More Reliable Neighborhood Contrastive Learning for Novel Class Discovery in Sensor-Based Human Activity Recognition.

Area of Science:

• Computer Science
• Machine Learning
• Sensor Data Analysis

Background:

• Human Activity Recognition (HAR) systems utilize sensor data for activity classification.
• Current HAR systems struggle with discovering novel activity classes in unlabeled data without supervision.
• This limitation hinders real-world applications where fully supervised settings are impractical.

Purpose of the Study:

• Introduce the Novel Class Discovery (NCD) problem for HAR.
• Enable classification of new activities from unlabeled sensor data using existing labeled data.
• Develop a robust framework for unsupervised activity discovery.

Main Methods:

• Propose an end-to-end framework: More Reliable Neighborhood Contrastive Learning (MRNCL).
• MRNCL is a lightweight variant of Neighborhood Contrastive Learning (NCL).
• Incorporate an effective similarity measure for identifying reliable k-nearest neighbors in the embedding space.

Main Results:

• MRNCL outperforms existing methods on the NCD task in sensor-based HAR.
• Demonstrated superior clustering performance for new activity class instances across three public datasets.
• The proposed model effectively utilizes labeled data to aid in the discovery of novel activities.

Conclusions:

• MRNCL provides a significant advancement in addressing the Novel Class Discovery problem for HAR.
• The framework's efficiency and improved neighbor identification enhance unsupervised learning capabilities.
• This research paves the way for more adaptive and practical HAR systems in real-world scenarios.