incrementalLearner

Train RRCF Model

Fit an RRCF model to the training data. Specify an anomaly contamination fraction of 0.001.

rng(0,"twister"); % For reproducibility
TTforest = rrcforest(Xtrain,ContaminationFraction=0.001);
details(TTforest)

  RobustRandomCutForest with properties:

        CollusiveDisplacement: 'maximal'
                  NumLearners: 100
    NumObservationsPerLearner: 256
                           Mu: []
                        Sigma: []
        CategoricalPredictors: [2 4 6 7 8 9 10 14 15]
        ContaminationFraction: 1.0000e-03
               ScoreThreshold: 55.5745
               PredictorNames: {'age'  'workClass'  'fnlwgt'  'education'  'education_num'  'marital_status'  'occupation'  'relationship'  'race'  'sex'  'capital_gain'  'capital_loss'  'hours_per_week'  'native_country'  'salary'}

TTforest is a RobustRandomCutForest model object representing a traditionally trained RRCF model. The software identifies nine variables in the data as categorical predictors because they contain string arrays.

Convert Trained Model

Convert the traditionally trained RRCF model to an RRCF model for incremental learning.

Incrementalforest = incrementalLearner(TTforest);

Incrementalforest is an incrementalRobustRandomCutForest model object that is ready for incremental learning and anomaly detection.

Fit Incremental Model and Detect Anomalies

Perform incremental learning on the Xstream data by using the fit function. To simulate a data stream, fit the model in chunks of 100 observations at a time. At each iteration:

n = numel(Xstream(:,1));
numObsPerChunk = 100;
nchunk = floor(n/numObsPerChunk);
medianscore = zeros(nchunk,1);
numAnom = zeros(nchunk,1);
threshold = zeros(nchunk,1);

% Incremental fitting
for j = 1:nchunk
    ibegin = min(n,numObsPerChunk*(j-1) + 1);
    iend = min(n,numObsPerChunk*j);
    idx = ibegin:iend;    
    [Incrementalforest,tf,scores] = fit(Incrementalforest,Xstream(idx,:));
    medianscore(j) = median(scores);
    numAnom(j) = sum(tf);
    threshold(j) = Incrementalforest.ScoreThreshold;
end

Analyze Incremental Model During Training

To see how the median score, score threshold, and number of detected anomalies per chunk evolve during training, plot them on separate tiles.

tiledlayout(3,1);
nexttile
plot(medianscore)
ylabel("Median Score")
xlabel("Iteration")
xlim([0 nchunk])
nexttile
plot(threshold)
ylabel("Score Threshold")
xlabel("Iteration")
xlim([0 nchunk])
nexttile
plot(numAnom,"+")
ylabel("Anomalies")
xlabel("Iteration")
xlim([0 nchunk])
ylim([0 max(numAnom)+0.2])

totalanomalies=sum(numAnom)

totalanomalies = 1

anomfrac= totalanomalies/n

anomfrac = 5.0000e-04

fit updates the model and returns the observation scores and the indices of observations with scores above the score threshold value as anomalies. A high score value indicates a normal observation, and a low value indicates an anomaly. The median score fluctuates between approximately 230 and 270. The score threshold rises from a value of 260 after the first iteration and steadily approaches 285 after 12 iterations. The software detected 4 anomalies in the Xstream data, yielding a total contamination fraction of 0.002.

Create Simulated Data Stream

Create a simulated data stream of observations representing a noisy sinusoid signal.

rng(0,"twister"); % For reproducibility
period = 100;
n = 2001+period;
sigma = 0.04;
a = linspace(1,n,n)';
b = sin(2*pi*(a-1)/period)+sigma*randn(n,1);

Introduce an anomalous region into the data stream. Plot the data stream portion that contains the anomalous region, and circle the anomalous data points.

c = 2*(sin(2*pi*(a-35)/period)+sigma*randn(n,1));
b(1150:1170) = c(1150:1170);
scatter(a,b,".")
xlim([900,1200])
xlabel("Observation")
hold on
scatter(a(1150:1170),b(1150:1170),"r")
hold off

Convert the single-featured data set b into a multi-featured data set by shingling [1] with a shingle size equal to the period of the signal. The $$i$$th shingled observation is a vector of $$k$$ features with values $$b_i$$, $$b_{i+1}$$, ..., $$b_{i+k-1}$$, where $$k$$ is the shingle size.

X = [];
shingleSize = period;
for i = 1:n-shingleSize
    X = [X;b(i:i+shingleSize-1)'];
end

Train Model and Perform Incremental Anomaly Detection

Fit a robust random cut forest model to the first 1000 shingled observations, specifying a contamination fraction of 0. Convert the model to an incrementalRobustRandomCutForest model object. Specify to keep the 100 most recent observations relevant for anomaly detection.

Mdl = rrcforest(X(1:1000,:),ContaminationFraction=0);
IncrementalMdl = incrementalLearner(Mdl,NumObservationsToKeep=100);

To simulate a data stream, process the full shingled data set in chunks of 100 observations at a time. At each iteration:

n = numel(X(:,1));
numObsPerChunk = 100;
nchunk = floor(n/numObsPerChunk);
anomIdx = [];
allscores = [];

% Incremental fitting
rng("default"); % For reproducibility
for j = 1:nchunk
    ibegin = min(n,numObsPerChunk*(j-1) + 1);
    iend = min(n,numObsPerChunk*j);
    idx = ibegin:iend;
    [isanom,scores] = isanomaly(IncrementalMdl,X(idx,:));
    allscores = [allscores;scores];
    anomIdx = [anomIdx;find(isanom)+ibegin-1];
    if (sum(isanom) < 3)
        IncrementalMdl = fit(IncrementalMdl,X(idx,:));
    end
end

Analyze Incremental Model During Training

At each iteration, the software calculates a score value for each observation in the data chunk. A negative score value with large magnitude indicates a normal observation, and a large positive value indicates an anomaly. Plot the anomaly score for the observations in the vicinity of the anomaly. Circle the scores of shingles that the software returns as anomalous.

figure
scatter(a(1:2000),allscores,".")
hold on
scatter(a(anomIdx),allscores(anomIdx),20,"or")
xlim([900,1200])
xlabel("Shingle")
ylabel("Score")
hold off

Because the introduced anomalous region begins at observation 1150, and the shingle size is 100, shingle 1051 is the first to show a high anomaly score. Some shingles between 1050 and 1170 have scores lying just below the anomaly score threshold, due to the noise in the sinusoidal signal. The shingle size affects the performance of the model by defining how many subsequent consecutive data points in the original time series the software uses to calculate the anomaly score for each shingle.

Plot the unshingled data and highlight the introduced anomalous region. Circle the observation number of the first element in each shingle returned by that the software as anomalous.

figure
xlim([900,1200])
ylim([-1.5 2])
rectangle(Position=[1150 -1.5 20 3.5],FaceColor=[0.9 0.9 0.9], ...
    EdgeColor=[0.9 0.9 0.9])
hold on
scatter(a,b,".")
scatter(a(anomIdx),b(anomIdx),20,"or")
xlabel("Observation")
hold off