Assembling Learning Approach with Weighted-Voting Label Assignment
Jinghui Qin, Hongling Liu · 2009
This paper proposes an Assembling Learning Approach (ALA) for multi classification concerned with weighted-voting label assignment strategy. This weighted-voting idea is reflected in two components of ALA: a Weighted SVMs method (WSVM) that identifies regular data label and a Locally Adaptive ANN (LAANN) that addresses the rejected case. Basic SVM of WSVM is equipped with confidence coefficient to its decision capacity, and these coefficients form weighted-max-wins decision rule. LAANN is based on an informative metric derived from the most discriminant directions that are revealed by SVM decision interfaces. It also adopts a weighted voting strategy to improve performance. Three strategies facilitate computational ease and adaptation: basic classifier is created in individually desired feature space, which is achieved by self-tuning hyper parameters adaptively; training set is reduced by a tuning support vector clustering (TSVC); and working set of LAANN is pre-specified. We present experimental evidence of classification performance improved by our schema over the state of the art on real datasets.