Informs2020 using machine learning to identify the factors of people's mobility
0 likes45 views
The document examines the complex factors influencing human migration and their impact on population growth in U.S. counties, while highlighting the necessity of modeling these factors for infrastructure planning. It describes the use of various data sources and machine learning techniques, including Principal Component Analysis (PCA) and Random Forest Regression (RFR), to identify and isolate significant variables affecting migration. The research aims to provide a framework to understand migration dynamics through public data and machine learning methodologies.
Informs2020 using machine learning to identify the factors of people's mobility
- 1. Using Machine Learning to Identify the Factors of People's Mobility INFORMS2020 Alexander Gilgur Jose Emmanuel Ramirez-Marquez The research performed by Jose E. Ramirez Marquez leading to these results has received funding from the National Science Foundation, CRISP Type 2 / Collaborative Research: Resilience Analytics: A Data-Driven Approach for Enhanced Interdependent Network Resilience, Award number 1541165.
- 2. Motivation and Problem Statement Migration is one of the key factors of population growth in US counties. Modeling it enables planning local infrastructure and its supply chains. 2
- 3. Why do people pick up and leave and go to the new places? ● Opportunities & Employment ● Infrastructure ● Affordability ● Stability ● Happiness & Unhappiness ● ... ... ... Migration is one of the key factors of population growth in US counties. Modeling it enables planning local infrastructure and its supply chains. Motivation and Problem Statement 3
- 4. Modeling the effects of migration factors is complicated ● Confounded Causality ● Multicollinearity ● Nonlinear relationships ● Differences between Domestic & International migration Migration is one of the key factors of population growth in US counties. Modeling it enables planning local infrastructure and its supply chains. Motivation and Problem Statement 4 Why do people pick up and leave and go to the new places? ● Opportunities & Employment ● Infrastructure ● Affordability ● Stability ● Happiness & Unhappiness ● ... ... ...
- 5. Modeling the effects of migration factors is complicated ● Confounded Causality ● Multicollinearity ● Nonlinear relationships ● Differences between Domestic & International migration We present a way to identify the factors that have an effect on domestic, international, and overall migration as first step to modeling migration Migration is one of the key factors of population growth in US counties. Modeling it enables planning local infrastructure and its supply chains. Motivation and Problem Statement 5 Why do people pick up and leave and go to the new places? ● Opportunities & Employment ● Infrastructure ● Affordability ● Stability ● Happiness & Unhappiness ● ... ... ...
- 7. Data Collection 1. Community Commons Map 2. Bureau of Labor Statistics Employment Data 3. Census Bureau Income Inequality Data 4. Census Population Data 5. CDC Cause of Death Data: Suicide Mortality 6. USDA Economic Research Service Education Data Government APIs are not very easy to use Data were downloaded as Excel or CSV files and saved to Dropbox They were later preprocessed for analysis 7
- 8. EDA: Target Variables’ Correlation Matrix 0: NETMIG 1: DOMESTICMIG 2: INTERNATIONALMIG 0: year 1: BIRTHS 2: CENSUSPOP 3: DEATHS 4: EMPLOYED 5: ESTIMATESBASE 6: GDP 7: GDP_LOG 8: LABOR_FORCE 9: NATURALINC 10: NPOPCHG_ 11: POPESTIMATE 12: POP_CDC 13: RBIRTH 14: RDEATH 15: RNATURALINC 16: SCDEATHS 17: SCRATE 18: SC_R_DEATH 19: UNEMPLOYED 20: UNEMPLOYMENT_RATE_PCT 21: bachelors_degree_plus 22: high_school_diploma 23: lt_high_school_diploma 24: mean_income__dollars 25: mean_to_median_income_ratio 26: median_income__dollars 27: pct__bachelors_degree_plus 28: pct__high_school_diploma 29: pct__lt_high_school_diploma 30: pct__some_college_or_assoc_degree 31: some_college_or_assoc_degree 8
- 9. EDA: Target Variables’ Correlation Matrix 0: NETMIG 1: DOMESTICMIG 2: INTERNATIONALMIG 0: year 1: BIRTHS 2: CENSUSPOP 3: DEATHS 4: EMPLOYED 5: ESTIMATESBASE 6: GDP 7: GDP_LOG 8: LABOR_FORCE 9: NATURALINC 10: NPOPCHG_ 11: POPESTIMATE 12: POP_CDC 13: RBIRTH 14: RDEATH 15: RNATURALINC 16: SCDEATHS 17: SCRATE 18: SC_R_DEATH 19: UNEMPLOYED 20: UNEMPLOYMENT_RATE_PCT 21: bachelors_degree_plus 22: high_school_diploma 23: lt_high_school_diploma 24: mean_income__dollars 25: mean_to_median_income_ratio 26: median_income__dollars 27: pct__bachelors_degree_plus 28: pct__high_school_diploma 29: pct__lt_high_school_diploma 30: pct__some_college_or_assoc_degree 31: some_college_or_assoc_degree USCB models population by migrations, births, and deaths => collinearity & data leakage Factors positively correlated with international migration are negatively correlated with domestic migration => low correlation with overall migration These variables are not all independent; we need to make them such. 9
