Meta-learning Recommendation of Default Hyper-parameter Values for SVMs in Classification Tasks

Machine learning algorithms have been investigated in several scenarios, one of them is the data classification. The predictive performance of the models induced by these algorithms is usually strongly affected by the values used for their hyper-parameters. Different approaches to define these values have been proposed, like the use of default values and optimization techniques. Although default values can result in models with good predictive performance, different implementations of the same machine learning algorithms use different default values, leading to models with clearly different predictive performance for the same dataset. Optimization techniques have been used to search for hyper-parameter values able to maximize the predictive performance of induced models for a given dataset, but with the drawback of a high computational cost. A compromise is to use an optimization technique to search for values that are suitable for a wide spectrum of datasets. This paper investigates the use of meta-learning to recommend default values for the induction of Support Vector Machine models for a new classification dataset. We compare the default values suggested by the Weka and LibSVM tools with default values optimized by meta-heuristics on a large range of datasets. This study covers only classification task, but we believe that similar ideas could be used in other related tasks. According to the experimental results, meta-models can accurately predict whether tool suggested or optimized default values should be used.