Knowledge graphs can provide relational features and causal structure for transfer learning.

Relational transfer learning extends ordinary feature transfer by moving relationship networks between a source and target domain. The RF-TL framework constructs knowledge graphs through an extended structural-equation model and uses causal analysis plus counterfactual inference to guide transfer. In experiments involving sleep-apnea questionnaire data and COVID-19 ICU-admission data, the proposed method reportedly achieved more accurate predictions with fewer input features than TCA and CORAL, though those results are task- and dataset-dependent.