LOCI: Local Correlation Integral (LOCI) algorithm with constant...

In DDoutlier: Distance & Density-Based Outlier Detection

Description

Function to calculate Local Correlation Integral (LOCI) as an outlier score for observations. Suggested by Papadimitriou, S., Gibbons, P. B., & Faloutsos, C. (2003). Uses a k number of nearest neighbors instead of a constant radius

Usage

LOCI(dataset, alpha = 0.5, nn = 20, k = 3)

Arguments

dataset	The dataset for which observations have a LOCI returned
alpha	The parameter setting the size of the sampling neighborhood, as a proportion of the counting neighborhood, for observations to identify other observations in their respective neighborhood. An alpha of 1 equals a sampling neighborhood the size of the counting neighborhood (the size of distance to nn). An alpha of 0.5 equals a sampling neighborhood half the size of the counting neighborhood
nn	The number of nearest neighbors to compare sampling neighborhood with. Original paper suggest a constant user-given radius that includes at least 20 neighbors in order to introduce statistical errors in MDEF. Default is 20
k	The number of standard deviations the sampling neighborhood of an observation should differ from the sampling neighborhood of neighboring observations, to be an outlier. Default is set to 3 as used in original papers experiments

Details

LOCI computes a counting neighborhood to the nn nearest observations, where the radius is equal to the outermost observation. Within the counting neighborhood each observation has a sampling neighborhood of which the size is determined by the alpha input parameter. LOCI returns an outlier score based on the standard deviation of the sampling neighborhood, called the multi-granularity deviation factor (MDEF). The LOCI function is useful for outlier detection in clustering and other multidimensional domains

Value

npar_pi	A vector of the number of observations within the sample neighborhood for observations
avg_npar	A vector of average number of observations within the sample neighborhood for neighboring observations
sd_npar	A vector of standard deviations for observations sample neighborhood
MDEF	A vector of the multi-granularity deviation factor (MDEF) for observations. The greater the MDEF, the greater the outlierness
norm_MDEF	A vector of normalized MDEF-values, being sd_npar/avg_npar
class	Classification of observations as inliers/outliers following the rule of k

Author(s)

Jacob H. Madsen

References

Papadimitriou, S., Gibbons, P. B., & Faloutsos, C. (2003). LOCI: Fast Outlier Detection Using the Local Correlation Integral. In International Conference on Data Engineering. pp. 315-326. DOI: 10.1109/ICDE.2003.1260802

Examples

# Create dataset
X <- iris[,1:4]

# Classify observations
cls_observations <- LOCI(dataset=X, alpha=0.5, nn=20, k=1)$class

# Remove outliers from dataset
X <- X[cls_observations=='Inlier',]

Example output

DDoutlier documentation built on May 1, 2019, 10:20 p.m.