LALE: A Python Library Simplifying Automated Machine Learning

0
1082
Technology vector created by freepik - www.freepik.com

Whenever any data scientist thinks of developing a pipeline, they try bringing automated machine learning into the picture to make the task easier. However, due to inconsistent syntax and limited support for advanced features like topology search or higher-order operators, the development becomes tedious. Introducing a solution to this, IBM Research, USA has published a research paper on ‘LALE’: high-level Python interfaces’ library, which simplifies automated machine learning. 

The research tends to overcome the following shortcomings of previous research on the inconsistency of Auto-ML libraries: 

  1. There is inconsistency in pipeline specification syntax across the manual and automated spectrum. 
  2. User needs to learn different syntax to rewrite the code while switching between various Auto-ML tools. 
  3. Previous tools do not optimize the topology of the pipeline. 
  4. Invalid configuration while combining different hyperparameters 
https://github.com/IBM/lale/blob/master/talks/2019-1105-lale.pdf

Characteristics of LALE: 

  • LALE helps in selecting algorithms and tune hyperparameters of pipelines, compatible with scikit-learn
  • LALE provides a highly consistent interface to existing tools such as Hyperopt, SMAC, and GridSearchCV for automation. 
  • LALE uses JSON schema for checking correctness. 
  • LALE has an expanding library of estimators and transformers for interoperability. 
  • LALE uses Python subclassing to implement lifecycle states
https://github.com/IBM/lale/blob/master/talks/2019-1105-lale.pdf

 Users can install LALE just like any other Python package and edit it with off-the-shelf Python tools such as Jupyter notebooks.

Source: https://arxiv.org/pdf/2007.01977.pdf

Github: https://github.com/ibm/lale

LEAVE A REPLY

Please enter your comment!
Please enter your name here

This site uses Akismet to reduce spam. Learn how your comment data is processed.