Multipatch Augmentation Learning Based on Dual-Policy Model for Time Series Classification

Area of Science:

• Machine Learning
• Data Science
• Time Series Analysis

Background:

• Multiple instance learning (MIL) for time series classification (TSC) suffers from instability due to anomalous instances impacting bag-level predictions.
• Selecting optimal series transformations and patch lengths for improved instance discriminability and stability is a significant challenge.

Purpose of the Study:

• To develop a novel TSC algorithm that adaptively selects optimal configurations for transformations and patch lengths.
• To enhance the stability and accuracy of MIL-based TSC by mitigating the impact of anomalous instances.

Main Methods:

• A dual-policy model is proposed to adaptively select transformation combinations and patch lengths from a predefined space.
• The multipatch augmentation learning based on dual-policy model (MPAL-DPM) algorithm employs a two-stage training process with three functional modules.
• Key modules include a policy model for transformation selection, a scalable patch for length fine-tuning, and multipatch classification (MPC) for feature extraction and instance weighting.

Main Results:

• MPAL-DPM demonstrates performance comparable to state-of-the-art classifiers on the UCR archive.
• The algorithm shows effectiveness in achieving stable and accurate time series classification.
• Experimental results confirm the robustness and scalability of the proposed MPAL-DPM approach.

Conclusions:

• The MPAL-DPM algorithm effectively addresses the instability issues in MIL-based TSC.
• Adaptive selection of transformations and patch lengths leads to more consistent and reliable predictions.
• The proposed method offers a robust and scalable solution for time series classification tasks.