Federated Multi-task Learning with Hierarchical Attention for Sensor Data Analytics
- Resource Type
- Conference
- Authors
- Chen, Yujing; Ning, Yue; Chai, Zheng; Rangwala, Huzefa
- Source
- 2020 International Joint Conference on Neural Networks (IJCNN) Neural Networks (IJCNN), 2020 International Joint Conference on. :1-8 Jul, 2020
- Subject
- Bioengineering
Computing and Processing
General Topics for Engineers
Power, Energy and Industry Applications
Robotics and Control Systems
Signal Processing and Analysis
Task analysis
Data models
Performance evaluation
Feature extraction
Correlation
Data analysis
Predictive models
Sensor analytics
Attention mechanism
Multitask learning
- Language
- ISSN
- 2161-4407
The past decade has been marked by the rapid emergence and proliferation of a myriad of small devices, such as smartphones and wearables. There is a critical need for analysis of multivariate temporal data obtained from sensors on these devices. Given the heterogeneity of sensor data, individual devices may not have sufficient quality data to learn an effective model. Factors such as skewed/varied data distributions bring more difficulties to the sensor data analytics. In this paper, we propose to leverage multi-task learning with attention mechanism to perform inductive knowledge transfer among related devices and improve generalization performance. We design a novel federated multi-task hierarchical attention model (FATHOM) that jointly trains classification/regression models from multiple distributed devices. The attention mechanism in the proposed model seeks to extract feature representations from inputs and to learn a shared representation across multiple devices to identify key features at each time step. The underlying temporal and nonlinear relationships are modeled using a combination of attention mechanism and long short-term memory (LSTM) networks. The proposed method outperforms a wide range of competitive baselines in both classification and regression settings on three unbalanced real-world datasets. It also allows for the visual characterization of key features learned at the input task level and the global temporal level.