An Agreement and Sparseness-based Learning Instance Selection and its Application to Subjective Speech Phenomena

Redundant instances in subjective speech phenomena may cause increased training time and performance degradation of a classifier like in other pattern recognition tasks. Instance selection, aiming at discarding some 'troublesome' instances and choosing the most informative ones, is a way to solve this issue. We thus propose a tailored algorithm based on human Agreement levels of labelling and class Sparseness for learning Instance Selection - ASIS for short. Extensive experiments on a standard speech emotion recognition task show the effectiveness of ASIS, indicating that by selecting only 30 % of the training set, the system performance significantly outperforms training on the whole training set without instance balancing. In terms of performance it remains comparable to the classifier trained with instance balancing, but at a fraction of the training material.