Estimates probabilistic labels by learning a generative label model from the provided noisy labels. This node is a key component for the realization of weak supervision approaches as popularized by Snorkel . The idea in weak supervision is that it is often possible to create a number of simple inaccurate models (e.g. simple rules or existing models for slightly different tasks) that can label unlabeled data and that the agreements and disagreements of these simple models can be analyzed to infer information of the true label. Our implementation is a TensorFlow based adaptation of the matrix completion approach proposed in this paper by the Snorkel team. We refer to the publication for details on the strategy.
Options
Input Ports
- Table containing label sources. A label source is just a nominal column. Note that missing values in a label source are interpreted as abstains i.e. it is assumed that a missing value indicates that the label source did decide not to label the corresponding row. Label sources without a set of possible values assigned are ignored during the computation and a corresponding warning is displayed on the node.
Output Ports
- Contains for each row in the input additional probabilistic label columns. That is for each of the possible classes, a column which gives the probability that a row is an instance of that class.
- Each row in this table gives the conditional probabilities that the label source displayed in the Label Source column takes on a specific value given the true label displayed in the Latent Label column.
Views
- This node has no views
Workflows
- No workflows found
Links
- No links available
Developers
You want to see the source code for this node? Click the following button and we’ll use our super-powers to find it for you.