Enhancing decision-level fusion through cluster-based partitioning of feature set

Feature set decomposition through cluster-based partitioning is the subject of this study. Approach is applied for the detection of mild laryngeal disorder from acoustic parameters of human voice using random forest (RF) as a base classifier. Observations of sustained phonation (audio recordings of vowel/a/) had clinical diagnosis and severity level (from 0 to 3), but only healthy (severity 0) and mildly pathological (severity 1) cases were used. Diverse feature set (made of 26 variously sized subsets) was extracted from the voice signal. Feature- and decision-level fusions showed improvement over the best individual feature subset, but accuracy of fusion strategies did not differ significantly. To boost accuracy of decision-level fusion, unsupervised decomposition for ensemble design was proposed. Decomposition was obtained by feature-space re-partitioning through clustering. Algorithms tested: a) basic k-Means; b) non-parametric MeanNN; c) adaptive affinity propagation. Clustering by k-Means significantly outperformed feature- and decision-level fusions.