HODMD for Identification of Spatio-Temporal Patterns in Covid19
Standard DMD
Standard DMD
Standard DMD Reinterpreted: DMD-1 Algorithm
Step 1: Dimensionality Reduction
Rescaled Temporal Modes
Reduced Snapshot Matrix
Standard DMD Reinterpreted: DMD-1 Algorithm
Step 1: Dimensionality Reduction
Step 2: Compute DMD Modes and Reduced Koopman Matrix
Standard DMD Reinterpreted: DMD-1 Algorithm
Step 1: Dimensionality Reduction
Step 2: Compute DMD Modes and Reduced Koopman Matrix
Standard DMD Reinterpreted: DMD-1 Algorithm
We Need to Ask the Question, When will the above assumption hold??
Intuitively, There should be Consistency with Spatial Resolution and Temporal Resolution.
This has Direct Effect on the data being "Linear"...
A Balance Between Spatial And Temporal Resolutions should be reached
Standard DMD Reinterpreted: DMD-1 Algorithm
We Need to Ask the Question, When will the above assumption hold??
Should There be a Constraint on Data being Considered? Such that Linearity is Fullfilled?
Standard DMD Reinterpreted: DMD-1 Algorithm
Should There be a Constraint on Data being Considered? Such that Linearity is Fullfilled?
Temporal Dimention
Spatial Dimention
Consider One Temporal Dimention
There are 2 points that can be connected by a line
Extending the Argument to higher Dimensions, In general, we might get,
Standard DMD Reinterpreted: DMD-1 Algorithm
Linear Consistency
Temporal Dimention
Spatial Dimention
Linearity
*Tu, J. H., Rowley, C. W., Luchtenburg, D. M., Brunton, S. L. & Kutz, J. N. On dynamic mode decomposition: Theory and applications. J. Comput. Dyn. 1, 391–421
"nonlinear data is inconsistent and inconsistent data is nonlinear"
Standard DMD Reinterpreted: DMD-1 Algorithm
Compatiblity Condition (Kim et al.)
Kim et al. : Kim, S., Kim, M., Lee, S. et al. Discovering spatiotemporal patterns of COVID-19 pandemic in South Korea. Sci Rep 11, 24470 (2021). https://doi.org/10.1038/s41598-021-03487-2
"The compatibility condition implies that the linearity of the data T is almost always guaranteed in case m≤n, which then leads to meaningful DMD results. for m>n, T will be in general inconsistent unless it is linear. As such, the direct and reliable DMD analysis of large time series data is not feasible in general." (Kim et al.)
Standard DMD Reinterpreted: DMD-1 Algorithm
Compatiblity Condition (Kim et al.)
Kim et al. : Kim, S., Kim, M., Lee, S. et al. Discovering spatiotemporal patterns of COVID-19 pandemic in South Korea. Sci Rep 11, 24470 (2021). https://doi.org/10.1038/s41598-021-03487-2
"The compatibility condition implies that the linearity of the data T is almost always guaranteed in case m≤n, which then leads to meaningful DMD results. for m>n, T will be in general inconsistent unless it is linear. As such, the direct and reliable DMD analysis of large time series data is not feasible in general." (Kim et al.)
Pitfalls of the DMD-1 Algorithm
Kim et al. : Kim, S., Kim, M., Lee, S. et al. Discovering spatiotemporal patterns of COVID-19 pandemic in South Korea. Sci Rep 11, 24470 (2021). https://doi.org/10.1038/s41598-021-03487-2
Compatable window DMD
Kim et al. : Kim, S., Kim, M., Lee, S. et al. Discovering spatiotemporal patterns of COVID-19 pandemic in South Korea. Sci Rep 11, 24470 (2021). https://doi.org/10.1038/s41598-021-03487-2
CwDMD chooses an adequate set of representative subdomains called windows, each containing a moderate size of time-series data that satisfies the compatibility condition
Kim et. al,
13.6 weeks
10.6 weeks
12.6 weeks
13.3 weeks
K
J = 17
Apply DMD to a subset of dataset which is compatable
Compatable window DMD
Kim et al. : Kim, S., Kim, M., Lee, S. et al. Discovering spatiotemporal patterns of COVID-19 pandemic in South Korea. Sci Rep 11, 24470 (2021). https://doi.org/10.1038/s41598-021-03487-2
CwDMD chooses an adequate set of representative subdomains called windows, each containing a moderate size of time-series data that satisfies the compatibility condition
Kim et. al,
13.6 weeks
10.6 weeks
12.6 weeks
13.3 weeks
K
J = 17
HODMD Motivation
Kim et al. : Kim, S., Kim, M., Lee, S. et al. Discovering spatiotemporal patterns of COVID-19 pandemic in South Korea. Sci Rep 11, 24470 (2021). https://doi.org/10.1038/s41598-021-03487-2
13.6 weeks
10.6 weeks
12.6 weeks
13.3 weeks
K
J = 17
Developments: On the Effect of Parameter d
Effect of Parameter d
Effect of Parameter d
For a prediction window of 20 days and train days of 40 days, d=20 gave best results
Effect of Parameter d
For a prediction window of 20 days and train days of 40 days, d=20 gave best results
Further, the literature says that parameter d should be scaled with the same scaling factor that scales the temporal or spatial resolution. [1]
For a prediction window of 20 days and train days of 40 days, d=20 gave best results
Effect of Parameter d
For a prediction window of 20 days and train days of 40 days, d=20 gave best results
The code has been refactored for reproducible research, with increased quality of code.
Rigorous Test suite is written to ensure good quality of code
Code is now modular
pip install -e ./src/
Install python package (Called hodmdExperiments for now)
from hodmdExperiments.experiment import Experiment
exp = Experiment(covid_data)
exp.fit()
Our experiments hence can be easily reproduced
Developments: Software Engineering perspective: On Enabling Reproducable research
Effect of Parameters
Interesting Metric:
We should aim to maximize the above metric
"Get maximum pred_days with minimum train_days"
For a given d, minimizing the below metric will aid our desired results
Effect of Parameters
Possible Directions.....
Bayesian Optimization for finding parameters, based on the above metric??
Other Possible Directions...
Other Possible Directions...
0.11753837-0.82817903j
0.62842673+0.54379097j
Faster growth Rate than China
-0.39535637 -0.70067206j
Text
Why is Re(Eig) negative?????
China
India
BNG
AFG
UK
Detailed Analysis need to be done