DIOPT: Extremely Fast Classification Using Lookups and Optimal Feature Discretization

摘要: For low dimensional classification problems we propose the novel DIOPT approach which considers construction of a discretized feature space. Predictions for all cells in this space are obtained by means reference classifier and class labels stored lookup table generated enumerating complete This then leads to extremely high throughput as inference consists only discretizing relevant features reading label from index corresponding concatenation bin indices. Since size is limited due memory constraints, selection optimal their respective discretization levels paramount. We particular supervised striving achieve maximal separation features, further employ purpose-built memetic algorithm search towards levels. The run time accuracy compared benchmark random forest decision tree classifiers several publicly available data sets. Orders magnitude improvements recorded runtime with insignificant or modest degradation many evaluated binary tasks.