Discrepancy-Based Active Learning for Domain Adaptation

This library offers several query methods for active learning, in the context of domain-expansion tasks :

• KMedoidsQuery: based on the K-medoids clustering algorithm;
• KMeansQuery: based on the K-means clustering algorithm;
• KCenterQuery: based on the k-centers algorithm;
• DiversityQuery: select target points with maximum average-distance to source points;
• RandomQuery: A baseline method that randomly selects samples from the target domain. To be used only in comparison with other methods.
• OrderedQuery: a basic method that should be used in conjonction with uncertainty predictions provided by the following approaches:
  • AADA: an hybrid active learning method for domain adaptation using a combination of entropy measure from a classifier and the outputs of a domain discriminator;
  • QBC: Query By Commitee;
  • BVSB: Best versus second best.;

Examples

A simple example with KMedoidsQuery is provided in notebooks/kmedoids_toy_data.ipynb.

An more comprehensive example is provided in notebooks/Superconductivity.ipynb, which demonstrates every method on the superconductivity benchmark dataset.