Main Content

incrementalConceptDriftDetector

Instantiate incremental concept drift detector

Since R2022a

    Description

    IncCDDetector = incrementalConceptDriftDetector() returns an incremental concept drift detector that utilizes the default method, Hoeffding's Bounds Drift Detection Method with moving average test (HDDMA).

    IncCDDetector = incrementalConceptDriftDetector(DetectionMethod) returns an incremental concept drift detector that utilizes the method DetectionMethod.

    example

    IncCDDetector = incrementalConceptDriftDetector(DetectionMethod,Name=Value) specifies additional options using one or more Name=Value arguments.

    example

    Examples

    collapse all

    Initiate the concept drift detector using the Drift Detection Method (DDM).

    incCDDetector = incrementalConceptDriftDetector("ddm");

    Create a random stream such that for the first 1000 observations, failure rate is 0.1 and after 1000 observations, failure rate increases to 0.6.

    rng(1234)  % For reproducibility
    numObservations = 3000;
    switchPeriod = 1000;
    
    for i = 1:numObservations
        if i <= switchPeriod
           failurerate = 0.1;
        else
           failurerate = 0.6;
        end
           X(i) = rand()<failurerate; % Value 1 represents failure
    end

    Preallocate variables for tracking drift status.

    status = zeros(numObservations,1);
    statusname = strings(numObservations,1);

    Continuously feed the data to the drift detector and perform incremental drift detection. At each iteration:

    • Update statistics of the drift detector and monitor for drift using the new data point with detectdrift. (Note: detectdrift checks for drift after the warmup period.)

    • Track and record the drift status for visualization purposes.

    • When a drift is detected, reset the incremental concept drift detector by using reset.

    for i = 1:numObservations     
        
        incCDDetector = detectdrift(incCDDetector,X(i));
        statusname(i) = string(incCDDetector.DriftStatus);
              
        if incCDDetector.DriftDetected
           status(i) = 2;
           incCDDetector = reset(incCDDetector); % If drift detected, reset the detector
           sprintf("Drift detected at Observation #%d. Detector reset.",i)
        elseif incCDDetector.WarningDetected
           status(i) = 1;
        else 
           status(i) = 0;
        end   
    end
    ans = 
    "Drift detected at Observation #1078. Detector reset."
    

    After the change in the failure rate at observation number 1000, detectdrift detects the shift at observation number 1078.

    Plot the drift status versus the observation number.

    gscatter(1:numObservations,status,statusname,'gyr','*',4,'on',"Observation number","Drift status")

    Figure contains an axes object. The axes object with xlabel Observation number, ylabel Drift status contains 3 objects of type line. One or more of the lines displays its values using only markers These objects represent Stable, Warning, Drift.

    Create a random stream such that the observations come from a normal distribution with standard deviation 0.75, but the mean changes over time. First 1000 observations come from a distribution with mean 2, the next 1000 come from a distribution with mean 4, and the following 1000 come from a distribution with mean 7.

    rng(1234) % For reproducibility
    numObservations = 3000;
    switchPeriod1 = 1000;
    switchPeriod2 = 2000;
    X = zeros([numObservations 1]);
    
    % Generate the data
    for i = 1:numObservations
       if i <= switchPeriod1
          X(i) = normrnd(2,0.75);
       elseif i <= switchPeriod2
          X(i) = normrnd(4,0.75);
       else
          X(i) = normrnd(7,0.75);
       end
    end

    In an incremental drift detection application, access to data stream and model update would happen consecutively. One would not collect the data first and then feed into the model. However, for the purpose of clarification, this example demonstrates the simulation of data separately.

