Abstract
In transfer learning, a source domain often carries diverse knowledge, anddifferent domains usually emphasize different types of knowledge. Differentfrom handling only a single type of knowledge from all domains in traditionaltransfer learning methods, we introduce an ensemble learning framework with aweak mode of convergence in the form of Statistical Invariant (SI) formulti-source transfer learning, formulated as Stochastic Ensemble Multi-SourceTransfer Learning Using Statistical Invariant (SETrLUSI). The proposed SIextracts and integrates various types of knowledge from both source and targetdomains, which not only effectively utilizes diverse knowledge but alsoaccelerates the convergence process. Further, SETrLUSI incorporates stochasticSI selection, proportional source domain sampling, and target domainbootstrapping, which improves training efficiency while enhancing modelstability. Experiments show that SETrLUSI has good convergence and outperformsrelated methods with a lower time cost.