Methods Of Classification And Sort Evaluation Using Mahalanobis-Taguchi System Based On Omni-Optimizer Algorithm And Applications

Data mining is an effective means of accessing knowledge and information from the data for further decision-making. The classification problem is one of the main issues in data mining because it aims to extract a classifier which can be used to predict the classes of objects whose class labels are unknown based on the data characteristics. Currently, many mature and effective classification methods with distinct characteristics have been thoroughly studied and widely used, but still have their own limitations and shortcomings. Feature selection, which aims to reduce the number of features (dimensions), can play a role of improving the classification accuracy, simplifying the problem and cost-saving in process of classification.The Mahalanobis-Taguchi System (MTS) is a new classifying and diagnostic tech-nique of pattern recognition using a collection of methods of Mahalanobis distance (MD), orthogonal arrays (OAs) and signal-to-noise ratios (SNR). Advantages of MTS, such as can determine the important features, do not need to make assumptions about data distribution and high classification speed, etc., make it a wide range of applications in areas such as industrial production, business management and pattern recognition. As a relatively new classification method, it also has some shortcomings and rigorous problems about theoret-ical basis and usage, such as:lack of rigor for feature screening, subjectivity of determin-ing the threshold and confining to the two-class classification. In addition to be used for classification and diagnosis, the characteristics of MTS makes it also be used to sort evalu-ation. For shortages of the MTS, the paper aims to make the MTS an effective and practic-al classification and sort evaluation technique with the focusing on improving the theory of MTS, and apply it to real problems. These research works of the paper are the following:(1) Research on MTS based on omni-optimizer algorithm for two-class classificationFor the inadequacy of traditional MTS in feature selection by OAs and SNR as well as determining the threshold by loss function, multi-objective optimization and om-ni-optimizer algorithm are alternative for improvements. Classification accuracy and the-larger-the-better SNR as well as dimensionality reduction are considered as optimiza-tion goals to build a multi-objective optimization model, which is used for feature selection and threshold determining, is solved by omni-optimizer algorithm. For evaluating the ef-fectiveness of the proposed method, a dataset experiment for comparison is implemented. Finally, the proposed method is applied to a real case about product quality inspection. The results show that the proposed method outperforms other well-known algorithms not only on classification accuracy but also on feature selection efficiency.(2) Research on MTS based on omni-optimizer algorithm and probabilistic thre-sholding model for imbalanced data classificationMTS establishes a classifier by constructing a continuous measurement scale rather than directly learning from the training set. Therefore, it is expected that the construction of an MTS model will not be influenced by data distribution, and this property is helpful to overcome the class imbalance problem. A probabilistic thresholding method is proposed to determine the classification threshold in MTS for imblanced data classification. Perfor-mance evaluation indicators of imbalanced data classification--g/F values and the-larger-the-better SNR as well as dimensionality reduction are considered as optimiza-tion goals to build a multi-objective optimization model, which is used for feature selection, is solved by omni-optimizer algorithm. For evaluating the effectiveness of the proposed method, a dataset experiment for comparison is implemented. The results show that the proposed method is effective not only on classification accuracy for imbalanced data but also on feature selection efficiency.(3) Research on multi-MTS for multi-class classificationMTS is a method for two-class classification, which can not be directly used for mul-ti-class classification. Two multi-class MTS classification methods----binary tree MTS (BT-MTS) and multi-Mahalanobis-space(MS) feature-selection MTS (MF-MTS) are stu-died. BT-MTS decomposes multi-class problem by a combination of binary tree and MTS. The process and steps of implementation of BT-MTS and the construction programs for binary tree are studied. MF-MTS establishes classifier using distance criterion by estab-lishing an individual MS space for each class and can optimize the feature space at the same time. The process and steps of implementation of MF-MTS are studied. In MF-MTS, classification accuracy and the-larger-the-better SNR as well as dimensionality reduction are considered as optimization goals to build a multi-objective optimization model, which is used for feature selection, is solved by omni-optimizer algorithm. For evaluating the ef-fectiveness of the proposed two methods, a dataset experiment for comparison is imple-mented. Finally, the proposed methods are applied to a real case about credit rating of en-terprises on platform of government investment and financing. According to the results, we can conclude that the two multi-class MTS classification methods have high classification accuracy, but the MF-MTS is more valuable for application.(4) Research on MTS based on omni-optimizer algorithm for sort evaluation Currently, MTS is mainly used for classification problems. In fact, the MTS can sort the samples for evaluation by calculating the Mahalanobis distances of samples to refer-ence space (RS), which represents the degree of deviation from the RS. In this paper, MTS for sort evaluation is studied including:procedures of MTS for sort evaluation; variable screening model based on the omni-optimizer algorithm; validity and discussion of the MTS for sort evaluation by examples and comparative analysis. The results show that the MTS for sort evaluation method does not need to determine the index weight, can maintain consistent evaluation reference, and can screen index, which is an effective evaluation me-thod. But the mechanisms for identifying the reference space require further studies and perfect.Based on the above research work, the main contributions and innovations of this pa-per are as follows:(1) With identifying different goals for classification and sort evaluation, optimization ideas are imported in the core issue of MTS feature selection and multi-objective optimiza-tion model as an alternative to OAs is proposed and studied innovatively, which is solved by omni-optimizer algorithms. It is a new method for feature selection.(2) According to the different purposes of classification. optimization or probabilistic model is used to determine threshold of MTS classification, which replaces the loss of function or exhaustive search method. It is a new threshold determining way.(3) The MTS class classification method has been successfully extended to imba-lanced data classification and multi-class classification by means of probabilistic thresholding model, binary tree and multi-MS and the effectiveness of the methods are proved. The me-thods are new for imbalanced data classification and multi-class classification.