This thesis explores the possibility of feature-driven time series pattern recognition from both practical and theoretical perspectives for predictive modelling in a situation where data are imbalanced, minority class examples are scarce, the ratio of feature dimension to sample size is high, and the class labels provided might not be optimized for the application. These problems are common in learning patient-specific patterns in medical and health domains, where labels provided by medical experts might not fit the goal of predictive modelling. Extracting informative labels for supervised learning is a difficult and time-consuming task. A novel strategy is proposed to solve the problems mentioned above, which aims to reduce human effort by automatically finding the earliest pattern that a classifier can recognize. The proposed algorithm locates and learns similar patterns across training examples that maximize the difference between both classes. This method ensures precise learning and boosts the performance of classifier by reducing the number of false positives. The performance of the algorithm was evaluated based on the classification results and the anticipation responses on the data provided by EPILEPSIAE, a European Epilepsy Database. An average false positive of 0.0519 per hour was achieved using the proposed algorithm with a sensitivity of 0.79 in anticipating seizures.
展开▼
