Time Series Segmentation Based on Stationarity Analysis to Improve New Samples Prediction

Area of Science:

• Data Science
• Machine Learning
• Time Series Analysis

Background:

• Time series data, prevalent in Internet of Things (IoT) applications, often suffers from unreliable readings, leading to inaccurate machine learning models.
• Low-quality training datasets hinder the performance of algorithms that rely on sequential data patterns.
• Identifying change points in time series is crucial for segmenting data and mitigating bias.

Purpose of the Study:

• To develop and evaluate novel methods for automatic time series segmentation.
• To improve the accuracy of machine learning models for prediction tasks using segmented time series data.
• To address challenges posed by noisy and unreliable data in IoT applications.

Main Methods:

• Two automatic time series segmentation frameworks were developed: Change Detector segmentation and ADF-based segmentation.
• The ADF-based segmentation utilizes a novel change detector derived from the Augmented Dickey-Fuller (ADF) test for data stationarity.
• The methods were tested on real-world IoT datasets and benchmarks.

Main Results:

• Both proposed segmentation methods demonstrated improvements in prediction tasks.
• The ADF-based segmentation, particularly when used with Long Short-Term Memory (LSTM) models, significantly reduced relative prediction error from 1 to 0.67.
• The segmentation techniques effectively handled noisy data, enhancing model reliability.

Conclusions:

• Automatic time series segmentation using ADF-based methods is effective in improving prediction accuracy for machine learning models.
• These techniques offer a robust solution for handling unreliable data in IoT and other sequential data applications.
• The proposed segmentation frameworks provide a valuable tool for enhancing the performance of time series analysis.