View source for Graph Neural Networks for Multivariate Time Series Data Analysis

{{StudentProjectTemplate
|Summary=Our project aims to analyze multivariate time series through the innovative application of graph neural networks.
|Keywords=multivariate time series, graph neural network, data imputation, time series prediction.
|TimeFrame=Fall 2024
|References=[1] Feng, Y., You, H., Zhang, Z., Ji, R., & Gao, Y. (2019, July). Hypergraph neural networks. In Proceedings of the AAAI conference on artificial intelligence (Vol. 33, No. 01, pp. 3558-3565).
[2] Zhang, C., Song, D., Huang, C., Swami, A., & Chawla, N. V. (2019, July). Heterogeneous graph neural network. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining (pp. 793-803).
|Supervisor=Guojun Liang, Prayag Tiwari,
|Level=Master
|Status=Open
}}
Our project aims to analyze multivariate time series through the innovative application of graph neural networks (GNN). By using this approach, we intend to uncover latent structures within time series data, which enable us to address challenges such as casual discovery, missing data imputation and accurate future prediction. Recently, there are two innovative GNNs: hyper GNN and Heterogeneous GNN, which can deal with more graph structure dataset, especially the spatio-temporal dataset.

Time series data is ubiquitous in practical scenarios, ranging from annual housing price trends to daily traffic flow patterns. However, traditional approaches to multivariate time series imputation have often focused solely on temporal patterns, disregarding the valuable spatial structures inherent in the data. Notably, the spatial structure does not simply represent physical (e.g., geographic) proximity, but rather indicates that considered sensors are related w.r.t. a generic (quantifiable) functional dependency.  As a result, mining the spatio-temporal relationship is an important key to the subsequent task. Several topics can be selected for the thesis:   
1)    Time Series Prediction  Accurate time series prediction is crucial for many applications like neuroscience, transportation, and healthcare. Exploring appropriate spatial and temporal characteristics is the key to achieving high performance. For instance, transportation datasets come with pre-defined graph structures. This topic involves employing graph neural networks and other pertinent machine learning algorithms to forecast future values based on historical data.    
2) Time Series Imputation: Incomplete data due to sensor reliability issues or network malfunctions is a common challenge in time series data collection. Consequently, effective time series imputation is pivotal within multivariate time series analysis. In this topic, time series imputation is to find out the latent structure of time series and impute the missing data by graph neural network or other relevant machine learning algorithms.    
3)   Causal discovery of time series Having insight into the causal associations in a complex system facilitates decision-making, e.g., for medical treatments, urban infrastructure improvements or financial investments. The amount of observational data grows, which enables the discovery of causal relationships between variables from observation of their behavior over time. Existing methods for causal discovery from time series data do not yet exploit the representational power of deep learning. In this topic, the aim of causal discovery is to find out the causal relationship between different nodes, by utilizing the power of graph neural networks and other machine learning techniques with time series data analysis.