kts 0.2.45

Competition-oriented framework for interactive feature engineering and building reproducible pipelines

kts

kts was created to simplify and unify process of solving machine learning competitions.

Key Principles

• Feature engineering is modular:
  • Each feature engineering function represents a block of features: takes an input dataframe and produces a new one with only new features
  • Such functions called feature constructors are then assebmled into feature sets which are then used for training models
  • Features are computed once, then they are loaded from cache
• A pair of a model and a feature set is validated using a given metric and cross validation splitter
• Once experiment is conducted, it is placed to your local leaderboard, trained models and sources are saved
• Each experiment is given an ID, which is used to access it
• Each experiment can produce predictions for any dataframe which has same columns as a training one: feature engineering is done automatically, then features are fed to trained models

Features

• We support features which should be computed differently for training set and validation set: they are implemented as simply as usual ones using special syntax (df.train and df.encoders attributes of a dataframe passed to function)
• User cache: you can store any objects to access them from other notebooks of your project with kts.save and kts.load
• Standard library: some common feature generation techniques are preimplemented, like target or one hot encoding; you can use their sources to borrow best practices of writing custom feature constructors in kts style
• Designed for multiple notebook environment: cache is synchronized between notebooks, e.g. you can change source of a feature constructor in one notebook and get it automatically changed in other one; same for objects
• Stacking is as simple as kts.stack(IDs): it creates a standard feature constructor which can be used for feature set creation
• Feature selection: select best features from an experiment using built-in feature importances calculator (sklearn-style) or permutation importance. You can also implement your own feature importance calculator using our base class

Getting started

Use $ pip3 install kts to install the latest version from PyPI.
Check kts-examples repo to learn basics.

Command line interface

Use it to create a new project:

$ mkdir project
$ cd project
$ kts init

or download an example from kts-examples repo:

$ kts example titanic

Contribution

Contact me in Telegram or ODS Slack to share any thoughts about the framework or examples. You're always welcome to propose new features or even implement them.

Acknowledgements

Core of the project was designed and implemented by the team of Mikhail Andronov, Roman Gorb and Nikita Konodyuk under the mentorship of Alexander Avdyushenko during a project practice held by Yandex and Higher School of Economics on 1-14 February 2019 at Educational Center «Sirius».