Neuro-Symbolic Class-Contrast Evidence Audit for Reliable Cross-Subject Wearable Activity Recognition

Area of Science:

• * Computer Science
• * Machine Learning
• * Wearable Technology

Background:

• * Reliable wearable activity recognition necessitates not only accurate class labels but also verifiable evidence supporting these labels.
• * Existing methods often lack auditable support, limiting trust and interpretability in cross-subject scenarios.

Purpose of the Study:

• * To introduce CC-NSIEA (Class-Contrast Neuro-Symbolic Inference Evidence Audit), a novel framework for label-preserving, evidence-auditable cross-subject wearable activity recognition.
• * To enhance the reliability and trustworthiness of activity recognition systems by incorporating an explicit evidence auditing mechanism.

Main Methods:

• * A Temporal Residual Perception Network generates activity labels, posterior probabilities, and temporal embeddings.
• * A read-only Training-Subject Evidence Memory stores historical sensor data and predictions.
• * A rule-based Evidence Consistency Audit integrates data validity, motion coherence, retrieval support, and class separation, with optional Class-Contrast Evidence Refinement.

Main Results:

• * The CC-NSIEA model achieved 90.13% accuracy and 90.55% macro-F1 on subject-disjoint data from the UCI HAR dataset.
• * CC-NSIEA demonstrated improvements in reliability metrics, increasing Error AUPRC from 0.423802 to 0.433057 and reducing AURC from 0.035941 to 0.035913 compared to a deterministic controller.
• * Statistical analysis confirmed a significant improvement in AUPRC (bootstrap interval [0.001595, 0.019547]).

Conclusions:

• * CC-NSIEA offers a robust, evidence-centered approach to wearable activity recognition, complementing confidence-based reliability estimation.
• * The neuro-symbolic architecture, combining neural prediction with rule-based auditing, enhances the transparency and auditability of recognized activities.
• * This method provides a significant advancement for cross-subject wearable activity recognition, particularly in applications demanding high reliability and verifiable results.