This article presents a decision-directed approach for classifying discrete data. In the clustering algorithm, probable clusters are initiated through the use of a sorting scheme based on the estimated probability distribution of the data and an arbitrary distance measure. The subsequent iterative reclassification procedures are directed by the estimated distribution of each class. The distribution estimation adopted is modified from the dependence tree procedure. The algorithm performance is then evaluated through the use of simulated and clinical data. Finally, the algorithm is applied to disease categorization and to signs and symptoms extraction for each disease class.

A method of data analysis which aims at the detection of the significant periodicities as well as of the irregular high frequency oscillations (perturbations) in biological time series (TS) is proposed. The significant frequency modes are detected by application of a statistical significance criterion on the periodogram derived from the given TS; they are used to construct the significant periodic series. The latter is then subtracted from the original TS in order to obtain the detrended TS defined as the perturbation series which is considered as the outcome of a stochastic process. Under the assumption of Gaussian characteristics for the stochastic process, evidence of which is available, a probabilistic basis for both univariate and bivariate analysis is provided. A ``surprisal'' measure defined as the reciprocal of the probability for an outcome of a stochastic process (or a pair of correlated porcesses) is introduced to account for both the average as well as the localized perturbation of the TS (or the correlation between the pair of TS). A dependence tree configuration that maximizes the overall mutual information or dependence among branches is proposed as an optimal representation of a multivariate system. Lag-correlation analysis and cospectrum tests are adopted for validation and modification of the configuration generated. Simulated systems are employed to test the extent of linear approximation of various nonlinear functions as well as the sensitivity of the proposed method when the simulated system is subject to disturbances of varied type and degree.

In this paper, issues concerning feature patterns in terms of both feature composition and feature interdependence are discussed, and the concepts of typicality and diversity of an ensemble are formulated. The features of the specimens investigated are organized in a two-dimensional array, called an observation matrix, with each row vector representing the ordered set of features of a specimen. An algorithm (based upon the proposed measures and statistical screening) is implemented for extracting feature patterns. In the algorithm, schemes for feature patterns and specimen reweighting are proposed to optimize the utilization of available information in the array, and to minimize possible bias caused by the uneven sampling of the ensemble. Two sets of real world data in the environmental and molecular biology areas are used to exemplify the physical meaning of the proposed measures as well as to demonstrate the operational feasibility and significance of this methodology in analyzing homologous ensemble which is subject to variable degrees of diversity.