A helper library for data science pipelines

"Something that is made of many parts and yet is somehow more than or different from the combination of its parts"

Machine learning (ML) has achieved considerable success in multiple fields, much of the pipeline surrounding machine learning can be generalised to a point that removes the 'engineering' of the data flows and pipelines leaving the human expert to work on the most difficult aspects:

1. Pre-processing the data
2. Feature Engineering
3. Select appropriate features

The goal of Gestalt is to remove the cumbersome parts of building a meta-learner data science pipeline (i.e the process of building a generalised stacker across folds).

The current roadmap for this module is as follows:

1. Build P.O.C for GeneralisedStacking class using pandas - Done
2. Build Stackers for, Regression and Classification problems (binary and multiclass) - Done
3. Create plugin wrapper to expose other algorithms to the stacking process (say R libraries) - Done
4. Support scipy.sparse data to allow both dense and sparse models to be run on the same folds -Done
5. Create an example of bayesian encoding as a transformer that runs across folds in line with the stacker.
6. Create Hyper-parameter autotuning class for the set of base models to be used in the metaleaner
7. Add loads of tests and documentation.

As with all alpha OSS projects things are under constant development and already one can see places where refactoring some of the core code would make sense (i.e a Class to remove the cumbersome _predict_# and _fit_# methods)

pip install -U git+https://github.com/mpearmain/gestalt

In the examples dir there is a set of examples showing various different use cases. To run the r wrapper for ranger you need to have a copy of rpy2 and a copy of the ranger R library locally. NOTE: You can even runu Vowpal Wabbit inside the stacker with the VWClasssifer and VWRegression sklearn wrappers from pip install vowwpalwabbit