OUTLIER DETECTION TECHNIQUE FOR HETEROGENEOUS DATA USING TRIMMED-MEAN ROBUST ESTIMATORS

Context. Fortunately, the most commonly used in parametric statistics assumptions such as normality, linearity, independence, are not always fulfilled real practice. The main reason for this is appearance of observations data samples that differ from bulk data, a result which sample becomes heterogeneous. application conditions generally accepted estimation procedures, example, mean, entails bias increasing and effectiveness decreasing estimates obtained. This, turn, raises problem finding possible solutions to processing sets include outliers, especially small samples. object study process detecting excluding anomalous objects heterogeneous sets.
 Objective. goal work develop procedure anomaly detection sets, rationale using number trimmed-mean robust estimators statistical measure location parameter distorted distribution models.
 Method. problems analysis (processing) containing sharply distinguished, suspicious considered. possibilities methods have been analyzed. A identification extraction outliers caused by measurement errors, hidden equipment defects, experimental conditions, etc. has proposed. proposed approach based on symmetric asymmetric truncation ranked set obtained initial statistics. For reasonable choice value coefficient, it use adaptive procedures. Observations fell into zone smallest lowest ordinal considered outliers.
 Results. allows, contrast traditional criteria identifying outlying observations, Smirnov (Grubbs) criterion, Dixon etc., split analyzed homogeneous component identify assuming their share total unknown.
 Conclusions. article proposes formation supposed zones set, built basis data. It complex procedures establish expected levels form region order dataset. final level dataset refined existing allow checking boundary (minimum maximum) outliers.