Generalizability of Voting Classifiers

The objective of the project was to give a robust classifier for fault diagnosis in machines using acoustic data. We proposed a new voting based classifier - Majority Vote Point classifier for this task since it is more generalized than even linear kernel SVMs. I established a theoretical proof of generalizability by formulating a strict upper bound on the VC dimension of the MVP classifier, and then estimated the value by showing convergence of the growth function of the classifier, reducing search space exponentially with combinatorial mathematics inspired heuristics. I also carried out a case study on fault diagnosis using acoustic features in a reciprocating air compressor and demonstrated the higher consistency of the MVP classifier compared to SVM on both ITER and mRMR feature selection strategies.

A presentation summarizing my work can be found here: [Slides]