Multiobjective Learning in the Model Space for Time Series Classification.

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Accurate time series classification relies on well-defined distance metrics.
• Existing methods like dynamic time warping or generative models have limitations in exploiting dynamic and label information simultaneously.

Purpose of the Study:

• To propose a novel multiobjective learning algorithm, multiobjective model-metric (MOMM) learning, for enhanced time series approximation and classification.
• To simultaneously utilize dynamic time series information and label information for improved distance metric learning.

Main Methods:

• Employed a recurrent network as a temporal filter to learn generative models for each time series representation.
• Developed a non-Euclidean space where label information is used to learn a distance metric.
• Optimized network size for parsimonious representations and simultaneously optimized data representation, model separation, and network size.

Main Results:

• MOMM learning demonstrated superior overall performance in both univariate and multivariate time series classification tasks.
• The proposed method also showed promising results in time series prediction.

Conclusions:

• MOMM learning offers a powerful approach to time series classification by integrating dynamic features and label information.
• The algorithm's ability to optimize multiple objectives leads to enhanced performance in both classification and prediction.