This dictionary learning variant is tailored for dealing with multivariate datasets and especially timeseries, where samples are matrices and the dataset is seen as a tensor. Dictionary Learning Algorithm (DLA) decompose input vector on a dictionary matrix with a sparse coefficient vector, see (a) on figure below. To handle multivariate data, a first approach called multichannel DLA, see (b) on figure below, is to decompose the matrix vector on a dictionary matrix but with sparse coefficient matrices, assuming that a multivariate sample could be seen as a collection of channels explained by the same dictionary. Nonetheless, multichannel DLA breaks the "spatial" coherence of multivariate samples, discarding the column-wise relationship existing in the samples. Multivariate DLA, (c), on figure below, decompose the matrix input on a tensor dictionary, where each atom is a matrix, with sparse coefficient vectors. In this case, the spatial relationship are directly encoded in the dictionary, as each atoms has the same dimension than an input samples.

(figure from Chevallier et al., 2014 )

To handle timeseries, two major modifications are brought to DLA:

1. extension to multivariate samples
2. shift-invariant approach, The first point is explained above. To implement the second one, there is two possibility, either slicing the input timeseries into small overlapping samples or to have atoms smaller than input samples, leading to a decomposition with sparse coefficients and offsets. In the latter case, the decomposition could be seen as sequence of kernels occuring at different time steps.

(figure from Smith & Lewicki, 2005)

The proposed implementation is an adaptation of the work of the following authors:

• Q. Barthélemy, A. Larue, A. Mayoue, D. Mercier, and J.I. Mars. Shift & 2D rotation invariant sparse coding for multi- variate signal. IEEE Trans. Signal Processing, 60:1597–1611, 2012.
• Q. Barthélemy, A. Larue, and J.I. Mars. Decomposition and dictionary learning for 3D trajectories. Signal Process., 98:423–437, 2014.
• Q. Barthélemy, C. Gouy-Pailler, Y. Isaac, A. Souloumiac, A. Larue, and J.I. Mars. Multivariate temporal dictionary learning for EEG. Journal of Neuroscience Methods, 215:19–28, 2013.

The only dependencies are scikit-learn, matplotlib, numpy and scipy.

No installation is required.

A straightforward example is:

import numpy as np
from mdla import MultivariateDictLearning
from mdla import multivariate_sparse_encode
from numpy.linalg import norm

rng_global = np.random.RandomState(0)
n_samples, n_features, n_dims = 10, 5, 3
X = rng_global.randn(n_samples, n_features, n_dims)

n_kernels = 8
dico = MultivariateDictLearning(n_kernels=n_kernels, max_iter=10).fit(X)
residual, code = multivariate_sparse_encode(X, dico)
print ('Objective error for each samples is:')
for i in range(len(residual)):
    print ('Sample', i, ':', norm(residual[i], 'fro') + len(code[i]))

• Chevallier, S., Barthelemy, Q., & Atif, J. (2014). Subspace metrics for multivariate dictionaries and application to EEG. In Acoustics, Speech and Signal Processing (ICASSP), IEEE International Conference on (pp. 7178-7182).
• Smith, E., & Lewicki, M. S. (2005). Efficient coding of time-relative structure using spikes. Neural Computation, 17(1), 19-45
• Chevallier, S., Barthélemy, Q., & Atif, J. (2014). On the need for metrics in dictionary learning assessment. In European Signal Processing Conference (EUSIPCO), pp. 1427-1431.