictionary that contains the losses of epochs, learning rate schedule and total number of training epochs

  sae._recording

  The implementation of TensorFlow also allows to produce the output of any arbitrary layer in the encoder. This can be done by calling

  Y = sae._monitor_layer(data, layer_num, network='encoder')

  where layer_num is the layer index of the encoder or decoder network.

• Generalized Additive Model (GAM) with Gradient Boosting:

  This package provides an implementation of a GAM algorithm proposed in the paper Intelligible models for classification and regression by Yin Lou, Rich Caruana and Johannes Gehrke (Proceedings KDD'12). This model has successfully been used in understanding clinical data for Predicting Pneumonia Risk and Hospital 30-day Readmission

  The basic idea is to learn the univariate shape function of an attribute, while optimizing the overall generalized model

  
  g(x) = f₁(x₁) + f₂(x₂) + ... + f_n(x_n)
  

  where g(x) is a nonlinear function such as the logit. The shape function f_i can be nonlinear themselves and are found using function approximation using a greedy gradient boosting machine.

  To instantiate a GAM use

  gam = GAM(max_leaf_nodes=10, min_samples_leaf=75)

  GAM leverages a scikit-learn sklearn.tree.DecisionTreeRegressor under the hood and hence exposes all its **kwargs. To train the GAM use

  gam.train(x_train, y_train, n_iter=10, display_step=2, leaning_rate=0.002)

  The algorithm can be trained iteratively, i.e. if it's not yet converged, calling train() again will use its last state to continue the training. Moreover, there is a training recording, that stores the number of epochs and various classification metrics (using the training set)

  gam._recording

• Bolasso:

  The Bolasso package provides and implementation of the Bolasso feature selection technique, based on the article Model consistent Lasso estimation through the bootstrap by F. R. Bach.

  This feature selection wrapper trains several sklearn LogisticRegressionCV classifiers with L1-penalty on a bootstrapped subset of the data. It keeps a running tap on the number of occurences a given feature appeared throughout all iterations.

  To instantiate the selector we run

  b = bl.Bolasso(bootstrap_fraction=0.5)

  We can then fit the selector using

  b.fit(data_df, target_series, epochs=5)

  If the results are not yet satisfactory, you can call this again and it continues to train the same Bolasso selector. To get the individual feature statistics we call

  b.get_feature_stats()