Learning Knowledge Sharing Schemes for Time Series Modeling
Open AccessIn IoT mobile sensing world, with the rapidly growing volume of Internet-connected sensory devices, the IoT generates massive data characterized by its velocity in terms of spatial and temporal dependency. Analyzing these mobile time series data appropriately can bring considerable socio-economic benefits such as target-advertising based on accurate prediction of cellular traffic data, real-time health status monitoring, anomaly detection for core infrastructures security and early warning based on the forecasting of heartbeat data from kinds of sensors, and underground passenger traffic prediction to avoid peak travel risks, and so on. In this work, we focus on how to further improve the generalization performance for machine learning based time series modeling approaches as a copious amount of IoT sensors in many real-life scenarios consecutively generate substantial volumes of time series data. In Machine Learning (ML), we normally focus on optimizing for a specific metric whether it is a score on a specific benchmark or a business KPI. To accomplish this, we usually train a single model or a group of models to perform the task at hand. The models are then fine-tuned and tweaked until their performance plateaus. While we can generally attain satisfactory results in this manner, we miss out on information that could help us improve our performance on the statistics we give concern. This information is derived specifically from the training signals of related tasks. We can improve the generalization of our model on our initial task by sharing representations between similar tasks. This method is also known as Multi-Task Learning (MTL), and the focus of this thesis will be on how to use the MTL paradigm to improve time series modeling procedures. Considering the existing spatial information of entities in many real-world applications, Graph Convolutional Networks (GCNs) have recently revealed discriminative power in learning graph representations due to their permutation-invariance, local connectivity, and compositionality. Graph neural networks allow each graph node to acknowledge its neighborhood context by propagating information through structures. We then present a graph-based framework for modeling geographical information in multivariate time series data. Moreover, other challenges such as privacy preservation and label data scarcity drive us devise an innovative federated learning-based framework in online learning fashion. A semi-supervised learning approach for time series modeling is also proposed in this work to overcome insufficient labeled data in many scenarios.
- All rights reserved
Notice to Authors
If you are the author of this work and you have any questions about the information on this page, please use the Contact form to get in touch with us.