Sensitivity Analysis for an Ensemble Transfer Learning scheme for Time Series Forecasting–Case-Study of a Shallow Neural Network Achitecture

Auteur(s): Kyamakya Kyandoghere Jean Marie Vianney
Auteur(s) (S/D): 2023 IST-Africa Conference (IST-Africa)
Titre de l'ouvrage: PROCEEDINGS OF 2023 IST-Africa Conference (IST-Africa)
Editeur: IEEE
Lieu: DOI: 10.23919/IST-Africa60249.2023.10187750
Année: 2023

Résumé

DOI: <a href="https://doi.org/10.23919/IST-Africa60249.2023.10187750">10.23919/IST-Africa60249.2023.10187750</a> Transfer learning (TL), applied in the context of time-series forecasting, is an important hot topic nowadays in machine learning. This paper addresses the gap identified in most recent survey papers that empirical studies, currently missing, are very necessary to come up with guidelines for TL approaches and TL method design selections that can be used by practitioners. In this perspective, this paper does overall suggest the skeleton of a comprehensive sensitivity analysis methodology for TL schemes w.r.t. to a given machine learning model at hand. As a first step, five relevant TL performance metrics are suggested and comprehensively defined. Then, the core steps of a comprehensive TL-related sensitivity analysis are formulated. For illustration, an mini-sensitivity analysis is conducted on an MLP shallow network. Although relatively small, this mini-sensitivity analysis does confirm the usability of the suggested methodology as it does already, nevertheless, highlight some interesting insights. The end-product of this project can culminate in designing a pretrained model that can be useful to time-series forecasting practitioners.