Predictive models on 1D signals in a small-data environment

Fabio Coppini (1) , Yiye Jiang (2) , Sonia Tabti

1 LPSM (UMR_8001) - Laboratoire de Probabilités, Statistique et Modélisation
2 IMB - Institut de Mathématiques de Bordeaux

This report is concerned by one project done during the Semaine d'Études Mathématiques et Entreprises (SEME) at the Institut de Mathématiques de Bordeaux. The subject, proposed by the company FieldBox.ai, concerns the use of machine learning algorithms and data augmentation techniques, applied to small datasets composed of 1D signals measurements. The target variable is supposed to be continuous, i.e., a regression problem. By first reviewing the literature and existing methods on data augmentation, we propose two procedures to tackle this problem: one allows to create synthetic observations for a specific range of target values and it is based on a perturbation method in Fourier/Wavelet space; the other is based on neural networks and uses a particular version of the Variational Autoencoder known as LSTM-VAE. Our methods are applied to an open dataset, available at the UCI repository, and show encouraging results for a common class of machine learning algorithms.

Mots clés: small-data; machine learning; inferring techniques; data augmentation; imputation; Variational Autoencoder; LSTM; Fast Fourier Transform; Discrete Wavelet Transform; time series; class imbalance

Domaines: Traitement du signal et de l'image [eess.SP]; Machine Learning [stat.ML]; Apprentissage [cs.LG]

Fabio Coppini, Yiye Jiang, Sonia Tabti. Predictive models on 1D signals in a small-data environment. [Research Report] IMB - Institut de Mathématiques de Bordeaux. 2021. ⟨hal-03211100⟩