Bad incrementa​lRegressio​nLinear prediction results using simple slope 1x data

Hello All,

I am a new user of the Statistics and Machine Learning Toolbox and want to test the IncrementalRegressionLinear blocks with simple synthetic data with a slope of 1:

To simulate streaming data (e.g. from sensors) I preprocessed the data as follows:

rng(0,"twister") % For reproducibility

% Extract time and value columns
time = data.time;
value = data.value;

% Split the data into training and test set
n = length(time);
split_idx = round(0.7 * n);
data_train = data(1:split_idx,:)

% Training set
time_train = time(1:split_idx);
value_train = value(1:split_idx);

% Test set
time_test = time(split_idx + 1:end);
value_test = value(split_idx + 1:end);

n = numel(time_train);
p = size(value_train,2); % Number of predictors



numObsPerChunk = 11;
nchunk = floor(n/numObsPerChunk);
for j = 1:nchunk
    ibegin = min(n,numObsPerChunk*(j-1) + 1);
    iend = min(n,numObsPerChunk*j);
    idx = ibegin:iend;   
    Xin(:,:,j) = value_train(idx,:);
    Yin(:,j) = time_train(idx);
    Xtestset(1,:,j) = 140;
    ytrue(1,:,j) = 361;
end

k = size(Xin,3); % Number of data chunks
t = 0:k-1;
X_ts = timeseries(Xin,t,InterpretSingleRowDataAs3D=true);
Y_ts = timeseries(Yin',t,InterpretSingleRowDataAs3D=true);
Xtest_ts = timeseries(Xtestset,t,InterpretSingleRowDataAs3D=true);
ytest_ts = timeseries(ytestset,t,InterpretSingleRowDataAs3D=true);
ytrue_ts = timeseries(ytrue,t,InterpretSingleRowDataAs3D=true);

% incrementalRegressionLinear

Mdl = incrementalRegressionLinear(NumPredictors=p,  Learner='svm',  Solver='scale-invariant', Shuffle=false  ,...
    Standardize=true,EstimationPeriod=110, MetricsWarmupPeriod=11, MetricsWindowSize=11);
linearMdl = Mdl;

This is the 'connected' simulink model:

And this the plotted result:

For a value of 140 the true y would be 361.

So i was wondering why the prediction is so bad? As the data is 100% linear with a slope of 1 would assume that the prediction is also 361. But it varies around 300-375.

Does anyone knows why? Do i have a complete missunderstanding or did i forget something?

It is based on the matlab example but simplyfied: https://de.mathworks.com/help/stats/configure-simulink-template-for-conditionally-enabled-regression.html?searchHighlight=conditional+incremental+learning&s_tid=srchtitle_support_results_7_conditional%2520incremental%2520learning

Best regards

Christoph

Answer 1

[Prathamesh]

Hi @Christoph,

I understand that you have a table that contains time and value. And you want to test the ‘IncrememntalRegressionLinearl’ block.

Below are the required changes that might solve the issue:

1. The plot of ‘data.time’ vs ‘data.value’ clearly shows a negative linear relationship. As ‘time’ increases the ‘value decreases’. Specifically the slope is ‘-1’.
2. In the Simulink model , you connected ‘value_train’ to the ‘x’ input of the ‘IncrementalRegressionLinear Fit’ block and ‘time_train’ to ‘y’ input. This means the model is learning to predict time based on value, not value based on time.

Hope this helps.