In dannymorris/ALSO: Attribute-Wise Learning for Scoring Outliers (ALSO) using Random Forests

knitr::opts_chunk$set(echo = TRUE, warning = FALSE, message = FALSE, 
                      cache = FALSE)

Overview

Attirubte-wise Learning for Scoring Outliers (ALSO) is an outlier detection technique for standard multidimensional datasets. For each feature in the dataset, a predictive model is constructed in which the reference feature is the target and the remaining features are predictors. When a feature is generally predictable with good accuracy, observations deviating significantly from predicted values may be outliers. Observations with significant deviations across many predictable features may be strong outliers.

This document will illustrate the use of the ALSO package. It was created for the sole purpose of making the ALSO technique easily available to R users. The package currently contains a single function, ALSO_RF().

library(ALSO) # devtools::install_github("dannymorris/ALSO")

library(dplyr)
library(corrplot)
library(plotly)

sessionInfo()

Data

states <- state.x77 %>% as_tibble()

states

This dataset is relatively small at 50 rows and 8 variables. Later on we'll explore a higher dimenional dataset and observe the results.

Center and scale variables for preprocessing

It is customary to standardize numeric variables prior to predictive modeling in order to eliminate the effects of differences in original measurement scales. Let's apply the common z-standardization technique.

states_scaled <- states %>%
    mutate_all(funs(scale))

Univariate density plots

par(mfrow = c(3,3),
    mar = c(3,3,2,2))

for (i in seq_along(states_scaled)) {
    pull(states_scaled[, i]) %>%
        density() %>%
        plot(main = colnames(states_scaled[, i]),
             xlab = "")
}

Population, Income, and Area show potential outliers in the right-side tails.

Bivariate linear correlations

pairs(states_scaled)

cor(states_scaled) %>% round(., 2)

The correlation plot and matrix show quite a few moderate to strong correlations in both positive and negative directions.

ALSO with random forest regressor

A random forest works quite nicely in the context of ALSO for a few reasons:

1. can do prediction or classification
2. robust due to ensembling
3. can detect nonlinear relationships

rf_also <- ALSO_RF(data = states_scaled, cross_validate = FALSE, 
                   scores_only = FALSE)

rf_also_cv <- ALSO_RF(data = states_scaled, cross_validate = TRUE, 
                      scores_only = FALSE)

Note that cross validation scoring (ensures that each point receives an out of sample score) is significantly more computationally intensive than the in-sample scoring.

microbenchmark::microbenchmark(
    ALSO_RF(data = states_scaled, cross_validate = FALSE),
    ALSO_RF(data = states_scaled, cross_validate = TRUE)
)

# Unit: milliseconds
#                                                                    expr
#  ALSO_RF(data = states_scaled, cross_validate = FALSE, scores_only = F)
#   ALSO_RF(data = states_scaled, cross_validate = TRUE, scores_only = F)
#        min        lq      mean    median        uq       max neval
#   421.8709  451.9757  482.1679  466.6313  481.8332  828.7061   100
#  2013.3989 2081.6085 2178.1072 2121.9938 2202.0590 2789.7553   100

Cross validation scoring takes rougly 4.5 times longer.

Outlier scores

plot(density(rf_also$scores), main = "Outlier Scores Density Estimate",
     xlab = "ALSO Score")

Present of outliers reflected in the severe right skewness of the distribution of outlier scores.

Prediction errors

Squared prediction errors for all n points across all k feature models.

rf_also$squared_prediction_errors

The squared prediction error matrix is multiplied by the feature weights to produce the total outlier score.

Feature weights

rf_also$adjusted_feature_weights

Visualization of outlier scores via principal components analysis

Principal components analysis (PCA) is an effective technique for visualizing multidimensional data in fewer dimensions.

pca_states <- princomp(states_scaled)

summary(pca_states)

It appears that the top 3 principal components explain nearly 80% of the variation in the original dataset. Let's inspect a 3-D scatterplot for potential outliers.

pca_states$scores %>%
    as_tibble() %>%
    mutate(outlier_scores = rf_also$scores) %>%
    plotly::plot_ly(x = ~Comp.1, y = ~Comp.2, z = ~Comp.3, 
                    color = ~outlier_scores, type = "scatter3d")

A comment on efficiency

Currently the ALSO_RF() function is not efficient for medium to large data sets. This is especially true for wider data sets with many features. The function was recently tested on a 30,000 x 30 dataset and required approximately 30 seconds to compute on a Delll i9 with 8GB of RAM without cross validation scoring.

dannymorris/ALSO documentation built on May 4, 2019, 7:42 p.m.