An Improved Self-Training Method for Positive Unlabeled Time Series Classification Using DTW Barycenter Averaging

Area of Science:

• Machine Learning
• Data Mining
• Time Series Analysis

Background:

• Traditional supervised time series classification (TSC) requires extensive labeled data, which is often impractical.
• Labeling large unlabeled datasets is time-consuming and requires domain expertise.
• Positive unlabeled time series classification (PUTSC) addresses this by labeling unlabeled data using a small labeled set.

Purpose of the Study:

• To address the limitations of existing self-training (ST) methods in PUTSC, particularly their sensitivity to initial labeled data.
• To propose a novel ST-based methodology, ST-average, for more robust and reliable PUTSC.
• To demonstrate the effectiveness of ST-average compared to existing popular methods.

Main Methods:

• The study proposes ST-average, a novel approach for PUTSC.
• This method utilizes a representative average sequence, computed via DTW barycenter averaging, for labeling unlabeled data.
• ST-average is designed to be insensitive to the initial labeled data and compatible with existing ST techniques.

Main Results:

• Experimental results on public datasets demonstrate that ST-average outperforms existing popular PUTSC methods.
• The proposed ST-average method shows improved reliability and robustness compared to traditional one-nearest-neighbor (1NN) based ST approaches.
• The average sequence used in ST-average is more representative than individual labeled sequences.

Conclusions:

• ST-average offers a more reliable and robust solution for positive unlabeled time series classification.
• The method overcomes the sensitivity issues associated with initial labeling in conventional ST approaches.
• ST-average provides a valuable advancement for efficiently classifying large unlabeled time series datasets.