SpIS: A stochastic approximation approach to minimal subset instance selection

Instance selection provides a means to enhance the efficacy and efficiency of machine learning tools when utilised for data mining. This study proposes Stochastic Perturbation Instance Selection (SpIS), a wrapper instance selection algorithm which uses two candidate solutions to traverse the reduction and performance criteria seeking a minimal subset specific to any given machine learning model in conjunction with any corresponding performance metric. The subset selected by SpIS provide a high quality, representative subset of the full data enabling better identification of insights, whilst providing comparable predictive performance with respect to the full dataset. Across 43 diverse classification datasets SpIS was evaluated on 6 different wrappers using 5-fold cross validation with respect to reduction rate and prediction accuracy. The mean results across all datasets and wrappers show that SpIS selects 3.10% of the dataset on average with statistically equivalent performance at a 5% level of significance compared to the full training set. • SpIS is a novel training set selection algorithm designed to work in conjunction with any given supervised machine learning algorithm and corresponding performance metric. • On average, a training set selected by SpIS utilising approximately 3% of the available training data provides statistically equivalent test accuracy compared to the full training set at a 5% level of significance. • SpIS provides a statistically significant improvement in classification accuracy over other instance selection algorithms which select a comparable number of instances at a 5% level of significance.