DeepContext: Parameterized Compatibility-Based Attention CNN for Human Context Recognition

The ubiquity of sensor-rich smartphones has increased interest in mobile context-aware sensing applications in domains such as ambient assisted living, remote health care, and sports injury detection. Recognizing the user's current context by analyzing their smartphone's sensor data is a critical problem for such applications. One of the major technical challenges for context recognition is reliable feature extraction due to coarse-grained labeling. In sensor data coarse-grained labeling, only certain parts of smartphone sensor data are truly representative of the assigned label, while their exact duration and location within the segment are unknown. To address this, we propose DeepContext, a deep learning based network architecture for recognizing a smartphone user's current context. DeepContext uses a Convolutional Neural Network (CNN) with parameterized compatibility-based attention to discover and focus on important parts of smartphone sensor data, mitigating coarse-grained weak labels and extracting salient discriminative features. DeepContext uses a joint-learning fusion strategy that utilizes both domain-specific handcrafted features and features that are autonomously generated by a Convolutional NeuralNetwork (CNN) . We demonstrate that DeepContext consistently outperforms prior state-of-the-art context recognition and human activity recognition deep learning models on smartphone context sensor data gathered from 100 participants by nearly 5% in Balanced Accuracy.