Christel Ruwet

Multivariate coefficients of variation

In the univariate setting, coefficients of variation are well-known and used to compare the variability of populations characterized by variables expressed in different units or having really different means. When dealing with more than one variable, the use of such a relative dispersion measure is much less common even though several generalizations of the coefficient of variation to the multivariate setting have been introduced in the literature. In this paper, the lack of robustness of the sample versions of the multivariate coefficients of variation (MCV) is illustrated by means of influence functions and some robust counterparts based either on the Minimum Covariance Determinant (MCD) estimator or on the S estimator are advocated. Then, focusing on two of the considered MCVs, a diagnostic tool is derived and its efficiency in detecting observations having an unduly large effect on variability is illustrated on a real-life data set. The influence functions are also used to compute asymptotic variances under elliptical distributions, yielding approximate confidence intervals. Finally, simulations are conducted in order to compare, in a finite sample setting, the performance of the classical and robust MCVs in terms of variability and in terms of coverage probability of the corresponding asymptotic confidence intervals.

Impact of Contamination on Training and Test Error Rates in Statistical Clustering

The k-means algorithm is one of the most common non hierarchical methods of clustering. It aims to construct clusters in order to minimize the within cluster sum of squared distances. However, as most estimators defined in terms of objective functions depending on global sums of squares, the k-means procedure is not robust with respect to atypical observations in the data. Alternative techniques have thus been introduced in the literature, e.g., the k-medoids method. The k-means and k-medoids methodologies are particular cases of the generalized k-means procedure. In this article, focus is on the error rate these clustering procedures achieve when one expects the data to be distributed according to a mixture distribution. Two different definitions of the error rate are under consideration, depending on the data at hand. It is shown that contamination may make one of these two error rates decrease even under optimal models. The consequence of this will be emphasized with the comparison of influence functions and breakdown points of these error rates.

The influence function of the TCLUST robust clustering procedure

The TCLUST procedure performs robust clustering with the aim of finding clusters with different scatter structures and weights. An Eigenvalues Ratio constraint is considered by TCLUST in order to achieve a wide range of clustering alternatives depending on the allowed differences among cluster scatter matrices. Moreover, this constraint avoids finding uninteresting spurious clusters. In order to guarantee the robustness of the method against the presence of outliers and background noise, the method allows for trimming of a given proportion of observations self-determined by the data. Based on this “impartial trimming”, the procedure is assumed to have good robustness properties. As it was done for the trimmed k-means method, this article studies robustness properties of the TCLUST procedure in the univariate case with two clusters by means of the influence function. The conclusion is that the TCLUST has a robustness behavior close to that of the trimmed k-means in spite of the fact that it addresses a more general clustering approach.