In this paper, the problem of a nonlinear model – specifically the hidden unit conditional random fields (HUCRFs) model, which has binary stochastic hidden units between the data and the labels – exhibiting unstable performance depending on the hyperparameter under consideration.
There are three main optimization search methods for hyperparameter tuning: manual search, grid search and random search. This study shows that HUCRFs’ unstable performance depends on the hyperparameter values used and its performance is based on tuning that draws on grid and random searches. All experiments conducted used the n-gram features – specifically, unigram, bigram, and trigram.
Naturally, selecting a list of hyperparameter values based on a researchers’ experience to find a set in which the best performance is exhibited is better than finding it from a probability distribution. Realistically, however, it is impossible to calculate using the parameters in all combinations. The present research indicates that the random search method has a better performance compared with the grid search method while requiring shorter computation time and a reduced cost.
In this paper, the issues affecting the performance of HUCRF, a nonlinear model with performance that varies depending on the hyperparameters, but performs better than CRF, has been examined.
This research was supported by Basic Science Research Program through the National Research Foundation of Korea (NRF) funded by the Ministry of Science, ICT and future Planning (2015R1A2A2A01007333), and by Hallym University Research Fund, 2015 (HRF-201512-013).
Yang, E.-S., Kim, J.D., Park, C.-Y., Song, H.-J. and Kim, Y.-S. (2017), "Hyperparameter tuning for hidden unit conditional random fields", Engineering Computations, Vol. 34 No. 6, pp. 2054-2062. https://doi.org/10.1108/EC-11-2015-0350Download as .RIS
Emerald Publishing Limited
Copyright © 2017, Emerald Publishing Limited