Metafeatures

Guide to metafeatures module

Overview

This module provides functionality to extract metafeatures for a dataset split (using only the training set). The main entrypoint for the functions in this module is Featurizer.py. The functions in this script can be used to extract metafeatures for the full metadataset and save them in a .csv file, and also to extract metafeatures given a new dataset.

Using the featurizer

To extract all the metafeatures for all of the datasets:

1. Run fetch_data.sh, which gets all dataset splits from the Google Cloud bucket and places them in a local folder, ../all_data.
2. Run Featurizer.py, which extracts metafeatures for all dataset splits in ../all_data. The resulting metafeatures are saved to Metafeatures.csv

To extract metafeatures for a dataset split specified in some location, use featurize_dataset_split() within Featurlizer.py.

Implementing new features

This section is only relevant if you wish to add metafeatures to the featurizer.

Featurizer.py keeps track of metafeatures to extract in two data structures: feature_func_lookup and all_features. A metafeature is obtained by calling a function which returns a dictionary with one or several metafeature values. If some metafeature is obtained by calling a function named foo with arguments given in an OrderedDict kwargs, then there will be an entry ("foo", kwargs) in all_features (the use of an OrderedDict is so that the feature can be converted into a string consistently). Furthermore, feature_func_lookup must contain an entry with the string "foo" mapping to function foo.

New features may be implemented by following this general workflow, as long as all_features and feature_func_lookup in Featurizer.py end up containing the necessary entries. The main sets of features currently implemented are given in Basic.py, Landmarkers.py, and DistributionFeatures.py, from which all_features and feature_func_lookup are currently populated.