Active Learning - Exploration / Exploitation Score

This workflows shows examples of Active Learning with the Exploration vs Exploitation strategy taken from Cebron et al, 2008, called Prototype Based Active Learning (PBAC). Exploration vs exploitation consists of selecting instances with missing labels based on the uncertainty of the model (exploitation) and density of its neighborhood of unlabeled instances (exploration). This metric is computed by Exploration/Exploitation Score Combiner node. The two measures are combined in a single metric, which is then used to select the instances to be labeled in the human-in-the-loop strategy.

Legacy Example - before KNIME Analytics Platform 4.1 (Dec. 2019):
This workflow shows an example of Active Learning. We read a simple dataset of images separated in two classes and calculate some features on them. Now the "Active Learning Loop" determines the best sample that could be manually labeled by a user and benefits most the separation of the classes. This workflow can be used to perform Active Learning on the KNIME Analytics Platform, however it cannot be deployed on the KNIME WebPortal as a remotely accessible web-based application.

Current Example - since KNIME AP 4.1 (Dec. 2019):
With KNIME Analytics Platform 4.1 the Active Learning extension was updated to support interactive JavaScript views for KNIME WebPortal. In this example you can interactively label the instances using KNIME WebPortal. More infos on KNIME WebPortal in the link below. The new Active Learning Loop is quite similar to the Recursive Loop, but enhanced with ports and instructions for active learning.

External resources

• Active learning for object classification - Nicolas Cebron et al - Data Min Knowl Disc (2009)
• Burr Settles, Active Learning Literature Survey, 2010 - Chapter 3.1 Uncertainty Sampling
• KNIME Blog - Labeling with Active Learning
• knime.com - KNIME WebPortal