Predicting Metropolitan Status from Structural Cost Shares Using Machine Learning

This repository contains the full implementation, analysis, and research paper for a machine learning study investigating whether U.S. county metropolitan status can be predicted solely from structural household expenditure shares and minimal demographic indicators. The research demonstrates that cost-share compositions encode strong geographic and socioeconomic signals, enabling accurate metro classification even when explicit geographic identifiers are missing.

---

Project Overview

The central research question is:

Can structural cost-share patterns alone reliably classify counties as metropolitan or non-metropolitan?

To answer this, the project uses normalized expenditure shares for:

• Housing
• Transportation
• Healthcare
• Food
• Childcare
• Taxes
• Other necessities

Along with:

• SNAP participation rate
• Family type
• State abbreviation

A complete machine-learning pipeline was constructed, including:

• Feature normalization and engineering
• Imputation, scaling, and one-hot encoding using a unified preprocessing pipeline
• SMOTE oversampling for class balance
• PCA for structural dimensionality analysis
• Comparative evaluation of six classifiers
• Aggregated feature importance across model families

The best-performing model, a Neural Network (MLP), achieved:

• Test ROC AUC: 0.9528
• Test Weighted F1 Score: 0.8864

These results confirm that cost-share structures alone provide a strong and interpretable signal of metropolitan economic conditions.

---

Visual Summary

PCA Projection of Cost-Share Features

This plot shows how counties cluster in PCA space using only structural expenditure shares and minimal demographic inputs. Metro and non-metro counties form partially separable clusters along PC1, reflecting underlying urban–rural cost gradients.

---

Machine Learning Pipeline Diagram

The full modeling workflow, from preprocessing through evaluation.

--

Repository Structure

📁 project-folder │ ├── 📄 metro_classification.ipynb │ └─ Full analysis: EDA, PCA, modeling pipeline, evaluation │ ├── 📄 metro_status_paper.pdf │ └─ Final IEEE-style research paper │ ├── 📄 README.md │ └─ Project overview and documentation │ └── 📁 data └── data.csv # County-level cost-share and demographic dataset

---

Methods Summary

Preprocessing Pipeline

A unified preprocessing pipeline was implemented using:

• Median imputation for numeric variables
• Standardization via StandardScaler
• Most-frequent imputation for categorical variables
• One-hot encoding with handle_unknown='ignore'

This ensures consistent, leakage-free transformations across all models.

Models Evaluated

A diverse set of classifiers was compared:

• Logistic Regression
• Decision Tree
• Random Forest
• Gradient Boosting
• Support Vector Classifier
• Neural Network (MLP)

Handling Class Imbalance

SMOTE oversampling was applied inside an ImbPipeline to ensure oversampling occurred only on training folds, preserving the integrity of validation and test sets.

Evaluation Metrics

The models were compared using:

• ROC AUC (primary metric)
• Weighted F1 score
• Confusion matrix
• Balanced vs. unbalanced performance comparison
• Combined feature importance across three model families

Validation Results

Rank	Model	Validation F1	Validation ROC AUC
1	NeuralNet (Unbalanced)	0.8908	0.9541
2	NeuralNet (Balanced)	0.8849	0.9525
3	SVC (Unbalanced)	0.8732	0.9291
4	RandomForest (Balanced)	0.8414	0.9152
5	RandomForest (Unbalanced)	0.8293	0.9120
6	GradientBoosting (Balanced)	0.8217	0.8889
7	GradientBoosting (Unbalanced)	0.8156	0.8851
8	LogReg (Balanced)	0.8157	0.8776
9	LogReg (Unbalanced)	0.8221	0.8770
10	DecisionTree (Balanced)	0.8107	0.8204
11	DecisionTree (Unbalanced)	0.8146	0.8093

---

Key Findings

• Structural expenditure shares convey rich socioeconomic information that aligns closely with metro–non-metro distinctions.
• Housing and other-necessities shares were the strongest positive predictors of metro status.
• Transportation and healthcare shares were the strongest negative predictors.
• PCA revealed clear clustering structure aligned with metropolitan status, despite no geographic inputs.
• Neural Networks provided the most accurate and generalizable decision boundary across nonlinear feature interactions.
• The final test performance confirms strong predictive capability in unseen county data.

---

How to Use This Repository

1. Open the Notebook

Run the complete analysis using:

metro_classification.ipynb

It includes:

• Data loading
• Feature engineering
• Exploratory data analysis
• PCA projection
• Model training and validation
• SMOTE balancing
• Final model evaluation
• Confusion matrix and ROC curve

2. View the Research Paper

The report is available here:

metro_status_paper.pdf

This document contains full methodological detail, figures, analysis, interpretation, discussion, and references.

---

Citation

If you use or reference this project, please cite:

M. J. Mungoshi, "Predicting Metropolitan Status from Structural Cost Shares Using Machine Learning," Clarkson University, 2025.

---

Author

Mutsa Jonah Mungoshi
Applied Data Science
Clarkson University
Email: mungosmj@clarkson.edu