Analytics work across HR, healthcare, and business.

The core challenge: only 16% of employees actually left, making this an imbalanced classification problem.

Employees who work overtime leave at roughly 3× the rate of those who don't.

Employees who left earned less on average, with a lower and more compressed income range.

Tenure-related variables (years at company, in role, with manager) cluster tightly together, as does income with job level.

The pruned decision tree: overtime is the very first split, followed by income, job role, and marital status - the same signals that show up across every model in this project.

Model accuracy compared:

Model accuracy on held-out data. Notice that Random Forest (87.0%) barely beat the do-nothing baseline (84.0%) - with only 237 of 1,470 employees actually leaving, a model can score well just by leaning on the majority class. Accuracy alone doesn't tell you whether a model is actually catching at-risk employees.

Rebalancing the training data traded 13 points of accuracy for nearly double the sensitivity - the model went from catching under half of employees who actually left to catching most of them. XGBoost's built-in class weighting reached a smaller version of the same trade-off without touching the training data at all.

Employees who work overtime left at roughly 3× the rate of those who don't (31% vs. 10%), the single strongest individual signal in the dataset. The estimated dollar figures above are illustrative - based on a commonly cited 50%-of-salary replacement-cost benchmark, not this specific organization's real costs.

SHAP values confirm it from a different angle: overtime and stock option level carry the most weight in the model's predictions, ahead of income, commute distance, and relationship satisfaction - a second model, a second method, the same story.

The core obstacle for award prediction: only 2.32% of employees ever won an award, an extreme class imbalance that misleads naive accuracy scores.

Promotion rates vary meaningfully by department - Technology promotes at nearly double the rate of Legal (10.8% vs. 5.1%).

The balanced Logistic Regression model reached an AUC of 0.74 - meaningfully better than chance, though far from a confident predictor on its own.

The numeric variables barely correlate with each other - age and length of service are the one real exception (0.66) - meaning each one carries mostly independent signal.

One variable dominates every other signal combined: average training score is by far the strongest predictor of promotion, followed by age, length of service, and last year's performance rating.

The unbalanced Logistic Regression model looked fine on paper but was functionally useless - it predicted "not promoted" almost every time and caught only 9% of employees who actually were. Random Forest handled the same imbalance far better on its own, reaching 93% accuracy. But award prediction told a harder story: with awards won by barely 2% of employees, K-Nearest Neighbors identified just 3 of 318 real award-winners, a reminder that some outcomes are too rare for a model to learn reliably without deliberately rebalancing the data first.

The target: roughly 60% of students in the dataset graduated, 40% did not.

Students who are both scholarship holders and up-to-date on tuition skew heavily toward graduating.

Most students' grades stay fairly stable between semesters, with a small tail of large drops.

The decision tree's first split is second-semester units approved, then tuition payment status and admission grade - a mix of academic performance and financial standing.

Model accuracy compared:

SVM with a linear kernel edged out every other model, including a Random Forest that badly underperformed here - a reminder that more complex models don't always win. Tuition payment status and second-semester units approved were the strongest predictors of graduation, pointing to financial standing and early academic momentum as the two biggest levers for student retention efforts.

The methodology here isn't specific to higher education - it's a general template for early-warning retention modeling. In HR, the same approach flags which new hires are at risk of leaving before their first review, using early engagement and performance signals the way this model uses first-semester grades and units approved. For an NGO or membership organization, the same logic applies to donor or beneficiary attrition: identifying who's disengaging early enough to intervene, rather than finding out only when they're already gone.

Hiring outpaced terminations every year, but both climbed together - the organization wasn't just growing, it was churning.

Turnover rate spiked to 73.3% in 2021 before easing back in 2022 - still well above a healthy benchmark.

Personal/family reasons and better opportunities elsewhere dominate the exit data - the first is hard to control, the second is not.

Tenure climbs with age: younger employees cluster at low tenure, while longer-tenured staff skew older.

Pay band tells the clearest story: employees earning under $35,000 had the shortest tenure by a wide margin, while the $105,000-$155,000 and $155,000+ bands held onto staff far longer.

Part-time employees left earlier than full-time employees overall, and among full-time staff with short tenure, the majority were female - a pattern worth a closer look alongside the compensation data above.

For a critical-care air medical transport operation, "chute time" - the minutes between a 911 dispatch and the aircraft leaving the ground - is one of the few variables an organization can actually influence, and every minute shaved off it is a minute closer to definitive care. This project analyzed five years of dispatch and weather data across Life Flight of Maine's fleet and crews to find out what actually moves that number, and who's doing it fastest.

