set.seed(1024)

library(fpc)
library(dplyr)
library(ggplot2)
library(DMwR2)

• Has clear ties with clustering
  • Clustering: find and group similar items
  • Anomaly Detection: find items which do not belong to any groups
• Types of outliers
  • Point outliers: a point out of the normal
  • Contextual outliers: a point out of the specific context
    • It is normal to have a heart rate of 80bpm
    • …unless you are dead.
  • Collective outliers: multiple points where only a few is ok
    • Multiple failed login attempts

• the boxplot rule

\([Q_1-1.5\times IQR, Q_3+1.5\times IQR]\)

• Grubb’s test

\(\displaystyle z=\frac{|x-\bar x|}{s_x}\)

\(\tau = t^2_{\alpha/(2N),N-2}\)

\(\displaystyle z\geq \frac{N-1}{\sqrt N} \sqrt{\frac {\tau} {N-2+\tau}}\)

case is an outlier if this inequality holds.

• implemented in package outliers as grubbs.test()

• For categorical variables there is no simple formula
• We need expert knowledge to compare the distribution of values
  • Then, we can label anomalies

• Types of detection
  • Supervised
  • Unsupervised
  • Semi-supervised

• Unsupervised
  • DBSCAN (we had covered last week)

dbscan.outliers <- function(data, ...) {
  require(fpc, quietly=TRUE)
  cl <- dbscan(data, ...)
  posOuts <- which(cl$cluster == 0)
  list(positions = posOuts,
       outliers = data[posOuts,],
       dbscanResults = cl)
  }

• Training data with manually labeled outliers is required
• Train a classification model with outliers being the target variable
• Use the model for detecting outliers in new training data

Major problem : Imbalance!

• Outliers are outliers, so they will be out numbered
• This imbalance creates problems for learning algorithms
  • If outliers are 2% in the set, labeling everything as normal has an accuracy of 98% !
  • Models usually ignore outliers: they are designed to detect regularities, not irregularities

• To fix imbalance
  • over sample outliers
  • under sample regulars
  • if supported by the ML method, use biased cost matrices

Using the data at hand, build a model which can be used to predict the value of a response variable based on the values of input variables.

• Almost all ML models are basically curve fitting algorithms
• If you fit a curve to the existing data points, you can use this curve to compute unknown/unobserved points

Mainly two types:

• Classification: nominal target variable
• Regression: numeric target variable

Ordinals may go into one of these categories.

Mostly, predictive analysis is curve fitting.

\(f(X_1, X_2, ..., X_k) \rightarrow Y\)

Overall approach:

1. First assume the shape of \(f\) (the type of model)
   • linear, logical, probabilistic, complex, ensemble
2. Based on the data, optimize \(f\)
3. Evaluate results

Why choose one model over another?

• Understandability / Readability
• Speed / Complexity
• Accuracy / Success of prediction

• Confusion matrix
  • A matrix displaying frequencies of observations for an interaction of predictions and ground truth
  • The predictions are the columns and the actual values are the rows

• a: the actual value is \(c_1\) and the prediction is \(c_1\)
• b: the actual value is \(c_1\) but the prediction is \(c_2\)
• d: the actual value is \(c_2\) but the prediction is \(c_1\)

• Error rate (aka. the 0/1 loss)

\[L_{0/1} = \frac{1}{N_{test}} \sum_{i=1}^{N_{test}}{I(\hat h(x_i) \neq y_i)}\]

where,

• \(N_{test}\) is the number of test cases.
• \(I(x)\) is an indicator function:
  • x is false \(\rightarrow I(x) = 0\)
    
  • x is true \(\rightarrow I(x) = 1\)
• \(\hat h(x_i)\) is the prediction for \(x_i\)
• \(y_i\) is the actual target value for observation i

• Accuracy

\[Acc = 1-L_{0/1}\]

\(\displaystyle Acc = \frac{a+e+i}{N_{test}}\)

• Cost/benefit matrix

• Provides flexible cost and benefit values for each type of prediction
  • Especially useful in imbalanced datasets
  • Also, fraud detection, outlier detection, etc.

• Utility is computed as \[U = \sum^{n_c}_{i=1}{\sum^{n_c}_{k=1}{CM_{i,k}\times CB_{i,k}}}\]

• CM: Confusion matrix

• CB: Cost/benefit matrix

• When you have a binary classification

• Precision: rate of correctly identified trues to all predicted as true.

Prec = \(\frac{TP}{TP+FP}\)

• Recall: rate of correctly identified trues to all actual trues.

Rec = \(\frac{TP}{TP+FN}\)

• You can aggregate precision and recall into one metric, the F-measure:

\(F_\beta = \frac{(\beta^2+1)\times Prec \times Rec}{\beta^2\times Prec + Rec}\)

• For numeric target variables, one frequently used metric is the mean squared error:

\(\displaystyle MSE = \frac{1}{N_{test}}{\sum^{N_{test}}_{i=1}{(\hat y_i - y_i)^2}}\)

• Or for the sake of unit compliance, use root mean squared error:

\(\displaystyle RMSE = \sqrt{MSE}\)

• Or, alternatively use mean absolute error:

\(\displaystyle MAE = \frac{1}{N_{test}}{\sum^{N_{test}}_{i=1}{|\hat y_i - y_i|}}\)

• You can use a baseline method to produce relative error metrics.
• A baseline method is something naive, such as the mean \(y\) value
• Normalized mean squared error:

\(\displaystyle NMSE = \frac{\sum^{N_{test}}_{i=1}{(\hat y_i - y_i)^2}}{\sum^{N_{test}}_{i=1}{(\bar y_i - y_i)^2}}\)

• We expect NMSE to be close to 0. A value of 1 means a performance as bad as the baseline.
• Also, Normalized mean absolute error

\(\displaystyle NMAE = \frac{\sum^{N_{test}}_{i=1}{|\hat y_i - y_i|}}{\sum^{N_{test}}_{i=1}{|\bar y_i - y_i|}}\)

• There are many implementations of these metrics
  • function mmetric in package rminer (Cortez, 2015)
  • functions classificationMetrics and regressionMetrics in package performanceEstimation (Torgo, 2014a)
  • function performance in package ROCR (Sing et al., 2009)
  • function performance in package mlr (Bischl et al., 2016)
• And, you can always compute them on the fly.

• Load house.data into R
• Apply clustering to the data
  • How many clusters seems to be the optimal?
• Apply anomaly detection to the data
  • Do you catch a few or many anomalies?