核心概念
A novel framework called Motion Code that jointly learns across different collections of noisy time series data by explicitly modeling the underlying stochastic processes and introducing the concept of the most informative timestamps.
摘要
The paper introduces a novel framework called Motion Code for robust time series classification and forecasting. The key ideas are:
-
Modeling each time series as an instance of an underlying stochastic process, which allows capturing the data dependence across timestamps and separating noise from true signals.
-
Handling multiple distinct dynamics in the data by assigning a signature vector (motion code) to each dynamical model and optimizing them jointly.
-
Introducing the concept of the most informative timestamps, which are a small subset of timestamps that minimizes the mismatch between the original data and the information reconstructed using only this subset. This allows an interpretable feature for visualizing the core dynamics.
-
The final Motion Code model can jointly learn across multiple time series collections, enabling simultaneous classification and forecasting without the need for separate models.
Extensive experiments on noisy sensor and device time series data demonstrate Motion Code's competitiveness against state-of-the-art time series classification and forecasting benchmarks, as well as its ability to handle uneven length and missing data.
統計資料
The data consists of collections of time series, where each collection corresponds to a particular underlying stochastic process.
The time series within each collection have varying lengths between 80 and 95 data points.
Gaussian noise with standard deviation 0.3 times the maximum absolute value is added to the original data to create a more challenging setting.
引述
"Unlike images, videos, text, or tabular data, finding a suitable mathematical concept to represent and study time series is a complicated task."
"To handle such settings, we first assign each dynamics a signature vector. We then propose the abstract concept of the most informative timestamps to infer a sparse approximation of the individual dynamics based on their assigned vectors."
"The final model, referred to as Motion Code, contains parameters that can fully capture different underlying dynamics in an integrated manner. This allows unmixing classification and generation of specific sub-type forecasting simultaneously."