A Repeatable ML Workflow for Bootcamp Students
For bootcamp students learning ML · Based on Kylie Ying ML for Everyone Framework
// TL;DR
Bootcamp students can use the ML for Everyone Framework as a reliable mental checklist that turns scattered lessons into one repeatable pipeline: inspect and encode data, split into train/validation/test, scale, handle imbalance, train multiple models, and evaluate with F1. Instead of re-Googling the order of steps for every assignment, you internalize a sequence that prevents the classic beginner mistakes — data leakage, oversampling test data, and reporting misleading accuracy. Use it on any tabular classification or regression project to produce work that holds up when a reviewer checks your methodology.
Why do bootcamp assignments feel so disorganized?
Because you learn concepts — encoding, scaling, KNN, neural networks — as isolated lessons, but nobody hands you the order to apply them in. The ML for Everyone Framework gives you that order as a fixed ten-step workflow. Once you internalize it, every new dataset becomes the same repeatable process: you stop asking 'what do I do first?' and start executing. This is the difference between memorizing algorithms and actually being able to ship a model.
What is the correct order of steps?
Follow the workflow exactly: (1) load and inspect the data, identifying X and y; (2) encode categorical features — one-hot for nominal, ordered integers for ordinal; (3) plot feature histograms by class to see what's discriminative; (4) shuffle, then split 60/20/20 into train/validation/test; (5) fit StandardScaler on training data only and transform the rest; (6) apply RandomOverSampler to the training set if classes are imbalanced; (7) train KNN, Naive Bayes, Logistic Regression, and SVC; (8) run classification_report for each; (9) compare F1 scores and tune hyperparameters on the validation set; (10) optionally escalate to a neural network. Memorize this sequence — it's the backbone of nearly every tabular ML project you'll face.
How do I avoid the mistakes graders look for?
The three mistakes that tank bootcamp grades are all preventable. First, never fit StandardScaler on your full dataset or test set — fit on training only, then transform. Fitting on test data is data leakage and reviewers will spot it immediately. Second, never oversample your validation or test sets; oversampling is training-only. Third, never use the test set to pick your model — that's the validation set's job, and the test set is used exactly once at the very end. Also remember to shuffle before splitting, or you'll bake ordering bias into your splits.
Why should I train multiple models instead of one?
Training four baselines — KNN, Naive Bayes, Logistic Regression, and SVC — teaches you that no single algorithm wins everywhere and gives you a comparison table that impresses reviewers. You'll see distance-based, probabilistic, and margin-based approaches behave differently on the same data. Compare their F1 scores, not accuracy, especially on imbalanced datasets. This habit also builds intuition: over a few projects you'll start predicting which model will win before you even run them.
When should I reach for a neural network?
Only after your classical models plateau. Beginners often jump straight to deep learning because it sounds impressive, but a neural network without activation functions collapses into a linear model, and networks overfit fast on the small datasets typical of bootcamp assignments. When you do build one, use ReLU in hidden layers, Sigmoid for binary output, and watch training versus validation loss per epoch to catch overfitting.
Next step
Pick a Kaggle-style tabular dataset and run the full ten-step workflow end to end without skipping any step, even if some feel unnecessary. Doing the complete sequence once, correctly, cements the pattern far better than half-doing it on ten datasets.
// FREQUENTLY ASKED QUESTIONS
Do I really need a validation set separate from the test set?
Yes. The validation set is for tuning hyperparameters and comparing models during development, while the test set is used exactly once at the end for your final reported score. If you tune on the test set, your reported performance is inflated and dishonest — a mistake reviewers and graders catch instantly. Keep all three splits fully separate.
What's the most common beginner mistake this framework prevents?
Data leakage from fitting StandardScaler on the entire dataset or the test set. Always fit the scaler on training data only, then transform validation and test with that same fitted scaler. The framework builds this habit into step five so your model's reported performance reflects true generalization rather than memorized test-set statistics.
Can I skip steps if my data looks clean?
Not while you're learning. Running the full ten-step workflow — even when scaling or oversampling feels unnecessary — cements the pattern and trains you to notice when a step actually matters. Skipping steps early leads to inconsistent habits and mistakes on messier real datasets. Do the complete sequence correctly once before you start optimizing for speed.