Beyond Information Distortion: Imaging Variable-Length Time Series Data for Classification.

Area of Science:

• Machine Learning
• Data Science
• Computer Vision

Background:

• Time series data are common in manufacturing and human activity recognition.
• Variable sample lengths in time series pose challenges for standard classification models.
• Current methods for handling variable lengths can degrade data integrity and model performance.

Purpose of the Study:

• To introduce a novel method, Time series Into Pixels (TIP), for effectively classifying variable-length time series data.
• To address the limitations of existing approaches that may compromise data integrity.
• To provide a robust and accurate solution for real-world time series classification.

Main Methods:

• Proposed Time series Into Pixels (TIP) method: mapping time series data points to pixels in a 2D representation.
• Utilized a LeNet-like 2D Convolutional Neural Network (CNN) for evaluating the TIP representation.
• Conducted extensive evaluations across 11 real-world benchmarks against 10 baseline models.

Main Results:

• TIP achieved 2-5% higher accuracy compared to baseline models.
• TIP demonstrated 10-25% higher macro average precision.
• TIP showed comparable performance on complex multivariate data and highlighted potential issues with length normalization.

Conclusions:

• The TIP method offers a significant advancement for handling variable-length time series data.
• TIP provides a robust and accurate approach to time series classification without compromising data integrity.
• The proposed method is effective across diverse real-world applications, with code publicly available.