Feature Selection and Classification of Post-Graduation Income of College Students in the United States

Ewan Wright, Qiang Hao, Khaled Rasheed, Yan LiuarXiv 2018 · arXiv

Introduction

While a college degree is often viewed as a primary driver of social mobility, the "graduate premium"—the earnings gap between college and high school graduates—is increasingly complex. Despite the gap doubling since 1979, many recent graduates face underemployment or loan defaults. This study investigates the specific institutional, demographic, and socioeconomic attributes that predict a student's income six years after graduation, moving beyond anecdotal evidence to identify the hidden drivers of financial success for students receiving federal financial aid.

Methodology

The researchers employed a two-phase machine learning pipeline consisting of feature selection and classification. The study utilized the 2015 U.S. Department of Education’s College Scorecard dataset, focusing on 1,429 cohorts of graduates from 2001, 2003, and 2005.

Feature Selection

Starting with 30 potential attributes, the study compared three distinct approaches to identify the most informative subset:

  • Filter Methods: Statistical techniques (e.g., Chi-square) that rank features independently of a specific model.
  • Forward Selection: An iterative wrapper approach that adds variables based on model improvement.
  • Genetic Algorithm (GA): An evolutionary search method that mimics natural selection to find an optimal combination of features.

Classification

Once features were selected, ten different algorithms across five categories (Bayesian, Function-based, Instance-based, Tree-based, and Rule-based) were tested to classify cohorts into four income brackets, ranging from "Very Low" (<$25k) to "High" (>$50k).

Results and Key Insights

The study found that the choice of feature selection method significantly impacts predictive accuracy. The Genetic Algorithm outperformed simpler methods, identifying a robust set of 22 features (see Figure 4). This suggests that post-graduation income is driven by a complex web of interacting factors rather than a few isolated variables.

Attributes selected by the Genetic Algorithm, which proved to be the most effective feature selection method.

In terms of classification, the Support Vector Machine (SVM) using a Pearson VII function kernel achieved the highest performance, with an accuracy of 0.753 and an F1-score of 0.745. As illustrated in Figure 1, SVM and K-Nearest Neighbor (KNN) models consistently outperformed other classifiers like Naïve Bayes or Decision Trees.

Comparison of F1-scores across different classification models.

Several key socioeconomic insights emerged from the data:

  • The Neighborhood Effect: One of the strongest predictors was the percentage of the population in a student’s home Zip Code with a professional degree, suggesting that social capital and environment significantly influence career trajectories.
  • The First-Gen Ceiling: A clear negative correlation was observed between the percentage of first-generation college students in a cohort and their post-graduation income, likely due to a lack of informed career support.
  • SAT Scores as Proxies: While SAT scores were strong predictors, they were highly correlated with parental income and neighborhood status, suggesting they may act as a proxy for existing socioeconomic privilege.

Conclusion and Limitations

The findings challenge the "great equalizer" narrative of higher education, suggesting that institutional and socioeconomic backgrounds often maintain existing social stratifications. However, the study is limited by a temporal lag—using data from 2001–2005 cohorts—and a selection bias, as it only includes students who received financial aid. Future research should utilize mediational models to further untangle the causal relationships between neighborhood status, institutional selectivity, and long-term financial outcomes.

Summary updated

BibTeX
@misc{wright2018feature,
  author = {Ewan Wright and Qiang Hao and Khaled Rasheed and Yan Liu},
  title = {{Feature Selection and Classification of Post-Graduation Income of College Students in the United States}},
  year = {2018},
  eprint = {1803.06615},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1803.06615},
}
Figure