Active Learning

In this python notebook, different Active Learning strategies combined with diversity criteria are compared in order to find the best set of labeled observations to train a SVM classifier.

Problem description

The Semeion Handwritten Digit Data Set is a database of handwritten digits. Each record represents a handwritten digit, orginally scanned with a resolution of 256 grays scale (28). The goal is to obtain the best possible result by minimizing the number of labeled observations used to train a classification model by selecting samples that provide the maximum information possible.

Active Learning strategies

• MS (margin sampling)
• MCLU (multi-class label uncertainty)
• SSC (significance space construction)
• nEQB (normalized entropy query bagging)

Diversity criteria

• MAO (most ambiguous and orthogonal)
• MAO lambda
• diversity by clustering.

Results