Transfer Learning : an Application of Unsupervised Domain Adaptation

Details

Supervisors

Saerens, Marco ; Lebichot, Bertrand

Faculty

Louvain School of Management

Degree label

Master [120] en ingénieur de gestion

Abstract

Transfer learning differs from machine learning thanks to its ability to take advantage of knowledge learned in a different but related task. It allows to reduce data collection cost and can also reduce the time required to solve the task at hand. The present work arises from a real-life situation: the classification of three rheumatologic disease variants. This task is complicated by the lack of labelled data. Hence, knowledge is retrieved from similar applications performed in Netherlands and Germany. This application is solved using the CRISP-DM methodology and a specific evaluation strategy is developed. The transfer methods are evaluated thanks to two tools: first, an analysis of the mutual transfer between Netherlands and Germany and; second, a comparison of the predicted probability distribution with the expected one (which is determined by surveys) when transferring to Belgium. As a result of this work, several models are proven to be most performing than a naïve transfer from Netherlands or Germany to Belgium. The most performing model found is an instance-based transfer: Kernel Mean Matching.