    Specify the drift warmup period as 50 observations and estimation period for the data input bounds as 100.

    driftWarmupPeriod = 50;
    estimationPeriod = 100;

    Initiate the incremental concept drift detector. Utilize the Hoeffding's bounds method with exponentially weighted moving average method (EWMA). Specify the input type and warmup period.

    incCDDetector = incrementalConceptDriftDetector("hddmw",InputType="continuous", ...
                    WarmupPeriod=driftWarmupPeriod,EstimationPeriod=estimationPeriod)
    incCDDetector = 
      HoeffdingDriftDetectionMethod
    
            PreviousDriftStatus: 'Stable'
                    DriftStatus: 'Stable'
                         IsWarm: 0
        NumTrainingObservations: 0
                    Alternative: 'greater'
                      InputType: 'continuous'
                     TestMethod: 'ewma'
    
    
      Properties, Methods
    
    

    incDDetector is a HoeffdingDriftDetectionMethod object. When you first create the object, properties such as DriftStatus, IsWarm, CutMean, and NumTrainingObservations are at their initial state. detectdrift updates them as you feed the data incrementally and monitor for drift.

    Preallocate the batch size and the variables to record drift status and the mean the drift detector computes with each income of data.

    status = zeros([numObservations 1]);
    statusname = strings([numObservations 1]);
    M = zeros([numObservations 1]);

    Simulate the data stream of one observation at a time and perform incremental drift detection. At each iteration:

    • Monitor for drift using the new data with detectdrift.

    • Track and record the drift status and the statistics for visualization purposes.

    • When a drift is detected, reset the incremental concept drift detector by using the function reset.

    for i = 1:numObservations
        
        incCDDetector = detectdrift(incCDDetector,X(i));
        
        M(i) = incCDDetector.Mean;
            
        if incCDDetector.DriftDetected
            status(i) = 2;
            statusname(i) = string(incCDDetector.DriftStatus);
            incCDDetector = reset(incCDDetector); % If drift detected, reset the detector
            sprintf("Drift detected at observation #%d. Detector reset.",i)
        elseif incCDDetector.WarningDetected
            status(i) = 1;
            statusname(i) = string(incCDDetector.DriftStatus);
            sprintf("Warning detected at observation #%d.",i)
        else 
            status(i) = 0;
            statusname(i) = string(incCDDetector.DriftStatus);
        end      
    end
    ans = 
    "Warning detected at observation #1024."
    
    ans = 
    "Warning detected at observation #1025."
    
    ans = 
    "Warning detected at observation #1026."
    
    ans = 
    "Warning detected at observation #1027."
    
    ans = 
    "Warning detected at observation #1028."
    
    ans = 
    "Warning detected at observation #1029."
    
    ans = 
    "Drift detected at observation #1030. Detector reset."
    
    ans = 
    "Warning detected at observation #2012."
    
    ans = 
    "Warning detected at observation #2013."
    
    ans = 
    "Warning detected at observation #2014."
    
    ans = 
    "Drift detected at observation #2015. Detector reset."
    

    Plot the drift status versus the observation number.

    gscatter(1:numObservations,status,statusname,'gyr','*',5,'on',"Number of observations","Drift status")

    Figure contains an axes object. The axes object contains 3 objects of type line. These objects represent Stable, Warning, Drift.

    Plot the mean values versus the number of observations.

    scatter(1:numObservations,M)

    Figure contains an axes object. The axes object contains an object of type scatter.

    You can see the increase in the sample mean from the plot. The mean value becomes larger and the software eventually detects the drift in the data. Once a drift is detected, reset the incremental drift detector. This also resets the mean value. In the plot, the observations where the sample mean is zero correspond to the estimation periods. There is an estimation period at the beginning and then twice after the drift detector is reset following the detection of a drift.

    Initiate the concept drift detector using the Drift Detection Method (DDM).

    incCDDetector = incrementalConceptDriftDetector("ddm",Alternative="less",WarmupPeriod=100);

    Create a random stream such that for the first 1000 observations, failure rate is 0.4 and after 1000 failure rate decreases to 0.1.

    rng(1234)  % For reproducibility
    numObservations = 3000;
    switchPeriod = 1000;
    for i = 1:numObservations
        if i <= switchPeriod
           failurerate = 0.4;
        else
           failurerate = 0.125;
        end
           X(i) = rand()<failurerate; % Value 1 represents failure
    end

    Preallocate variables for tracking drift status and the optimal mean and optimal standard deviation value.

    optmean = zeros(numObservations,1);
    optstddev = zeros(numObservations,1);
    status = zeros(numObservations,1);
    statusname = strings(numObservations,1);

