Page 63 - Flytxt
P. 63
and optimizing Machine Learning limited budget and resources making
pipelines to solve data science it challenging. To this end, Bayesian
problems. In this context, the level optimization is a promising strategy.
of automation could vary depending Bayesian optimization has advantages
upon the complexity and scale of the over other naive parameter search
problem. Basic level of automation strategies such as grid search and
aims at automatically discovering an random search, especially when time
optimal set of hyperparameters for hungry algorithms such as Support
a given Machine Learning algorithm Vector Machine (SVM) and deep
with respect to a given data set. The learning models are used in an AutoML
next level of automation focuses at framework. It intelligently searches
discovering an optimal combination the parameter space using a Gaussian
of Machine Learning algorithm and its process to determine the next best
hyper parameters which works best on parameter combination to evaluate.
a given data set. A more advanced level AutoML systems could use Bayesian
of automation is to discover an optimal Optimization in the joint space of design
end-to-end model pipeline which choices namely, data preprocessing,
includes a data preprocessing step, a feature preprocessing, algorithm and
feature preprocessing step, an algorithm hyper parameter selection to discover
selection, and hyper parameter tuning an optimal model pipeline for a given
step. However, performing all these problem. This will result in considerable
levels of automation require multiple increase in efficiency when it comes to
iteration of model’s execution under deploying packaged analytics models,
INSIGHTZ - VOLUME 03, 2018 63
