nathanntg/lin-train

Linear Regression Feature Selection and Trainer

lin-train

Linear Regression Feature Selection and Trainer

This is based on a linear regression trainer and feature selection class initially developed to help analyze and make predictions for the MIT Big Data Challenge. The trainer can use any provided solver to perform a linear regression (by default, it uses the numpy provided linear least squares regression). The training class provides a simple way to do feature selection over a large feature space. The trainer does k-fold cross validation to find features that improve validation scores. When complete, the class has the model coefficients as well as a score.

Dependencies: Python 2.7, numpy

Usage:

from lintrain import Trainer

t = Trainer(x, y) # x is a matrix, y is a vector or single column matrix
t.debug = 2 # print detailed debugging information regarding the feature selection
t.run_forward_selection() # run forward feature selection

print t.column_indices # indices of columns in "x" that were selected
print t.fit # linear coefficients
print t.score # score identified by the trainer

novel_y_prime = t.apply_to_vector(novel_x) # applies column selection and plugs values into linear equation

Methods

The following attributes are available for instances of the Trainer class.

• run_forward_selection(initial_columns=None, initial_score) Will begin considering each feature (column in X) sequentially, and continue adding columns to the linear Regression as long as the k-fold cross-validation score improves. You can optionally specify initial columns, which will be used as a starting place (initial columns will remain in the regression, since forward selection only adds columns).

• run_backward_selection(initial_columns=None, initial_score) Will begin considering each feature (column in X) sequentially, removing columns from the linear Regression as long as the k-fold cross-validation score improves. You can optionally specify an initial set, which will be used as a starting place (the final result will be a subset of the initial columns). By default, the backward selection uses all columns as the initial set.

• run_bidirectional_selection(initial_columns=None, initial_score) Will consider adding OR removing features (columns in X), and will continue to do so as long as the addition or removal improves the k-fold cross-validation score. You can optionally specify initial columns, which will be used as a starting place. By default, the the initial column set is empty, so it will always begin by adding columns.

• select_columns_from_matrix(p_x) Takes a matrix where rows represent entries and columns represent features, and returns a matrix with just the features (columns) that were selected by the feature selection algorithm. Must be called after running the trainer class. The order of returned columns matches the order of the features in column_indices.

• select_columns_from_vector(a_x) Takes a vector where entries represent features related to a single entry, and returnsa vector with just the features (columns) that were selected by the feature selection algorithm. Must be called after running the trainer class. The order of returned columns matches the order of the features in column_indices.

• apply_to_matrix(p_x) Applies the feature selection process to novel values and returning a vector with the value predicted. This first selects the features from the matrix of novel entries. Using the dot product, it multiplies the coefficients and sums the response to return the predicted values.

• apply_to_vector(a_x) Applies the feature selection process to novel entry and returning a single value (the prediction). This first selects the features from the vector of features. Using the dot product, it multiplies the coefficients and sums the response to return the predicted value.

Attributes

The following attributes are available for instances of the Trainer class.

• debug Allows printing of information about the training process. Can be 0 (no debugging), 1 (minimal debugging) or 2 (detailed debugging). Minimal debugging prints final scores and such data, while detailed debugging prints individual feature (column) additions and removals.

• number_of_folds The number of folds to use in the k-fold cross-validation. Defaults to 5.

• solver An instance of a class that inherits from solvers.solver.Solver, which will implement the actual linear training. An instance of solver offers to functions. The first calculates the model parameters, based on the training data set x and y; that is: p = f(x, y). The second applies the model parameters to predict new values of y; that is: y_hat = g(x, p). By default, the system uses the provided LeastSquares class that uses the linear least squares regression.

• scorer An instance of a class that inherits from scorers.scorer.Scorer, which will use a set of predicted values (y_prime) and actual values (y) to calculate a score representing how closely the predictions match the actual values. By default, the system uses the provided MeanAbsolute class that calculates this as the Mean Absolute Error (MAE).

• score_ After running, this contains the score for the final fit. This is calculated based on the scorer provided.

• column_indices After running, this contains the column indices for the matrix X representing the features that were selected to best minimize the predictions without over-fitting to the data provided. (Depending on how the class is called, features are either iteratively added or removed as long as the k-fold cross-validation score continues to improve).

• fit The linear coefficients that correspond with each column in column_indices.

Multiprocess

To decrease the time for training, a Trainer class that runs across multiple Python processes is available. It is a drop-in replacement for the Trainer function in the example above, and will default to using one process per CPU.

To use, replace the import line above with the following:

from lintrain import ParallelTrainer as Trainer

To change the number of processes used, an added attribute is available:

• number_of_processes An integer specifying the number of processes to use (defaults to the number of CPUs available.

Utilities

lin-train includes a few utility functions to help generate features, specifically turning discrete integer features into columns of binary features representing different potential discrete values. This is useful for turning day of the week or time of day values into a set of binary values (e.g., "is Monday", "is Tuesday", etc).

• utilities.add_int_as_categories(A, int_val, val_max, val_min=0, step=1) Appends new columns onto the matrix A based on the integers contained in the vector or column matrix int_val (note that int_val must contain the same number of entries as the number of rows in A, as each integer value corresponds with one entry in the matrix). Integers in int_val must be discrete and between val_min and val_max (inclusive). Binary columns (either 0 or 1) are appended to A representing each discrete value based on the step size.

• utilities.int_as_categories(int_val, val_max, val_min=0, step=1) Creates a list of binary values representing the discrete value int_val as binary features. The integer in int_val must be discrete and between val_min and val_max (inclusive). The binary columns (either 0 or 1) returned represent each discrete value based on the step size.