Calculating Autocorrelation

Autocorrelation

Autocorrelation is a useful statistic for assessing mixing of a Markov chain. Keanu provides a method of calculating autocorrelation on samples.

Example

With a network defined, we can get the autocorrelation vertex A. The result is a tensor containing the autocorrelation at varying lags.

NetworkSamples posteriorSamples = MetropolisHastings.withDefaultConfig().getPosteriorSamples(
    bayesNet,
    bayesNet.getLatentVertices(),
    100
);
DoubleTensor autocorrelation = posteriorSamples.getDoubleTensorSamples(A).getAutocorrelation();

When the samples are tensors, we need to specify the tensor index on which to calculate the autocorrelation. For example, if the sample shape is [1,5] we can evaluate the autocorrelation at index [0,1].

NetworkSamples posteriorSamples = MetropolisHastings.withDefaultConfig().getPosteriorSamples(
    bayesNet,
    bayesNet.getLatentVertices(),
    100
);
DoubleTensor autocorrelation = posteriorSamples.getDoubleTensorSamples(A).getAutocorrelation(0,1);