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 classier. 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 signicantly. 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 anity propagation. Clustering by k-Means signicantly outperformed feature- and decision-level fusions.

Ämnesord

TEKNIK OCH TEKNOLOGIER  -- Elektroteknik och elektronik -- Annan elektroteknik och elektronik (hsv//swe)

ENGINEERING AND TECHNOLOGY  -- Electrical Engineering, Electronic Engineering, Information Engineering -- Other Electrical Engineering, Electronic Engineering, Information Engineering (hsv//eng)

Nyckelord

random forest

ensemble of classiers

feature-space decomposition

clustering

k-Means

MeanNN

anity propagation

pathological voice

Publikations- och innehållstyp

ref (ämneskategori)

art (ämneskategori)