@Conference{Marinov2022,
  author    = {Dobromir Marinov and Daniel Karapetyan},
  title     = {Faster Hyper-Parameter Optimisation via Partial Cross-Validation},
  year      = {2022},
  publisher = {OR64, Warwick, UK},
  abstract  = {It has been observed that machine learning (ML) systems are usually sensitive to the values of the hyperparameters (HPs), and HP optimisation has been used to set their values. The objective function in HP optimisation is the accuracy of the ML system provided a specific set of HPs. However, the measured accuracy is highly dependent on the training and test datasets; hence cross-validation is commonly used to reduce the amount of noise in the objective function.

We designed an approach to use partial cross-validation for the less successful sets of HPs. The system uses statistical analysis to identify the sets of HPs that are unlikely to achieve the highest performance. We embedded this approach into the Random HP optimisation method and tested it on a large set of small datasets for SVM. We found out that our approach can reduce the Random HP optimisation time budget by about 30%.},
  keywords  = {Hyper-parameter optimisation, cross-validation},
}