COVID-19 has impacted the social and economic activity of faculty and students at Purdue. The Protect Purdue initiative has implemented a range of measures that keeps classrooms safe. However, the risk outside of the classroom, in bars, restaurants, churches, gyms, grocery stores, etc. remains. Thus, the "social concentration" of people in places in our region creates concerns. An increase in foot traffic* at public places around Purdue, in Tippecanoe County, suggest the normalization of daily routines.
The task here is to build a statistical model using the historical movement, COVID-19 incident and policy intervention data collected from week 11 to week 43 to predict the foot traffic in week 44 for 1804 places of interest (POIs) in Tippecanoe County. Since the time series data given here is short (31 datapoints), using neural networks and attention mechanisms might overfit the data. Hence, I opted for simpler and faster approaches like ARIMA models. Of all the submissions I made, the one with the best performance is of an autoregressive model. That was better than using optimized Random Forest Regressors. This showed that low-level models are capable of handling short time series data better than complex black box models.
The report here shows various visualizations prepared from the given datasets. The objective here is to communicate my analyses and aid the policymakers in their decisions to curb the infection spread.
* foot traffic value here denotes the number of people visit a particular POI in a week.
Use slider to move between weeks. Press Play button to see full animation
Click on blue location icon to see the % change
Click on blue location icon to see the predicted foot traffic for week 44
These plots show the visit counts for top 50 busiest POIs for which the week 44 forecast showed high foot traffic. X axis shows week number and Y axis shows visit counts data. Historic data is shown as blue line. Week 44 forecast is shown as orange dot.
Most of these busiest POIs are either educational institutions or shopping centers. So, controlling foot traffic in those regions will help in curbing the infection spread.
Apart from the visit counts, various other metrics were collected and published. These metrics can provide an idea about the overall change in foot traffic in each POI. For example, the plots below show the average visit concentration and average median dwell time for each week.
It is evident from these plots that visit concentration and median dwell time has reduced as the lockdown progressed. This shows that there is a decent amount of compliance to the lockdown among the public.
Time series data for a POI is of short length - 33 data points. And the goal is to forecast the next immediate datapoint. Initial exploratory visualizations showed that the time series for foot traffic is stationary (having constant statistical parameters). Models like ARIMA can capture the trend well enough to make good predictions without overfitting the data.
This hackathon was organized with the purpose to build a model that can help in policy-making decisions of the public health officials. Using simpler models like Autoregression can help non-technical people to understand the working of the model and help in instilling confidence on using data science to address public issues. Therefore, model interpretability is given equal importance as accuracy.
The time series model that scored best in the hackathon is a simple Auto Regression model of order 1, which uses foot traffic data as the endogenous variable. A python library called 'statsmodels' was used to build the model. Variations of the model were tried but the model with the least complexity was able to outperform all other complex models. This shows that not every problem needs a deep neural network solution; even simple and naive models can answer important questions if used properly.
Surprisingly, extrapolation of time series data using Fourier Transform has outperformed Random Forest Regressors, one of the powerful regression models in machine learning, in this task. A python function was written from scratch to deconstruct and reconstruct a time series using Fourier Transform. The forecasted points are added during reconstruction of the fourier series. The number of harmonics (terms) used in the fourier series played a vital role in generating accurate forecasts.
From these maps and plots, we can confidently say that shopping centers and educational institutions are the busiest areas in Tippicanoe country during the pandemic. It is also worthy to note that these places did not get crowded simultaneously. During the first half of the data, shopping centers received high foot traffic whereas the visit counts in Purdue University started to increase only in the second half of the data. After week 30, food outlets like McDonald's and Chick-fil-A also received huge number of visitors.
To conclude, locations like Walmart, Purdue University, Sam's Club, McDonald's can be classified as high-risk areas and public health officials should focus more on these locations in order reduce the infection spread.