Setting Parameters for Support Vector Machines using Transfer Learning

Abstract

Machine Learning algorithms have a broad applicability, although generally a huge effort is necessary to find a good configuration for a given task. The tuning of free parameters, for example, is a task that directly affects the algorithm’s performance but is often carried out as an ad hoc process. An alternative approach is to define the problem as a search in the parameters space, which can be computationally expensive and slow. Furthermore, to apply an algorithm to a different problem, all the work must be done from scratch. Transfer learning can be used to avoid this rework. In this paper we propose an approach to tune the parameter by means of transfer learning. The idea is to use data complexity characterization measures to evaluate the similarity among datasets and evaluate whether they share similar configurations of good parameters. To compare our approach, four performance measures were used: area under ROC Curve (AUC), accuracy, F1 and area under Precision-Recall Curve. Results show that the proposed approach may reduce the search space for the parameter tuning by exploiting parameter recommendation of similar datasets and provide competitive performance compared to other widely used techniques.