Bagged OutlierTrees

Bagged OutlierTrees is an explainable unsupervised outlier detection method based on an ensemble implementation of the existing OutlierTree procedure (Cortes, 2020). This implementation takes advantage of bootstrap aggregating (bagging) to improve robustness by reducing the possible masking effect and subsequent high variance (similarly to Isolation Forest), hence the name “Bagged OutlierTrees”.

To learn more about the base procedure OutlierTree (Cortes, 2020), please refer to <arXiv:2001.00636> (the corresponding GitHub repository can be found here). This repository and its documentation are heavily based on the latter to ensure consistency and ease-of-use between the packages.

Important Note

Bagged OutlierTrees does not support missing values: if not taken care of, rows with missing values are automatically removed while fitting the model or predicting on new data.

Installation

You can install the released version of bagged.outliertrees from CRAN with:

install.packages("bagged.outliertrees")

Or, install the development version of bagged.outliertrees from GitHub with:

# install.packages("devtools")
devtools::install_github("RafaJPSantos/bagged.outliertrees")

Example

This is a basic example which shows you how to find outliers in the hypothyroid dataset:

library(bagged.outliertrees)

### example dataset with interesting outliers
data(hypothyroid)

### fit a Bagged OutlierTrees model
model <- bagged.outliertrees(hypothyroid,
  ntrees = 100,
  subsampling_rate = 0.75,
  z_outlier = 5,
  nthreads = 1
)

### use the fitted model to find outliers in the training dataset
outliers <- predict(model,
  newdata = hypothyroid,
  min_outlier_score = 0.5,
  nthreads = 1
)

### print the top-5 outliers in human-readable format
print(outliers, outliers_print = 5)
#> Reporting top 5 outliers [out of 28 found]
#> 
#> row [1438] - suspicious column: [FTI] - suspicious value: [394.495412844037]
#>  distribution: 99.93% <= [292.4685] - [mean: 109.7603] - [sd: 30.5332] - [norm. obs: 956]
#> 
#> 
#> row [623] - suspicious column: [age] - suspicious value: [455]
#>  distribution: 99.92% <= [91.8] - [mean: 53.2439] - [sd: 18.9342] - [norm. obs: 956]
#> 
#> 
#> row [1157] - suspicious column: [T3] - suspicious value: [7]
#>  distribution: 99.79% <= [5.4] - [mean: 2.0285] - [sd: 0.7776] - [norm. obs: 955]
#> 
#> 
#> row [745] - suspicious column: [T4U] - suspicious value: [2.12]
#>  distribution: 99.91% <= [1.7176] - [mean: 0.9977] - [sd: 0.1511] - [norm. obs: 698]
#>      [age] > [37.5789] (value: 87)
#> 
#> 
#> row [1425] - suspicious column: [FTI] - suspicious value: [161.290322580645]
#>  distribution: 98.99% <= [103.4513] - [mean: 62.781] - [sd: 18.0421] - [norm. obs: 85]
#>      [TT4] <= [98.6733] (value: 50)

References

• outliertree
• Cortes, David. “Explainable outlier detection through decision tree conditioning.” arXiv preprint arXiv:2001.00636 (2020).
• GritBot software