| Type: | Package |
| Title: | A Collection of Outlier Ensemble Algorithms |
| Version: | 0.1.3 |
| Maintainer: | Sevvandi Kandanaarachchi <sevvandik@gmail.com> |
| Description: | Ensemble functions for outlier/anomaly detection. There is a new ensemble method proposed using Item Response Theory. Existing outlier ensemble methods from Schubert et al (2012) <doi:10.1137/1.9781611972825.90>, Chiang et al (2017) <doi:10.1016/j.jal.2016.12.002> and Aggarwal and Sathe (2015) <doi:10.1145/2830544.2830549> are also included. |
| License: | GPL (≥ 3) |
| Encoding: | UTF-8 |
| Depends: | R (≥ 3.5) |
| Imports: | airt, EstCRM, psych, apcluster |
| RoxygenNote: | 7.3.2 |
| Suggests: | dbscan, knitr, rmarkdown, ggplot2 |
| VignetteBuilder: | knitr |
| URL: | https://sevvandi.github.io/outlierensembles/ |
| LazyData: | true |
| NeedsCompilation: | no |
| Packaged: | 2025-03-27 03:14:32 UTC; kan092 |
| Author: | Sevvandi Kandanaarachchi
 [aut, cre] |
| Repository: | CRAN |
| Date/Publication: | 2025-03-27 04:10:01 UTC |
A dataset containing anomaly scores
Description
This dataset contains anomaly scores for the annulus data
Usage
Y
Format
A dataframe of 803 rows and 9 columns
- Dimension 1
The different anomaly detection methods
- Dimension 2
Each column has the anomaly scores from that method for the annulus dataset.
Old Faithful anomaly scores
Description
This dataset contains anomaly scores for the Old Faithful dataset
Usage
Yfaithful
Format
A dataframe of 272 rows and 9 columns
- Dimension 1
The different anomaly detection methods
- Dimension 2
Each column has the anomaly scores from that method for the old faithful dataset.
Uses the mean as the ensemble score
Description
This function uses the mean as the ensemble score.
Usage
average_ensemble(X)
Arguments
X |
The input data containing the outlier scores in a dataframe, matrix or tibble format. Rows contain observations and columns contain outlier detection methods. |
Value
The ensemble scores.
Examples
set.seed(123)
if (requireNamespace("dbscan", quietly = TRUE)) {
X <- data.frame(x1 = rnorm(200), x2 = rnorm(200))
X[199, ] <- c(4, 4)
X[200, ] <- c(-3, 5)
# Using different parameters of lof for anomaly detection
y1 <- dbscan::lof(X, minPts = 10)
y2 <- dbscan::lof(X, minPts = 20)
knnobj <- dbscan::kNN(X, k = 20)
# Using different KNN distances as anomaly scores
y3 <- knnobj$dist[ ,10]
y4 <- knnobj$dist[ ,20]
# Dense points are less anomalous. Hence 1 - pointdensity is used.
y5 <- 1 - dbscan::pointdensity(X, eps = 0.8, type = "gaussian")
y6 <- 1 - dbscan::pointdensity(X, eps = 0.5, type = "gaussian")
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6)
ens <- average_ensemble(Y)
ens
}
Computes an ensemble score using the greedy algorithm proposed by Schubert et al (2012)
Description
This function computes an ensemble score using the greedy algorithm in the paper titled Evaluation of Outlier Rankings and Outlier Scores by Schubert et al (2012) <doi:10.1137/1.9781611972825.90>. The greedy ensemble is detailed in Section 4.3.
Usage
greedy_ensemble(X, kk = 5)
Arguments
X |
The input data containing the outlier scores in a dataframe, matrix or tibble format. Rows contain observations and columns contain outlier detection methods. |
kk |
The number of estimated outliers. |
Value
A list with the components:
scores |
The ensemble scores. |
methods |
The methods that are chosen for the ensemble. |
chosen |
The chosen subset of original anomaly scores. |
Examples
set.seed(123)
if (requireNamespace("dbscan", quietly = TRUE)){
X <- data.frame(x1 = rnorm(200), x2 = rnorm(200))
X[199, ] <- c(4, 4)
X[200, ] <- c(-3, 5)
# Using different parameters of lof for anomaly detection
y1 <- dbscan::lof(X, minPts = 10)
y2 <- dbscan::lof(X, minPts = 20)
knnobj <- dbscan::kNN(X, k = 20)
# Using different KNN distances as anomaly scores
y3 <- knnobj$dist[ ,10]
y4 <- knnobj$dist[ ,20]
# Dense points are less anomalous. Hence 1 - pointdensity is used.
y5 <- 1 - dbscan::pointdensity(X, eps = 0.8, type = "gaussian")
y6 <- 1 - dbscan::pointdensity(X, eps = 0.5, type = "gaussian")
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6)
ens <- greedy_ensemble(Y, kk=5)
ens$scores
}
Computes an ensemble score using inverse cluster weighted averaging method by Chiang et al (2017)
Description
This function computes an ensemble score using inverse cluster weighted averaging in the paper titled A Study on Anomaly Detection Ensembles by Chiang et al (2017) <doi:10.1016/j.jal.2016.12.002>. The ensemble is detailed in Algorithm 2.
Usage
icwa_ensemble(X)
Arguments
X |
The input data containing the outlier scores in a dataframe, matrix or tibble format. Rows contain observations and columns contain outlier detection methods. |
Value
The ensemble scores.
