Feature Engineering

methods

• normalizing and standardizing

• non separable problem

• technics list

  • lab Features Transform case1

  • lab Features Transform case2

kernels

The idea behind the kernel trick is to implicitly map data to a higher-dimensional space where it becomes linearly separable, enabling linear algorithms to solve nonlinear problems. This mapping is done through a kernel function, which computes the inner product between the images of two data points in this higher-dimensional space. Essentially, the kernel function measures the similarity between pairs of data points.

• recap dual

• define kernel

• lab kernel types demo

• mnist example

• lab fashion-mnist

Feature Engineering Tools

• preprocessing

• pipe

• grid search demo

• autouml

summary

* continuous features
    * combination of features 
        * polynominal
        * normalization and standardization
        * change of coordinates
    * transforms:
        * kernels
        * STFT
        * wavelet
        * dictionaries

* discrete features
    * encoding (one-hot)
    * grouping (valid combinations)    

* missing value
    * dropping (features/ samples)
    * imputer by a model - nearest / interpolation / classifier / regressor

* data / time
    * day of the week/month/year

* text
    * steamming
    * lemmatization

* feature selction

* dimensionality reduction

* automatic feature generation (autoML)

* stemming vs lemmatiztion

TBD

https://www.kaggle.com/code/willkoehrsen/automated-feature-engineering-tutorial https://www.kaggle.com/code/willkoehrsen/automated-feature-engineering-basics