    Continuously feed the data to the drift detector and monitor for any potential change. Record the drift status for visualization purposes.

    for i = 1:numObservations     
        
        incCDDetector = detectdrift(incCDDetector,X(i)); 
    
        statusname(i) = string(incCDDetector.DriftStatus);
        optmean(i) = incCDDetector.OptimalMean;
        optstddev(i) = incCDDetector.OptimalStandardDeviation;
    
        if incCDDetector.DriftDetected
           status(i) = 2;
           incCDDetector = reset(incCDDetector); % If drift detected, reset the detector
           sprintf("Drift detected at Observation #%d. Detector reset.",i)
        elseif incCDDetector.WarningDetected
           status(i) = 1;
        else 
           status(i) = 0;
        end   
    end
    ans = 
    "Drift detected at Observation #1107. Detector reset."
    

    After the change in the failure rate at observation number 1000, detectdrift detects the shift at observation number 1096.

    Plot the change in the optimal mean and optimal standard deviation.

    tiledlayout(2,1);
    ax1 = nexttile;
    plot(ax1,1:numObservations,optmean)
    ax2 = nexttile;
    plot(ax2,1:numObservations,optstddev)

    Figure contains 2 axes objects. Axes object 1 contains an object of type line. Axes object 2 contains an object of type line.

    Plot the drift status versus the observation number.

    figure();
    gscatter(1:numObservations,status,statusname,'gyr','*',4,'on',"Observation number","Drift status")

    Figure contains an axes object. The axes object with xlabel Observation number, ylabel Drift status contains 3 objects of type line. One or more of the lines displays its values using only markers These objects represent Stable, Warning, Drift.

    detectdrift concludes on a warning status for multiple observations before it decides on a drift.

    Input Arguments

    collapse all

    Incremental drift detection method, specified as one of the following.

    Detection MethodDefinition
    "ddm"Drift Detection Method (DDM)
    "hddma"Hoeffding's Bounds Drift Detection Method with moving average test (HDDMA)
    "hddmw"Hoeffding's Bounds Drift Detection Method with exponentially weighted moving average (EWMA) test (HDDMW)

    Name-Value Arguments

    Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

    Example: Alternative="less",InputType="continuous",InputBounds=[-1,1],ForgettingFactor=0.075 specifies the alternative hypothesis as less, that is, left-sided, the input data type as continuous data, lower and upper bounds of the input data as [-1,1] and the value of the forgetting factor for the HDDMW method as 0.075.

    General Options

    collapse all

    Type of alternative hypothesis for determining drift status, specified as one of "unequal", "greater", or "less". Given two test statistics F1(x) and F2(x),

    • "greater" tests for a drift in the positive direction, that is, F1(x)>F2(x).

      In this case, the null hypothesis is F1(x)F2(x).

    • "less" tests for a drift in the negative direction, that is, F1(x)<F2(x).

      In this case, the null hypothesis is F1(x)F2(x).

    • "unequal" tests for a drift in the either direction, that is, F1(x)F2(x).

      In this case, the null hypothesis is F1(x)=F2(x).

      "unequal" is for the HDDMA and HDDMW methods only.

    For each type of test, detectdrift updates the statistics and checks whether it can reject the null hypothesis in favor of the alternative at the significance level of WarningThreshold or DriftThreshold. If it rejects the null hypothesis at the significance level of WarningThreshold, then it updates the DriftStatus to 'Warning'. If it rejects the null hypothesis at the DriftThreshold, then it updates the DriftStatus to 'Drift'.

    Example: Alternative="less"

    Type of input to the drift detector, specified as either "binary" or "continuous".

    Example: InputType="continuous"

    Number of observations used for drift detector to warm up, specified as a nonnegative integer. Until the end of the warmup period, detectdrift trains the drift detector using the incoming data and updates the internal statistics, but does not check for the drift status. After the software reaches the warmup period, that is, once the drift detector is warm, it starts checking for any changes in the drift status.

    Example: WarmupPeriod=50

    Data Types: double | single

    Options for DDM

    collapse all

    Number of standard deviations for drift limit, specified as a nonnegative scalar value. This is the number of standard deviations the overall test statistic value can be away from the optimal test statistic value before the drift detector sets the drift status to drift. Default value of 3 corresponds to a 99.7% confidence level [1].

