Make it easy to analyze tabular datasets with lots of features in scikit-learn.

Although scikit-learn has selectors that let you select features based on their data types, sometimes you want to select features based on mutual information. In order to do this, metadata needs to be computed, similar to how tensorflow extended, and specifically, tensorflow data validation does it. On top of that mutual information needs be computed for each feature to determine which features have predictive power for the problem at hand.

Tabular knowledge makes it easier to understand data in a scikit-learn/pandas environment and provides functions to make it easy to filter out features that do not have any predictive power.

A working example of its use is available in src/client/client.py which implements a typical workflow that can lead to a baseline model. Specifically, computing metadata, fine tuning semantics of features, analyzing mutual information, basic encoding with pipelines, followed by a model.

It may be possible to build policy driven pipelines in the future. For example include features with mutual information above a certain threshold, where the selector behaves like make_column_selector in scikit-learn. Secondly, the encoding could also be declarative based on model type.