A Generalized Framework for Multi-Criteria Classifiers with Automated Learning: application on FLIR Ship Imagery

Supervised classification often consists in assigning a set of entities (e.g. alternatives, images, projects, subjects) into pre-defined and homogeneous categories. Categories are known a priori either by defining profiles limit between them or by a set of typical profiles (reference prototypes or elements) for each category. Ordinal Classification (or Sorting) usually refers to an order relationship between the categories, and nominal classification otherwise. Recently, a variety of classification methods–based on Artificial Intelligence (AI) and Operations Research (OR) techniques–have been proposed to solve classification problems [41]. Neural Networks (NN), Machine Learning (ML), Rough Sets (RS), Fuzzy Sets (FS) and Multi-Criteria Decision Analysis (MCDA) were used for the development and the validation of these methods. This paper focuses on classification methods based on MCDA methodology. In this paper we use Multi-Criteria Classifiers (MCCs) to designate supervised classification methods based on MCDA methodology. The most MCCs are based on either outranking or multi-attribute utility approaches. Roy and Moscarola [35], Masaglia and Ostanello [24], Yu [42], Perny [31], Belacel [3] and Henriet [15] have proposed MCCs based on the outranking approach, while M.H.DIS (Multi-group Hierarchical DIScrimination) method [40] and UTADIS (UTilités Additives DIScriminantes) method and its variants ([21], [39], [10]) are typical methods based on multi-attribute utility theory. This paper focuses essentially on outranking-based nominal MCCs where there is no order relationship between the categories. These MCCs are based on concordance/discordance concepts. Limitation of outranking-based methods is due to the large number of parameters (e.g. discrimination thresholds, weights, reference alternatives, etc.) required. In MCDA context, these parameters are generally elicited using interactive approaches from the decision-maker to articulate his relational preference system: it’s the Direct Elicitation Approach (DEA). However, it is difficult for the decision-maker to provide such information in a coherent way when the number of these parameters is considerable. Indirect Elicitation Approach (IEA) or Automatic Learning Methods (ALMs) might be the solution to elicit automatically the values of these parameters based on a training set of pre-assigned examples. These two elicitation approaches will be discussed in Section 3. This paper makes three main contributions. First, we propose a generalized framework for Nominal Concordance/Discordance-based MCC (NCD-based MCCs). Second, we develop an ALM based on RealCoded Genetic Algorithm (RCGA) to estimate the parameters of NCD-based MCCs. Then we illustrate and assess the performance of the proposed approach on selected NCD-based MCCs. Even if the purpose of the comparison might be seen limited, we present exper-