Laurent Oudre

Researcher in signal processing, pattern recognition and machine learning

Machine Learning for Time Series (Master MVA)

Teaching material and outline of the course Machine Learning for Time Series (Master MVA) during 2020-2021.

Course description

In many application contexts (health, economy, advertising...), the data collected takes the form of time series. The fundamental challenge then consists in choosing an adapted representation, allowing to take into account the temporal information as well as possible.

Machine Learning for time series gathers a large number of unsupervised or supervised tasks such as prediction, classification, completion/interpolation, query by content/indexation, clustering, segmentation/change-point detection or anomaly detection. But in reality, most of work for a data scientist dealing with temporal data consists in a series hidden tasks such as:

• Understand the data: know where they come from, how they were acquired, what are their characteristics, interact with domain-experts and understand their problems
• Improve the data: find accurate representation spaces where the events of interest can be seen, consolidate the data (denoising, detrending, detection/removal of outliers)
• Model the data: physical/statistical or expert-based models, simple, adaptive and interpretable models
• Extract information from the data: find repetitive patterns, features of interest

This course aims to provide an overview of ML techniques to study time series, in different tasks such as pattern extraction and recognition, anomaly detection, prediction, interpolation etc. The course will mostly focus on these often poorly-documented hidden tasks and introduce several recent ML methods that will help the future data scientist to mine, but above all to understand time series. The course will be widely illustrated with real data and problems from current challenges and will emphasize aspects related to data understanding and interpretation. Note that in its current form, the course will only briefly discuss Deep Learning algorithms.

Course sessions (preliminary outline)

Introduction
Session 1: Pattern Recognition and Detection Problem statement Comparing time series Euclidean distance Normalized Euclidean distance Dynamic Time Warping (DTW) Detecting patterns in time series Euclidean distance DTW Learning patterns from time series Distance-based pattern extraction Dictionary-based pattern extraction
Session 2: Feature Extraction and Selection Problem statement Feature extraction Spectral features Statistical features Deep Learning features Other features Feature selection Unsupervised setting Supervised setting
Session 3: Models and Representation Learning Problem statement Standard models Signal+Noise model Sinusoidal model Trend+Seasonality model AR, MA and ARMA models Hidden Markov model Representation learning Standard representations Notion of sparsity Sparse coding Dictionary learning	N/A
Session 4: Denoising, Detrending and Interpolation Problem statement Denoising Outlier removal Detrending Handling missing data	N/A
Session 5: Change-point and Anomaly Detection Problem statement Change point detection Optimal resolution Approximate methods Supervised approaches Anomaly detection	N/A
Session 6: Multivariate and multimodal time series Problem statement Tensor-based approaches Low-rank approximations Graph-based approaches Graph Fourier Transform (GFT)	N/A

Additional ressources

Useful references
List of possible topics/projects