Active Learning Brief Introduction

Active Learn introduction.

Why?

In some case, it is very expensive to label samples, or labels are extremely unbalance. E.g. risk control.

Active Learning can reduce label cost campared with random samples labeling.

How?

  1. Random sampling.
  2. Label samples, train model.
  3. Predict, get middle samples (score is round 0.5).
  4. repeat 2 and 3 until label resource is spent or model is enough good or samples can not be labeled.

Diffcults

Don’t use Active Learning in Financial case becuase bad samples are very sensitive to profit.

Internet case can use Active Learning. E.g. CV and NLP

Written on September 23, 2019