A Collection of Outlier Ensemble Algorithms.

outlierensembles

outlierensembles provides a collection of outlier/anomaly detection ensembles. Given the anomaly scores of different anomaly detection methods, the following ensemble techniques can be used to construct an ensemble score:

1. Item Response Theory based ensemble discussed in Kandanaarachchi (2021)
2. Greedy ensemble discussed in Schubert et al. (2012)
3. Inverse Cluster Weighted Averaging (ICWA) method discussed in Chiang (2017)
4. Using Maximum scores discussed in Aggarwal and Sathe (2015)
5. Using a threshold sum discussed in Aggarwal and Sathe (2015)
6. Using the mean as the ensemble score

Installation

You can install the released version of outlierensembles from CRAN with:

install.packages("outlierensembles")

And the development version from GitHub with:

# install.packages("devtools")
devtools::install_github("sevvandi/outlierensembles")

Example

We use 7 anomaly detection methods from DDoutlier R package as our base methods. You can use any anomaly detection method you want to build the ensemble. First, we construct the IRT ensemble. The colors show the ensemble scores.

faithfulu <- scale(faithful)

y1 <- DDoutlier::KNN_AGG(faithfulu)
y2 <- DDoutlier::LOF(faithfulu)
y3 <- DDoutlier::COF(faithfulu)
y4 <- DDoutlier::INFLO(faithfulu)
y5 <- DDoutlier::KDEOS(faithfulu)
y6 <- DDoutlier::LDF(faithfulu)
y7 <- DDoutlier::LDOF(faithfulu)
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6, y7)
ens1 <- irt_ensemble(Y)
#> Warning in sqrt(diag(solve(Hess))): NaNs produced
df <- cbind.data.frame(faithful, ens1$scores)
colnames(df)[3] <- "IRT"
ggplot(df, aes(eruptions, waiting)) + geom_point(aes(color=IRT))  +  scale_color_gradient(low="yellow", high="red") 

Then we do the greedy ensemble.

ens2 <- greedy_ensemble(Y)
df <- cbind.data.frame(faithful, ens2$scores)
colnames(df)[3] <- "Greedy"
ggplot(df, aes(eruptions, waiting)) + geom_point(aes(color=Greedy)) +  scale_color_gradient(low="yellow", high="red") 

We do the ICWA ensemble next.

ens3 <- icwa_ensemble(Y)
df <- cbind.data.frame(faithful, ens3)
colnames(df)[3] <- "ICWA"
ggplot(df, aes(eruptions, waiting)) + geom_point(aes(color=ICWA)) +  scale_color_gradient(low="yellow", high="red") 

Next, we use the maximum scores to build the ensemble.

ens4 <- max_ensemble(Y)
df <- cbind.data.frame(faithful, ens4)
colnames(df)[3] <- "Max"
ggplot(df, aes(eruptions, waiting)) + geom_point(aes(color=Max)) +  scale_color_gradient(low="yellow", high="red") 

Then, we use the a threshold sum to construct the ensemble.

ens5 <- threshold_ensemble(Y)
df <- cbind.data.frame(faithful, ens5)
colnames(df)[3] <- "Threshold"
ggplot(df, aes(eruptions, waiting)) + geom_point(aes(color=Threshold)) +  scale_color_gradient(low="yellow", high="red") 

Finally, we use the mean values as the ensemble score.

ens6 <- average_ensemble(Y)
df <- cbind.data.frame(faithful, ens6)
colnames(df)[3] <- "Average"
ggplot(df, aes(eruptions, waiting)) + geom_point(aes(color=Average)) +  scale_color_gradient(low="yellow", high="red") 

References