Introduction:

In this project, I explore how feature selection, model interpretability, and evaluation choices shape text classification systems. Using email spam detection as a case study, I examined how different words contribute to classification decisions, how adjusting thresholds affects false positives and false negatives, and how ambiguity in labeled data complicates the idea “ground truth.”

A central theme of this project was the exploration of Machine Learning techniques. Rather than treating the model’s output as definitive, I investigated why certain features mattered, how small changes could flip classifications, and what tradeoffs emerge when models are deployed in real-world contexts like content moderation. This approach highlights why data analysis is not just about improving metrics, but about understanding model behavior and consequences.

The visualization revealed that the word “font” had the strongest correlation with spam emails, while words such as “free” had much weaker correlations. This insight helped guide feature selection by identifying which words were actually informative and which were not, challenging initial assumptions about spam language.

I tested different word features by incorporating them into exploratory analysis and observing how they affected model performance. This process revealed that many words appeared very rarely in both spam and ham emails, limiting their usefulness as predictive features. I was surprised to find that even commonly associated spam words did not consistently distinguish spam from non-spam. This reinforced the idea that better performance often comes from better features, not more complex models.

This visualization highlights that classification is not a binary decision but a continuum. Adjusting the threshold allows the model to prioritize catching more spam at the cost of misclassifying legitimate emails, or vice versa. This tradeoff mirrors real-world decision-making in domains like medical screening and content moderation. 

I examined how removing a single feature could flip an email’s classification and reflected on how ambiguity in spam/ham labels affects model evaluation.

This analysis showed that simple models can be highly interpretable, making it easier to understand why predictions change. At the same time, it emphasized that labeled data is not always an objective ground truth—different people may reasonably disagree on whether an email is spam. This ambiguity directly impacts how we interpret model performance and fairness. 

This project reinforced that building a classification model is only the beginning; understanding why it behaves the way it does is where meaningful insight emerges. Through feature exploration, correlation analysis, and threshold tuning, the model evolved from a simple classifier into a lens for examining how assumptions about language, labels, and evaluation shape outcomes.

One of the most important lessons was how easily intuition can be challenged by data. Words that seemed like obvious spam indicators turned out to be weak signals, while less obvious features carried stronger predictive power. Similarly, small changes—such as removing a single feature or adjusting a probability threshold—were enough to flip classifications, revealing how sensitive models can be and how critical interpretability is in understanding those decisions.

This project also highlighted that data is rarely a perfect reflection of truth. Ambiguity in spam labels showed that disagreement is often inherent in human-generated data, and that models ultimately learn patterns shaped by these subjective decisions. As a result, evaluating performance requires more than optimizing metrics; it requires considering tradeoffs, consequences, and context.

Overall, this analysis reflects why I enjoy data science: curiosity-driven exploration transforms models from black boxes into understandable systems. By asking deeper questions about features, evaluation, and interpretation, data analysis becomes a process of learning—not just about the data, but about the systems and people affected by it.

This comparison shows that while logistic regression only slightly improves accuracy over the zero predictor, it is still meaningfully better because it identifies some spam emails correctly.

This comparison shows that while logistic regression only slightly improves accuracy over the zero predictor, it is still meaningfully better because it identifies some spam emails correctly.

Accuracy alone can be misleading in spam detection due to class imbalance. Precision measures how often emails predicted as spam are actually spam, while recall captures how effectively the model identifies spam emails. The false positive rate quantifies how often legitimate emails are incorrectly labeled as spam. Together, these metrics provide a more complete picture of model behavior and help evaluate the tradeoffs between blocking spam and avoiding the misclassification of legitimate messages.

Rather than relying on a single accuracy score, this project examines how a classification model behaves in practice by breaking down its predictions and errors. The word-frequency visualization reveals that many commonly assumed “spam indicators” appear infrequently in both spam and ham emails, which limits their usefulness as predictive features. This highlights an important lesson: model performance is often constrained by feature quality rather than the choice of algorithm.

To better understand prediction outcomes, the logistic regression model is evaluated using a confusion matrix and derived metrics such as precision, recall, and false positive rate. Precision helps quantify how often emails flagged as spam are truly spam, while recall captures how effectively the model identifies spam emails at all. The false positive rate provides insight into how often legitimate emails are incorrectly filtered, a key concern for user experience.

Comparing these results with a zero predictor emphasizes why accuracy alone is insufficient. Although the zero predictor achieves relatively high accuracy due to class imbalance, it completely fails to detect spam. The logistic regression model, while only modestly improving accuracy, demonstrates more practical value by correctly identifying some spam emails. Together, these analyses illustrate how thoughtful evaluation and curiosity-driven exploration uncover model limitations and tradeoffs that would otherwise remain hidden.