- 10. Making Variables Independent: Principal Component Analysis PCA finds the linear combinations (principal components, or PCs) of original variables that maximize the variances of the principal components This results in covariance being 0 => principal components are independent. 10
- 11. Principal Component Analysis and Dimensionality Reduction Problem - thresholds are arbitrary EV Threshold = 0.01 25 PCs 13 PCs EV Threshold = 0.05 5 PCs 11
- 12. Dimensionality Reduction: Random Forest Regression Random Forest Regression (RFR) is one of the most robust methods for modeling multidimensional data. It builds a forest of numerical decision trees and ensembles the decision paths from all trees by majority voting. It is non-parametric => it does not rely on knowledge of an underlying model. We can calculate feature importance for any model, enabling dimensionality reduction by identifying the critical features for further analysis. Due to the way RFR operates, feature importance evaluation based on RFR is model-agnostic, ensuring that if further modeling fails to identify a feature as important, we can rely on RFR-based feature importance for model validation. 12
- 13. EDA: ● Identify Correlated Features PCA: ● Combine Features into PCs RFR: ● Identify Important PCs Features Principal Components The EDA to PCA to RFR Daisy Chain 13
- 14. EDA: ● Identify Correlated Features PCA: ● Combine Features into PCs RFR: ● Identify Important PCs Features Principal Components Ytest PredictiononXtest R2 test = 0.78 The EDA to PCA to RFR Daisy Chain 14
- 15. EDA: ● Identify Correlated Features PCA: ● Combine Features into PCs RFR: ● Identify Important PCs Features Principal Components Ytest PredictiononXtest R2 test = 0.78 Important Features Ytest PredictiononXtest R2 test = 0.78 R2 stabilized ? The EDA to PCA to RFR Daisy Chain 15
- 16. The EDA to PCA to RFR Daisy Chain EDA: ● Identify Correlated Features PCA: ● Combine Features into PCs RFR: ● Identify Important PCs Features Principal Components Ytest PredictiononXtest R2 test = 0.78 Important Features PC Importances Principal Component PC Importance R2 stabilized ? 16
- 17. EDA: ● Identify Correlated Features PCA: ● Combine Features into PCs RFR: ● Identify Important PCs Features Principal Components Ytest PredictiononXtest R2 test = 0.78 Important Features PC Importances Principal Component PC Importance R2 stabilized ?
- 18. Going Back to Original Features 18 We have come full circle: 1) Identified the correlated variables 2) Applied PCA to rotate the data into a form where principal components are orthogonal. 3) Used Random-Forest Regression (RFR) to perform final PC selection based on their importance. Now we need to 4) Transform the selected PCs back to the original features. 5) Use these features in modeling.
- 19. Further Work 19 ● With the features identified by the PCA->RFR process, build the model of migration in the form of Poiseuille's Equation: ● Infer the pressure and the resistance terms and model them based on the features: ○ Discontent ○ Socioeconomic data ○ GDP ○ Education level ○ Population volume ● Build a model predicting movements of large masses of population
- 20. Conclusions 20 ● Mass migration is a high complexity multivariate problem ● Its features interact, overlap, and exhibit multicollinearity in many ways ● Properly modeling mass migrations requires understanding of the underlying processes, which is impossible to achieve without dimensionality reduction ● We have described how this can be achieved based on publicly available data and using the standard Machine-Learning techniques in a creative manner.
- 21. Thank You! 21
- 22. Appendix 22
- 23. Emigration: Mass Transfer Out Information about other places Possibility of “Flight” Affordability of “Flight” “Flight” Discontent High Stress Levels Navier-Stokes Equation: Poiseuille’s Equation: ------- Effluent rate is proportional to pressure difference and inversely proportional to resistance 23
- 24. Information about other places Possibility of “Flight” Affordability of “Flight” “Flight” Immigration: Diffusion into New Host Community Affordability of Living Media Sentiment New Host Community Acceptance of Strangers 24
- 25. Absorption of Community Characteristics characteristics of the incoming community Newcomers bring with them features that can be new to the host community. As time passes by, they retain fewer of such characteristics, while members of the community that has adopted them become penetrated with the new features. Likewise, features of the host community become characteristics of its new members. In the new members of the community In the members of the host community Time Saturation Time Saturation characteristics of the host community In the members of the host community In the new members of the community (a) (b) 25