Share of missions flown by aircraft:

LF1 and LF2 carried the bulk of the workload, together flying more than three-quarters of all missions between 2019 and 2023. LF1's heavier use cut both ways: it also swung between the fleet's shortest and longest chute times, suggesting that volume alone doesn't guarantee consistency - crew and conditions matter just as much as the aircraft.

Average chute time in minutes from dispatch to airborne - lower is faster. Conditions mattered as much as crew: chute times ran roughly 2 minutes longer at night, and climbed further in colder, lower-visibility weather, a reminder that the hardest missions to fly fast are often the ones where speed matters most. Based on consistently faster times across conditions, the team recommended the crew members above as internal trainers - turning the analysis into a concrete, actionable staffing recommendation rather than just a finding.

Most of the richest signal in an organization - exit interview notes, open-ended engagement survey responses, candidate feedback - sits in free text that never gets systematically read, let alone analyzed. This project built the full pipeline for turning that kind of unstructured text into a quantified, visual summary: cleaning and normalizing a corpus, stripping noise and stop words, building a term-document frequency matrix, and surfacing the dominant terms as both a ranked chart and a word cloud. The practice corpus here is a classic literary text, chosen to stress-test the pipeline on dense, unfamiliar language - but the technique transfers directly to any body of open-ended HR text.

The cleaned, stop-word-free term frequencies rendered as a word cloud - word size reflects frequency, giving a reader the corpus's dominant themes at a glance, with no need to read the source text line by line.

The same signal as a ranked bar chart - the top term appeared 307 times, more than double the tenth-ranked term (133), the kind of sharp drop-off that makes a handful of themes easy to prioritize.

The pipeline - lowercase conversion, punctuation and number stripping, custom and standard stop-word removal, then term-document frequency counting - is the same sequence used to mine open-ended HR text at scale. Run against exit-interview comments, it surfaces recurring complaint themes without a human reading every response line by line. Run against engagement-survey free text, it turns hundreds of open comments into a handful of dominant, trackable themes in minutes instead of days - the kind of leverage that turns unstructured feedback from an ignored data source into a regular input for leadership decisions.

A fairly balanced dataset: 162 of 300 patients (54%) show no heart disease, 138 (46%) do - close enough to even that accuracy alone is a meaningful measure here.

Patients with heart disease skew older, peaking in their late 50s and 60s, while disease-free patients are more evenly spread across their 40s and 50s.

Cholesterol does trend upward with age, but the overlap between patients with and without heart disease is substantial - neither variable alone cleanly separates the two groups.

The clearest visual signal in the dataset: patients with heart disease (green) cluster heavily at chest pain type 4 and higher ST depression values, while disease-free patients sit mostly at zero.

All three models performed respectably - AUCs ranged from 0.79 to 0.85 - but logistic regression won on every metric that mattered: highest accuracy (86%), highest specificity (94.3%), and the clearest interpretability. Rather than a black box, it named its reasoning: being male, chest pain types 2 and 4, elevated ST depression, and especially the number of major vessels colored by fluoroscopy (ca) came back as statistically significant predictors, with ca the single strongest (p < 0.0001). For a clinical screening context, a model that both performs well and tells you why is worth more than a marginal accuracy gain from a less transparent one.

The logistic regression model's ROC curve - an AUC of 0.848 confirms strong discrimination between patients with and without heart disease, well above the diagonal line representing random guessing.

A near-even split - 51.8% edible, 48.2% poisonous - meaning a model has to actually learn the distinguishing traits rather than lean on class imbalance.

Cap shape and color alone are nearly useless for telling edible and poisonous mushrooms apart - edible (green) and poisonous (orange) points sit tangled together across almost every combination.

Habitat offers a weak clue at best: paths (p) skew heavily poisonous and grasses/woods (d) lean edible, but no habitat is a clean separator on its own.

One feature stands apart from every other: odor alone produces 8 perfect splits - cleanly sorting specimens into 100% edible or 100% poisonous groups - nearly 1.5× the next-best feature.

The pruned decision tree in action: a single odor split correctly flags 46% of specimens as poisonous with no further questions needed, and resolves the rest with one follow-up question about spore print color.

All three decision tree variants - a full tree, an optimized 60/40 split, and a pruned tree - landed between 99.32% and 99.48% accuracy, with 100% specificity across the board: not one poisonous mushroom was ever misclassified as edible in any model. The handful of errors all ran the safer direction, occasionally calling an edible mushroom poisonous rather than the reverse. With odor alone sorting roughly half the dataset perfectly, the practical takeaway is refreshingly simple: a foul or pungent odor is one of the most reliable, no-equipment-needed warning signs a forager or field guide can use.