Examples
set.seed(123)
if (requireNamespace("dbscan", quietly = TRUE)) {
X <- data.frame(x1 = rnorm(200), x2 = rnorm(200))
X[199, ] <- c(4, 4)
X[200, ] <- c(-3, 5)
# Using different parameters of lof for anomaly detection
y1 <- dbscan::lof(X, minPts = 10)
y2 <- dbscan::lof(X, minPts = 20)
knnobj <- dbscan::kNN(X, k = 20)
# Using different KNN distances as anomaly scores
y3 <- knnobj$dist[ ,10]
y4 <- knnobj$dist[ ,20]
# Dense points are less anomalous. Hence 1 - pointdensity is used.
y5 <- 1 - dbscan::pointdensity(X, eps = 0.8, type = "gaussian")
y6 <- 1 - dbscan::pointdensity(X, eps = 0.5, type = "gaussian")
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6)
ens <- icwa_ensemble(Y)
ens
}
Computes an ensemble score using Item Response Theory
Description
This function computes an ensemble score using Item Response Theory (IRT). This was proposed as an ensemble method for anomaly/outlier detection in Kandanaarachchi (2021) <doi:10.13140/RG.2.2.18355.96801>.
Usage
irt_ensemble(X)
Arguments
X |
The input data containing the outlier scores in a dataframe, matrix or tibble format. Rows contain observations and columns contain outlier detection methods. |
Details
For outlier detection, higher ensemble scores indicate higher levels of anomalousness. This ensemble uses IRT's latent trait to uncover the hidden ground truth, which is used as the ensemble score. It uses the R packages airt and EstCRM to fit the IRT models. It can also be used for other ensembling tasks.
Value
A list with the components:
scores |
The ensemble scores. |
model |
The IRT model. |
Examples
set.seed(123)
if (requireNamespace("dbscan", quietly = TRUE)) {
X <- data.frame(x1 = rnorm(200), x2 = rnorm(200))
X[199, ] <- c(4, 4)
X[200, ] <- c(-3, 5)
# Using different parameters of lof for anomaly detection
y1 <- dbscan::lof(X, minPts = 10)
y2 <- dbscan::lof(X, minPts = 20)
knnobj <- dbscan::kNN(X, k = 20)
# Using different KNN distances as anomaly scores
y3 <- knnobj$dist[ ,10]
y4 <- knnobj$dist[ ,20]
# Dense points are less anomalous. Hence 1 - pointdensity is used.
y5 <- 1 - dbscan::pointdensity(X, eps = 0.8, type = "gaussian")
y6 <- 1 - dbscan::pointdensity(X, eps = 0.5, type = "gaussian")
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6)
ens <- irt_ensemble(Y)
ens$scores
}
Computes an ensemble score using the maximum score of each observation
Description
This function computes an ensemble score using the maximum score for each observation as detailed in Aggarwal and Sathe (2015) <doi:10.1145/2830544.2830549>.
Usage
max_ensemble(X)
Arguments
X |
The input data containing the outlier scores in a dataframe, matrix or tibble format. Rows contain observations and columns contain outlier detection methods. |
Value
The ensemble scores.
Examples
set.seed(123)
if (requireNamespace("dbscan", quietly = TRUE)) {
X <- data.frame(x1 = rnorm(200), x2 = rnorm(200))
X[199, ] <- c(4, 4)
X[200, ] <- c(-3, 5)
# Using different parameters of lof for anomaly detection
y1 <- dbscan::lof(X, minPts = 10)
y2 <- dbscan::lof(X, minPts = 20)
knnobj <- dbscan::kNN(X, k = 20)
# Using different KNN distances as anomaly scores
y3 <- knnobj$dist[ ,10]
y4 <- knnobj$dist[ ,20]
# Dense points are less anomalous. Hence 1 - pointdensity is used.
y5 <- 1 - dbscan::pointdensity(X, eps = 0.8, type = "gaussian")
y6 <- 1 - dbscan::pointdensity(X, eps = 0.5, type = "gaussian")
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6)
ens <- max_ensemble(Y)
ens
}
Computes an ensemble score by aggregating values above the mean
Description
This function computes an ensemble score by aggregating values above the mean as detailed in Aggarwal and Sathe (2015) <doi:10.1145/2830544.2830549>.
Usage
threshold_ensemble(X)
Arguments
X |
The input data containing the outlier scores in a dataframe, matrix or tibble format. Rows contain observations and columns contain outlier detection methods. |
Value
The ensemble scores.
Examples
set.seed(123)
if (requireNamespace("dbscan", quietly = TRUE)) {
X <- data.frame(x1 = rnorm(200), x2 = rnorm(200))
X[199, ] <- c(4, 4)
X[200, ] <- c(-3, 5)
# Using different parameters of lof for anomaly detection
y1 <- dbscan::lof(X, minPts = 10)
y2 <- dbscan::lof(X, minPts = 20)
knnobj <- dbscan::kNN(X, k = 20)
# Using different KNN distances as anomaly scores
y3 <- knnobj$dist[ ,10]
y4 <- knnobj$dist[ ,20]
# Dense points are less anomalous. Hence 1 - pointdensity is used.
y5 <- 1 - dbscan::pointdensity(X, eps = 0.8, type = "gaussian")
y6 <- 1 - dbscan::pointdensity(X, eps = 0.5, type = "gaussian")
Y <- cbind.data.frame(y1, y2, y3, y4, y5, y6)
ens <- threshold_ensemble(Y)
ens
}