TL;DR
This paper introduces mlCAF, a framework that fuses multi-level, cross-domain contextual information (physical activity, nutrition, and clinical data) to create a richer understanding of human behavior and lifestyle. By extending existing context-awareness models with semantic reasoning and fusion techniques, mlCAF aims to improve personalized health and wellness services, achieving a notable 92.65% accuracy in context recognition.
What Problem Does This Paper Address?
In ubiquitous and pervasive computing, accurately identifying user context and behavior is crucial for developing effective applications, especially in health and wellness. While many systems focus on single domains (like physical activity), real-world human behavior is influenced by a complex interplay of factors across multiple domains – physical activity, nutrition, and clinical data. The challenge lies in effectively integrating and fusing this heterogeneous, multi-level information to derive abstract, meaningful insights. Existing approaches often struggle with this cross-domain fusion, leading to a limited understanding of user behavior and lifestyle, which is particularly problematic for managing chronic diseases like diabetes.
Methods
The core of this work is the mlCAF (Multi-Level Context-aware Framework). It builds upon the Mining Minds Context Ontology (MMCO), which was extended to include nutrition and clinical domains alongside physical activity.
Key methodological components include:
- Ontology Evolution: The MMCO was expanded from modeling only physical activity to incorporating 57 food items and clinical data (blood glucose and blood pressure levels). This created a more comprehensive semantic model.
- Vertical Fusioning (VF): This process infers High-Level Contexts (HLCs) from Low-Level Contexts (LLCs) within a single domain. For example, inferring "Fats" (N-HLC) from "eating" (Activity LLC) and "restaurant" (Location LLC). This is achieved using OWL reasoning.
- Horizontal Fusioning (HF): This technique fuses vertically fused contexts across different domains and over time to infer more abstract behavioral contexts and lifestyle patterns. It utilizes SWRL (Semantic Web Rule Language) and SQWRL (Semantic Query-Enhanced Web Rule Language) for rule-based reasoning and querying.
- Semantic Reasoning: The framework leverages OWL-DL reasoning (using the Pellet reasoner) for inferring relationships and classifications within the ontology, and SWRL/SQWRL for more complex rule-based inference and behavioral pattern identification.
Datasets
The study used data collected from 20 adult volunteers in a real-world environment. This dataset comprised approximately 141,872 Low-Level Contexts and 36,276 recognized High-Level Contexts. The paper does not introduce a new benchmark dataset but rather uses this collected data for evaluation.
Key Insights and Ideas
- The Power of Cross-Domain Fusion: The most significant insight is that combining data from physical activity, nutrition, and clinical measurements provides a much richer and more accurate picture of human behavior than any single domain alone. This is crucial for understanding complex health-related behaviors.
- Ontologies as a Unifying Framework: Using ontologies (like MMCO) provides a structured, semantic way to represent and integrate diverse contextual information, enabling sophisticated reasoning and inference.
- Hierarchical Context Fusion: The distinction between Vertical Fusioning (within-domain) and Horizontal Fusioning (cross-domain, temporal) effectively breaks down the complex task of context understanding into manageable, hierarchical steps.
- Behavioral Patterns from Temporal Data: Horizontal fusion, by considering contexts over time, allows for the identification of more complex behavioral patterns and lifestyle trends, moving beyond immediate situational awareness.
Key Results
- High Context Recognition Accuracy: The mlCAF framework achieved an average context recognition accuracy of 92.65% for the collected cross-domain contexts.
- Reduced Data Volume: Vertical fusioning effectively reduces the volume of low-level context data that needs to be communicated, improving efficiency. For instance, out of 36,276 HLCs, only 4144 were communicated for behavior modeling.
- Effective Behavioral Identification: The framework successfully identified various behavioral contexts, such as "Sedentary Behavior," "Lightly Active," "Moderately Active," and "Very Active," as well as "Meal Frequency."
Limitations and Future Work
- Missing LLCs Impact Accuracy: The paper notes that while HLCs are rarely misclassified, missing Low-Level Contexts (LLCs) can lead to declined accuracies for High-Level Contexts, particularly for nutrition-related HLCs.
- Data Requirements: The authors acknowledge that there is still a need for more relevant data within specific timeframes for more accurate PA-HLC, N-HLC, and C-HLC context analysis.
- Future Work:
- Extend SWRL/SQWRL for lifestyle modeling and prediction analysis using data-driven approaches.
- Apply inductive reasoning on gathered datasets (A-Box assertions).
- Further investigate more efficient context fusioning mechanisms through semantic technologies.
Summary updated