Less is more: Efficient behavioral context recognition using Dissimilarity-Based Query Strategy.

Area of Science:

• Ubiquitous Computing
• Machine Learning
• Human-Computer Interaction

Background:

• Smartphone sensors generate vast unlabeled data streams, offering potential for behavioral context recognition.
• Accurate context recognition is crucial for applications in disease prevention and independent living.
• Label acquisition for sensor data remains a significant challenge due to user dependency.

Purpose of the Study:

• To propose a novel approach, Dissimilarity-Based Query Strategy (DBQS), for behavioral context recognition using smartphone sensor data.
• To address the challenge of label acquisition by leveraging active learning for selective sampling.
• To improve the efficiency and accuracy of training models for context recognition in natural environments.

Main Methods:

• The proposed Dissimilarity-Based Query Strategy (DBQS) utilizes active learning for selective sampling of informative and diverse sensor data.
• DBQS overcomes model stagnation by prioritizing new and distinct samples not previously explored.
• Temporal information within the data is exploited to further enhance dataset diversity.

Main Results:

• The DBQS approach demonstrated improved overall average Balanced Accuracy (BA) by 6% on a public dataset.
• The method achieved this improvement with a 13% reduction in the required training data.
• Experimentation confirmed the effectiveness of the approach in natural environment settings.

Conclusions:

• The Dissimilarity-Based Query Strategy (DBQS) offers an effective solution for behavioral context recognition from unlabeled sensor data.
• Leveraging active learning and temporal information enhances model training efficiency and accuracy.
• This approach holds significant promise for applications requiring reliable context-aware systems.