[20240708_LabSeminar_Huy]Covid19Dynamics.pptx

Forecasting COVID-19 Dynamics: Clustering, Generalized Spatiotemporal Attention, and Impacts of Mobility and Geographic Proximity

1. Quang-Huy Tran Network Science Lab Dept. of Artificial Intelligence The Catholic University of Korea E-mail: [email protected] 2024-07- 08 Forecasting COVID-19 Dynamics: Clustering, Generalized Spatiotemporal Attention, and Impacts of Mobility and Geographic Proximity Tong Shen et al. ICDE- 202 3 : The IEEE 39th International Conference on Data Engineering

2. OUTLINE MOTIVATION METHODOLOGY EXPERIMENT & RESULT CONCLUSION

3. MOTIVATION The COVID-19 pandemic has been dramatically impacting people’s lives across the globe in the last two years . Negative effect to economy. Hence, understand the dynamics of the pandemic and forecast its future trends is crucial to allow government agencies and policymakers to adopt proper restriction strategies. Plan and prepare necessary resources for combating foreseeable outbreaks. Contain the virus while minimizing the adverse effects on the economy. Overview and Challenges

4. MOTIVATION Challenges: D ynamics of COVID-19 manifest extremely complex spatial and temporal dependencies. temporal misalignment: a similar trend at different times in two regions. M any exogenous factors dramatically influence the dynamics of COVID-19 . Population mobility , geographically adjacent areas, national and local health advisories and restrictions (mask, vaccine coverage, quarantine policies, etc ) Overview and Challenges S electing a proper granularity is critical to the success of forecasting COVID-19 dynamics. State-level dynamics are relatively stable and thus less difficult to predict. P redicting county-level dynamics pays attention to the peculiarities of individual regions. Small population is generally vulnerable to noise in prediction.

5. INTRODUCTION P ropose generalized spatiotemporal attention : A ddress temporal misalignment property of COVID-19 dynamics enables learning between the epidemic dynamics at different regions and times. flexibly accounts for any intra- and inter- region information. Contribution P ropose a new algorithm that clusters neighboring counties into county groups: Based on geographic and population information . P rovide an appropriate, fine granularity (i.e., county group) for analyzing and forecasting the dynamics of COVID-19 without suffering much from the data noise. P ropose a new deep learning model COVID-Forecaster : C apture complex temporal and spatial dependencies of the COVID-19 dynamics. C onsider the impact of mobility, geographic proximity, and other helpful auxiliary information.

6. METHODOLOGY Clustering and Problem Definition Clustering : Set a population threshold to determine the number of county groups within a state. For county in state, map to feature space   Normalized latitude, longitude, and population of .   Apply agglomerative hierarchical clustering algorithm to cluster these counties into the target number of county groups inside the state.

7. METHODOLOGY Clustering and Problem Definition Given dataset: COVID-19 diagnosis case at location and date as . Weekly mobility data between county and county at week as . A sparse directed weighted graph , is county set and represents at time t. Auxiliary Data (ICU bed capacity, mask usage, mobility level, etc ) at location and date as .   Problem: Predict COVID-19 dynamics in the next t days.

8. METHODOLOGY Main Architecture

9. METHODOLOGY Data Preprocessing and Temporal Encoding Module To consistent , use to represent , a concatenation of variable matrix X from time step t − k to time step t − 1 . Mobility data from region i to region j at timestep t.   Preprocessing: map the input signal to embedding space for further processing COVID19 diagnosis data : 1DConv . Auxiliary data: a multiple-layer perceptron Merge the embedding of auxiliary data and virus-spreading data with another set of 1D convolution layers:

10. METHODOLOGY Temporal Attention Encoder Intra-region aggregation : Multi-head self-attention of Transformer for every single location from .   Int er -region aggregation : With L regions * T time steps, apply multi-head self-attention to time step t embedding as query and flatten all regions’ temporal embeddings as key and value with linear transformation for efficiency. Concatenate the embeddings from intra-region and inter-region:

11. METHODOLOGY Spatial Attention Decoder Positional Encoding with Geo-location: Geographic proximity: Regions geographically close to each other usually have a similar phase in the epidemiological process. Treat all the locations in a 2D M × M grid with latitude and longitude of center position. Adopt a relative positional bias to represent the geographic proximity between grids instead.   After mapping all the historical information to embedding space : Decoder digests the embeddings and predicts the future trend. Design decoder that considers the population mobility and geographic information of regions for better prediction.

