Uncertainty Quantification for Deep Context-Aware Mobile Activity Recognition and Unknown Context Discovery

Zepeng Huo, Arash Pakbin, Xiaohan Chen, Nathan C. Hurley, Ye Yuan, Xiaoning Qian, Zhangyang Wang, Shuai Huang, Bobak Mortazavi

Abstract

Human Activity Recognition (HAR) using mobile sensors is often hindered by environmental context, where the same activity produces different signal signatures depending on the surroundings. We present a novel α-β framework that addresses this by first identifying the environmental context and then applying specialized activity recognition models. Furthermore, by integrating Uncertainty Quantification (UQ) through Maximum Entropy Learning (MEL), the system can detect and reject unknown contexts, preventing overconfident erroneous predictions. Experimental results show a 10% improvement in accuracy and significantly higher sensitivity in discovering novel environments.

Introduction

Human activity recognition (HAR) is a cornerstone of mobile health and context-aware computing. However, standard deep learning models often struggle with context dependency—the phenomenon where sensor data for an activity like "walking" varies significantly between environments (e.g., walking on a treadmill vs. walking in a moving bus). Additionally, real-world deployments frequently encounter "unknown" contexts not present in training data. Standard models tend to be overconfident, forcing labels onto these novel situations. This paper proposes an integrative framework to solve these issues through context-specific modeling and robust uncertainty quantification.

The α-β Framework

The proposed architecture consists of two primary components designed to decouple environmental sensing from activity classification.

The integrated α-β framework architecture.
  1. The α Network (Context Detector): This network serves as a gating mechanism. It analyzes incoming sensor data to determine the probability distribution of the current environmental context (e.g., "in-car," "on-bicycle," or "shopping").
  2. The β Networks (Activity Experts): The system maintains a library of dedicated β networks, each specialized for a specific known context. Once the α network identifies the context, the corresponding β network (or a weighted combination) predicts the specific activity, such as sitting, standing, or walking.

To prevent "model collapse," where the system defaults to a single β network, the authors utilize K-means clustering on CNN-embedded features to initialize the context discovery process during pre-training. This initialization improved accuracy from 87% to 96% on benchmark datasets.

Uncertainty Quantification and Unknown Context Discovery

A critical feature of this framework is its ability to handle "unknown" contexts. This is achieved through Maximum Entropy Learning (MEL). Unlike standard training that seeks a single set of optimal parameters, MEL learns a distribution of parameters, allowing the model to quantify its own uncertainty.

When the model encounters sensor data from an environment it does not recognize, the calculated uncertainty exceeds a predefined threshold. This triggers a "rejection" option, signaling that the system has encountered a novel context. This mechanism is vital for real-world reliability, as it allows the system to flag data for future learning rather than providing a false activity label.

Experimental Evaluation and Key Results

The framework was evaluated using the UCI OPPORTUNITY benchmark and a new, realistic in-house dataset featuring 32 activities across various noisy environments, such as "walking with one shoe off" to simulate limping.

Probability distribution of known vs. unknown contexts.

Key findings include:

  • Accuracy Gains: The α-β framework achieved an approximately 10% improvement in accuracy and F-scores over standard baseline models.
  • Robust Context Discovery: In "leave-one-out" testing, the UQ-enabled model demonstrated high sensitivity (0.62–0.75) in identifying unseen contexts, far outperforming standard probabilistic baselines (0.17–0.34).
  • Stability: MEL provided a more robust way to quantify uncertainty than traditional softmax probabilities, which are often prone to overconfidence.

Conclusion

The α-β framework provides a robust solution for context-aware mobile activity recognition. By explicitly modeling environmental context and quantifying uncertainty, the system achieves higher precision and the ability to safely navigate unknown real-world scenarios. Future work will focus on defining more flexible context ontologies and optimizing the computational efficiency of the MEL-based parameter distributions for real-time mobile deployment.

Summary updated

BibTeX
@inproceedings{huo2020uncertainty,
  author = {Zepeng Huo and Arash Pakbin and Xiaohan Chen and Nathan C. Hurley and Ye Yuan and Xiaoning Qian and Zhangyang Wang and Shuai Huang and Bobak Mortazavi},
  title = {{Uncertainty Quantification for Deep Context-Aware Mobile Activity Recognition and Unknown Context Discovery}},
  booktitle = {International Conference on Artificial Intelligence and Statistics},
  year = {2020},
  url = {https://openreview.net/forum?id=eaE9tpVcvnS},
}
Figure