MiniRocketClassifier<T>

Models & Types · AiDotNet.Classification.TimeSeries

Implements MiniRocket for time series classification.

Inheritance: ClassifierBase<T> → MiniRocketClassifier<T>

Implements: IGeneratedParameterRegistrar<T>, ITimeSeriesClassifier<T>

For Beginners

MiniRocket is a faster, simpler version of ROCKET that uses deterministic kernels with fixed weights. It achieves similar accuracy to ROCKET while being much faster and more memory efficient.

How It Works

Key differences from ROCKET:

  • Uses only PPV (proportion of positive values) features, not max values
  • Kernels have fixed weights from {-1, 2} (instead of random weights)
  • Uses a fixed set of 84 base kernels (instead of random kernels)
  • Biases are computed from quantiles of convolution outputs

How MiniRocket works:

  • Define 84 deterministic kernel patterns using weights from {-1, 2}
  • For each kernel, compute convolution at multiple dilations
  • Compute bias values from quantiles of convolution outputs
  • Extract PPV features using each (kernel, dilation, bias) combination
  • Train a linear classifier on the extracted features

Reference: Dempster et al., “MiniRocket: A Very Fast (Almost) Deterministic Transform for Time Series Classification” (2021)

Example

// Create MiniRocket classifier with deterministic kernels for fast time series classification
var options = new MiniRocketOptions<double>();
            
// Prepare time series data: rows are samples, columns are time steps
var features = new Matrix<double>(4, 5);
features[0, 0] = 1.0; features[0, 1] = 1.2; features[0, 2] = 1.5; features[0, 3] = 1.3; features[0, 4] = 1.1;
features[1, 0] = 1.1; features[1, 1] = 1.3; features[1, 2] = 1.4; features[1, 3] = 1.2; features[1, 4] = 1.0;
features[2, 0] = 2.0; features[2, 1] = 2.5; features[2, 2] = 2.3; features[2, 3] = 2.8; features[2, 4] = 3.0;
features[3, 0] = 2.1; features[3, 1] = 2.4; features[3, 2] = 2.6; features[3, 3] = 2.9; features[3, 4] = 3.1;
var labels = new Vector<double>(new double[] { 0, 0, 1, 1 });
            
// Train: extract PPV features using fixed {-1, 2} kernels and fit classifier
var result = new AiModelBuilder<double, Matrix<double>, Vector<double>>()
    .ConfigureModel(new MiniRocketClassifier<double>(options))
    .Build(features, labels);
            
// Predict class for new time series
var newSample = new Matrix<double>(1, 5);
newSample[0, 0] = 1.0; newSample[0, 1] = 1.1; newSample[0, 2] = 1.3; newSample[0, 3] = 1.2; newSample[0, 4] = 1.0;
var predictions = result.Predict(newSample);
// Result is available in the returned value

Constructors

Constructor Summary
MiniRocketClassifier(MiniRocketOptions<T> options) Creates a new MiniRocket classifier.

Properties

Property Summary
bool SupportsVariableLengths { get; } Gets whether this classifier supports variable-length sequences.
int NumChannels { get; set; } Gets the number of channels (variables) in the time series.
int SequenceLength { get; set; } Gets the expected sequence length.

Methods

Method Summary
IFullModel<T, Matrix<T>, Vector<T>> WithParameters(Vector<T> parameters)
Matrix<T> PredictSequenceProbabilities(Tensor<T> sequences) Predicts class probabilities for time series sequences.
ModelOptions GetOptions()
Vector<T> Predict(Matrix<T> input)
Vector<T> PredictSequences(Tensor<T> sequences) Predicts class labels for time series sequences.
void Train(Matrix<T> x, Vector<T> y)
void TrainOnSequences(Tensor<T> sequences, Vector<T> labels) Trains the MiniRocket classifier on time series sequences.