Feature Extraction

Transforms the orignal set of features into a into a different dimension of less features

Feature creation #

Creating feature which has more info that the original dataset.

Methodologies #

  • Feature extraction

    • Domain specific
    • In multimedia, low-level (pitch, tonality), mid-level (fourier, wavelet), high-level(tag, genre)

  • Mapping data to new space

    • Fourier trans (freq domain)
    • Wavelet trans (freq + time domain)
    • Scale-Invariant Feature trans (SIFT) (Capute important/minuate points)

  • Feature construction

    • Create dummy features
      • Converting categorical data to numerical (neural N/w works only on numerical)
    • Create dervived features
      • Customer time duration on a webpage

  • Tricks

    • Handling Time, date and address difference
    • Handling sales and marketing data; Create proportion like sales by marketing, by person
    • Handling large volume of transaction - Document data given log transformation
    • Special object as influencing person