Towards a Transfer Learning Strategy in Full Model Selection Algorithm for Temporal Data Mining

Abstract

Traditional machine learning techniques were designed for training for scratch depend on the current feature-space distribution. In many real applications, the fact to obtain new data for training and rebuilds models could become expensive or impossible. Therefore, from a lifelong machine learning conceptualization, transfer learning can be indeed bene cial to speed up the time it takes to develop and train a model by reusing an isolated pre-training setting as a starting point for another target domain, especially when multiple tasks and hyper parameter optimization are considered, such as a full model selection approach. This document presents an early transfer learning strategy based on a decision tree powered by full models for temporal databases trained in an isolated way with di erent search methods. The proposed transfer learning strategy is capable to suggesting the starting point and the search method adopted by the full model selection approach.