Transductive vs Inductive learning

Induction From Nature of Statistical Learning Theory

  • Induction: Learning a general model from specific examples.
  • Deduction: Using a model to make predictions.
  • Transduction: Using specific examples to make predictions.

Transduction example: Nearest neighbour algorithm is an example of Transduction. We only take the nearest neighbour values to obtain the value of the new sample.

Induction example: Supervised learning tasks are Induction \rightarrow Deduction as they learn a model from example and then make predictions at new values.

Is semi-supervised (SS) learning induction or transduction?

Yes. It depends on the task.

In semi-supervised learning we have labelled and unlabelled training data.

  • If our task is to predict an unlabelled (test) dataset from labelled and unlabelled training data, then we have Induction \rightarrow Deduction.

    SS induction example: we have labelled unlabelled images of MRI scans (1 = cancer, 0 =no cancer) and we want to create a Convolutional Neural Network that predicts the labels of new unseen MRI scans.

  • If our task is to label the unlabelled training data, given labelled and unlabelled training data then we have Transduction.

    SS transduction example: We may have proteins represented as nodes in a Graph Neural Network, with protein-protein interactions as edges. Our labels are (1 = influenced by drug, 0 = not influenced by drug). The semi-supervised problem of transduction here is predicting if our unlabelled proteins (which the network is trained on!) are influenced by the drug.

    However, we may have a Graph Neural Network (GNN) where each node is either labelled or unlabelled and our task is to

Therefore depending on the task, semi-supervised learning may involve either Induction or Transduction