12. METHODOLOGY Spatial Attention Decoder where means all neighbors of in the mobility graph.   where is normalized column-wise adjacency matrix of the mobility network in time t-1 and is Hadamard product .   Spatial Attention with Mobility Graph : Encoder perform well in short-term prediction but hard to forecast well in a longer horizon. To address, infer the regional dependencies by assuming the disease transmission rates are correlated with population mobility. A causal model to predict the future trend of one region conditioned on its own mobility level and all other regions that have mobility connection: Apply self-attention module to handle the interaction between mobility connect regions MLP in concatenation of embedding of current auxiliary data and final embedding of attention  

13. METHODOLOGY Spatial-Temporal Graph Learning Loss Function : Where H is the length of the forecasting horizon and is a weight factor to balance two loss terms.  

14. EXPERIMENT AND RESULT EXPERIMENT SETTINGs Dataset: COVID-19 data and mobility data in US. Baselines: Spatial-Temporal GNN: STGCN [1], A3TGCN [2], GConvGRU [3], ACTS[4], and MPNNLSTM[5]. Deep Learning: GNND[6] and SpaceTimeFormer [7] . [1] Yu, B., Yin, H., & Zhu, Z. (2017). Spatio -temporal graph convolutional networks: A deep learning framework for traffic forecasting. arXiv preprint arXiv:1709.04875. [2] Bai, J., Zhu, J., Song, Y., Zhao, L., Hou, Z., Du, R., & Li, H. (2021). A3t-gcn: Attention temporal graph convolutional network for traffic forecasting. ISPRS International Journal of Geo-Information, 10(7), 485. [3] Seo, Y., Defferrard , M., Vandergheynst , P., & Bresson, X. (2018). Structured sequence modeling with graph convolutional recurrent networks. In Neural Information Processing: 25th International Conference, ICONIP 2018, Siem Reap, Cambodia, December 13-16, 2018, Proceedings, Part I 25 (pp. 362-373). Springer International Publishing. [4] Jin, X., Wang, Y. X., & Yan, X. (2021). Inter-series attention model for COVID-19 forecasting. In Proceedings of the 2021 SIAM International Conference on Data Mining (SDM) (pp. 495-503). Society for Industrial and Applied Mathematics. [5] Panagopoulos, G., Nikolentzos , G., & Vazirgiannis , M. (2021, May). Transfer graph neural networks for pandemic forecasting. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 35, No. 6, pp. 4838-4845). [6] Sesti , N., Garau -Luis, J. J., Crawley, E., & Cameron, B. (2021). Integrating LSTMS and GNNS for covid-19 forecasting. arXiv preprint arXiv:2108.10052. [7] Grigsby, J., Wang, Z., Nguyen, N., & Qi, Y. (2021). Long-range transformers for dynamic spatiotemporal forecasting. arXiv preprint arXiv:2109.12218. Measurement : Mean Absolute Percentage Error (MAPE) and the Weight Absolute Percentage Error (WAPE)

15. EXPERIMENT AND RESULT RESULT – Overall Perfor mance County-group level(clustering algorithm)

16. EXPERIMENT AND RESULT RESULT – Overall Perfor mance County level (no clustering algorithm)

17. EXPERIMENT AND RESULT RESULT – Visualization

18. CONCLUSION To achieve good forecasting accuracy in epidemic Models capture the complex spatial and temporal dependencies (solving the temporal misalignment problem) . consider the impacts of relevant intra- and inter region information (population mobility and geographic proximity). Summarization Propose COVID-Forecaster for COVID-19 dynamics Employs generalized spatiotemporal attention: capture the temporal dependency between the epidemic dynamics of different regions at different times . provides a general framework to encode any intra- and inter-region information. Investigate prediction at what granularity respects the characteristics of various regions avoiding the effects of noise. provide a clustering algorithm to find a proper granularity.