A comparison of two methods of discriminant analysis applied to binary data

Area of Science:

• Statistics
• Machine Learning
• Data Mining

Background:

• Classical linear discriminant analysis (LDA) is a parametric method typically suited for continuous variables and assumes normal distributions with identical covariance matrices.
• Kernel discriminant analysis (KDA) is a nonparametric method, particularly effective for binary data.
• Evaluating classification performance requires careful consideration of error rate estimation methods.

Purpose of the Study:

• To compare the performance of classical linear discriminant analysis and kernel discriminant analysis on multivariate binary data.
• To assess the accuracy of apparent error rates versus estimated true error rates for both methods.

Main Methods:

• Application of classical linear discriminant analysis to multivariate binary datasets.
• Application of kernel discriminant analysis to the same multivariate binary datasets.
• Estimation of true error rates using independent test sets and the leaving-one-out method.

Main Results:

• Kernel discriminant analysis consistently yielded lower apparent error rates than classical linear discriminant analysis.
• No significant difference in true error rates was observed between KDA and LDA when estimated via test sets or leaving-one-out validation.
• Apparent error rates may overestimate classification performance, especially for binary data.

Conclusions:

• While kernel discriminant analysis appears superior based on apparent error rates for binary data, its true predictive performance is comparable to classical linear discriminant analysis.
• The choice of error rate estimation method is crucial for accurately comparing classification algorithms.
• Both LDA and KDA can be applied to binary data, but their performance differences are not significant when true error rates are considered.