A clustering and principal component approach to exemplar based machine learning for classification identification.

摘要

Classifying detections is an important field of study in many disciplines. Typically, data can be represented in the form of a multidimensional vector defined within some hyperspace (e.g. One may have the sepal length, sepal width, petal length and petal width of an iris flower). One can view many classification problems as processing an unknown data vector in some way that produces an output which correctly categorizes it (e.g. Is the iris flower Iris Setosa, Iris Versicolour or Iris Virginica)? Exemplar based machine learning techniques tackle these problems by learning from representative training data. Several popular algorithms employing these techniques in various ways have been developed and published in the literature. This study explores and develops an innovative exemplar based machine learning methodology which combines clustering techniques with the tools of principal components analysis (PCA) to tackle this problem. Through clustering the methodology segments each classification's arbitrary multidimensional complex shape of training data in a way which can be adequately generalized using the tools of PCA. This generalization is then applied toward the development of an exemplar based machine learning algorithm capable of classifying unknown data. The methodology was applied to twenty one real world data sets obtained from the University of California at Irvine data repository and the results were compared to those of other research methods. The overall accuracy results equaled or exceeded the absolute best of any other method found by the author for twelve out of the twenty one data sets tested.; The development of a measure of confidence for each classification declared for any given unknown is discussed. Concepts are then proposed which would allow one to decrease the amount of information presented to a user based on the confidence level that the classification was made correctly. This confidence based filtering offers the potential of further increasing the overall accuracy of the algorithm. To highlight, the results suggest that the proposed methodology has a high degree of real world applicability and could be used over a wide range of application domains yielding highly competitive accuracies.