The KNIME Active Learning plugin comprises a set of KNIME nodes for modular active learning and novelty detection in KNIME. Active learning methods use feedback from the user to selectively sample training data.
KNIME Active Learning can be installed form the KNIME-Labs update site (minimum version is KNIME Analytics Platform 3.1).
KNIME Active Learning models the active learning process with the Active Learn Loop. The management of the data takes place in the Active Learn Loop Start, the labeling (assigning class labels to rows) in the node end. The creation of the query for the oracle takes place inside the the loop.
This example illustrates the active learning process with KNIME Active Learning:
- It starts with the Active Learn Loop Start node and ends with one of the Active Learn Loop End nodes.
- Each unlabeled row is assigned a score in the Score module.
- In the Select module, one (or more) rows are selected for labeling.
- The selected rows are then assigned a class label in the Active Learn Loop End node.
Examples on KNIME Hub
You can view and download all the nodes and Active Learning workflows from the KNIME Hub.
Active Learn Loop
The "Active Learn Loop" nodes provide the framework for the active learning process. Each active learning process starts with the Active Learn Loop Start node and ends with one of the Active Learn Loop End nodes:
- Active Learn Loop End: This node provides an interface for a human oracle to label the selected rows.
- Auto Active Learn Loop End: This node provides an automated oracle for fully labeled datasets. It can be used for verification and testing.
Scorer nodes are nodes which calculate a score for each row that describes its relevance for the active learning process. KNIME Active Learning provides scorer nodes grouped in the following categories:
- Uncertainty: Nodes in this category calculate their score based on a class probability distribution which is an configurable output of many predictor nodes.
- Density: Nodes in this category calculate and update a score initially based on the density of the featurespace.
- Novelty Detection: Nodes in this category calculate their score based on novelty detection methods, e.g. a Kernel Null Foley-Sammon Transformation.
- Combiner: Nodes in this category calculate aggregation scores out of the combination of scores calculated by other scorers.
- All in one: Nodes in this category provide scorers which package modular algorithms into a fixed package for increased performance.
The "Element Selector Node" selects the n elements with the highest score.
KNIME Active Learning is available under the same GPL - Licence as KNIME.