    DriftThreshold value must be strictly greater than the WarningThreshold value.

    Example: DriftThreshold=2.5

    Data Types: double | single

    Number of standard deviations for warning limit, specified as a nonnegative scalar value. This is the number of standard deviations the overall test statistic value can be away from the optimal test statistic value before the drift detector sets the drift status to warning. Default value of 2 corresponds to a 95% confidence level [1].

    WarningThreshold value must be strictly smaller than the DriftThreshold value.

    Example: WarningThreshold=1.75

    Data Types: double | single

    Options for HDDMA and HDDMW

    collapse all

    Threshold to determine if drift exists, specified as a nonnegative scalar value from 0 to 1. It is the significance level the drift detector uses for calculating the allowed error between a random variable and its expected value in Hoeffding's inequality and McDiarmid's inequality before it sets the drift status to drift [2].

    DriftThreshold value must be strictly smaller than the WarningThreshold value.

    Example: DriftThreshold=0.003

    Data Types: double | single

    Number of observations used to estimate the input bounds for continuous data, specified as a nonnegative integer. That is, when InputType is "continuous" and you did not specify the InputBounds value, the software uses EstimationPeriod number of observations to estimate the input bounds. After the estimation period, the software starts the warmup period.

    If you specify the InputBounds value or InputType is "binary", then the software ignores EstimationPeriod.

    Default value is 100 when there is a need for estimating the input bounds. Otherwise, default value is 0.

    Example: EstimationPeriod=150

    Data Types: double | single

    Lower and upper bounds of continuous input data, specified as a numeric vector of size 2.

    If InputType is "continuous" and you do not specify the InputBounds value, then detectdrift estimates the bounds from the data during the estimation period. Specify the number of observations to estimate the data input bounds by using EstimationPeriod.

    If InputType is "binary", then the drift detector sets the InputBounds value to [0,1] and the software ignores the InputBounds name-value argument.

    HDDM uses Hoeffding's inequality and McDiarmid's inequality for drift detection and these inequalities assume bounded inputs [2].

    Example: InputBounds=[-1 1]

    Data Types: double | single

    Note

    This option is only for the exponentially weighted moving average (EWMA) method (corresponding to DetectionMethod value set as "hddmw").

    Forgetting factor in the HDDMW method, specified as a scalar value from 0 to 1. Forgetting factor is the λ in the EWMA statistic X^t=λXt+(1λ)X^t1[2]. Forgetting factor determines how much the current prediction of mean is influenced by the past observations. A higher value of ForgettingFactor attains more weight to the current observations and less value to the past observations.

    Example: ForgettingFactor=0.075

    Data Types: double | single

    Threshold to determine warning versus drift, specified as a nonnegative scalar value from 0 to 1. It is the significance level the drift detector uses for calculating the allowed error between a random variable and its expected value in Hoeffding's inequality and McDiarmid's inequality before it sets the drift status to warning [2].

    WarningThreshold value must be strictly greater than DriftThreshold value.

    Example: WarningThreshold=0.007

    Data Types: double | single

    Output Arguments

    collapse all

    Incremental concept drift detector, specified as either DriftDetectionMethod or HoeffdingDriftDetectionMethod object. For more information on these objects and their properties, see the corresponding reference pages.

    References

    [1] Gama, Joao, Pedro Medas, Gladys Castillo, and Pedro P. Rodrigues. “Learning with drift detection.“ In Brazilian symposium on artificial intelligence, pp. 286-295. Berlin, Heidelberg: Springer. 2004, September.

    [2] Frias-Blanco, Isvani, Jose del Campo-Ávila, Ramos-Jimenez Gonzalo, Rafael Morales-Bueno, Augustin Ortiz-Diaz, and Yaile Caballero-Mota. “Online and non-parametric drift detection methods based on Hoeffding's bounds.“ IEEE Transactions on Knowledge and Data Engineering, Vol. 27, No. 3, pp.810-823. 2014.

    Version History

    Introduced in R2022a