A team of researchers, led by Prof. Jaesik Choi in the School of Electrical and Computer Engineering at UNIST recently had their paper accepted to International Conference on Machine Learning (ICML 2016).
Their paper titled “Automatic Construction of Nonparametric Relational Regression Models for Multiple Time Series” was accepted for presentation at the 33rd International Conference on Machine Learning, which was held on June 22, 2016 in New York, USA.
The paper proposes an extension of a general kernel learning framework to handle structure that could be present across multiple datasets. The idea is to exploit some sort of multi-task learning in the sense that it is expected that the different datasets share some part of the covariance function. At the same time, the method proposed by the authors considers a part of the covariance function to be specific of each dataset.
This figure shows the data analysis of General Electric stock after the 9.11 attacks. While Relational Multi-Kernel Learning (RKL) model (c) and (d) finds and explains the sudden drop in stock, the Compositional Kernel Learning (CKL) method, (a) and (b) do not.
In the study, Prof. Choi and his team proposed a nonparametric Bayesian framework that finds shared structure throughout the multiple sets of data. To validate their approach, the team applied this model to a synthetic data and several real world data, including stock data, the house market, and currency exchange rate data. The results show that the relational kernel learning methods find more accurate models for regression problems on real-world data sets.
The resulting Relational Kernel Learning (RKL) method provide a way of finding a shared kernel function that can describe multiple data with better Bayesian Information Criterion (BIC). Prof. Choi notes that “We expect that this research finding will get lots of attention from the financial industry.”
This research has been supported by Basic Science Research Program through the National Research Foundation of Korea (NRF), funded by the Korean Ministry of Science, ICT and Future Planning.
ICML is the leading international machine learning conference of the International Machine Learning Society (IMLS), providing a venue for the presentation and discussion of current research in the field of machine learning.
Journal Reference
Yunseong Hwang, Anh Tong, and Jaesik Choi. “Automatic Construction of Nonparametric Relational Regression Models for Multiple Time Series,” Proceedings of The 33rd International Conference on Machine Learning, pp. 3030–3039, (2016).
