What needed to be understood.
Loan acceptance was the minority outcome, so overall accuracy alone could make a weak model appear successful. The workflow needed to consider class balance and the business cost of missed prospects and false positives.
How the analysis was built.
- Prepared customer variables and separated the target from predictors.
- Built k-nearest-neighbor classification processes in RapidMiner.
- Compared model outcomes using validation data and a confusion matrix.
- Interpreted predictions as a targeting aid rather than an automatic lending decision.
What the evidence showed.
- The exercise showed how neighborhood-based classification can identify patterns among customers with similar profiles.
- Class imbalance made minority-class performance more important than raw accuracy.
- Model usefulness depended on balancing campaign reach, conversion opportunity, and contact cost.
What should happen next.
- Evaluate precision, recall, and lift for likely acceptors in addition to accuracy.
- Compare k-NN against logistic regression and tree-based models.
- Use the model for marketing prioritization only, with separate compliant underwriting controls.
What this project does—and does not—prove.
The academic dataset and model were not validated for real lending decisions, fairness, regulatory compliance, or current customer populations.