Evidential framework for Error Correcting Output Code classification. (August 2018)

• Record Type: Journal Article
• Authors: Lachaize, Marie; Hégarat-Mascle, Sylvie Le; Aldea, Emanuel; Maitrot, Aude; Reynaud, Roger
• Is Part Of: Engineering applications of artificial intelligence. Volume 73(2018)
• Page Start: 10
• Page End: 21
• Publication Date: 2018-08
• Subjects: Classification -- Error Coding Output Codes -- Belief Function Theory -- Hyperspectral data; Engineering -- Data processing -- Periodicals; Artificial intelligence -- Periodicals; Expert systems (Computer science) -- Periodicals; Ingénierie -- Informatique -- Périodiques; Intelligence artificielle -- Périodiques; Systèmes experts (Informatique) -- Périodiques; Artificial intelligence; Engineering -- Data processing; Expert systems (Computer science); Periodicals; 620.00285
• DOI: 10.1016/j.engappai.2018.04.019
• Languages: English
• ISSNs: 0952-1976

Abstract

Abstract: The Error Correcting Output Codes offer a proper matrix framework to model the decomposition of a multiclass classification problem into simpler subproblems. How to perform the decomposition to best fit the data while using a small number of classifiers has been a research hotspot, as well as the decoding part, which deals with the subproblem combination. In this work, we propose an evidential unified framework that handles both the coding and decoding steps. Using the Belief Function Theory, we propose an efficient modelling, where each dichotomizer in the ECOC strategy is considered as an independent information source. This framework allows us to easily model the refutation information provided by sparse dichotomizers and also to derive measures to detect tricky samples for which additional dichotomizers could be needed to ensure decisions. Our approach was tested on hyperspectral data used to classify nine different types of material. According to the results obtained, our approach allows us to achieve top performance using compact ECOC while presenting a high level of modularity. Highlights: BF framework provides elegant modeling of classifier information in ECOC approach. Classifier coding and decoding steps are handled simultaneously using BF operators. Refutation modeling of sparse dichotomizer response avoids bias in the decoding step. Imprecision or conflict measures are used as drift indicators versus training step. Statistics on class ambiguities allow for dynamic choice of additional dichotomizers.