scikit-learn
Data · Test
Build baseline models and evaluate them with consistent train/test splits.
All capabilities · Machine learning & data science
Train and evaluate classical ML models
Scikit-learn workflow: features, train/test split, cross-validation, metrics, overfitting.
Explore 5 tools for this projectMarket relevance
Which roles ask for this — and how often
Share of job postings in India, per role, that name this capability.
What employers mean
You should be able to…
- Load and clean a tabular dataset (missing values, outliers, encoding categoricals)
- Engineer features and build a scikit-learn Pipeline that doesn't leak test data
- Do a correct train/validation/test split and k-fold cross-validation
- Train and compare a linear model, a tree, and a boosted model (XGBoost/LightGBM)
- Pick and justify the right metric (accuracy vs precision/recall/F1/AUC/RMSE) for the business problem
- Diagnose overfitting vs underfitting from learning curves and fix it
- Tune hyperparameters with GridSearchCV/RandomizedSearchCV or Optuna
- Explain feature importance / SHAP values to a non-technical stakeholder
Needs first: Write production-quality Python for AI work
Learn — free, link-checked
The few resources that matter
Read · beginner · 25 min · scikit-learn.org
Getting Started
The canonical Pipeline/GridSearchCV/metrics pattern straight from the library docs -- copy-correct code beats half-remembered tutorials. — scikit-learn
Course · beginner · 180 min · kaggle.com
Intro to Machine Learning
Hands-on, notebook-based: train your first scikit-learn model and submit to a Kaggle competition the same afternoon. — Kaggle Learn
Watch · beginner · 500 min · youtube.com
Machine Learning
Clear, visual explanations of bias/variance, regularization, trees, boosting and metrics -- the intuition interviewers probe for beyond sklearn API calls. — StatQuest with Josh Starmer
Course · beginner · 600 min · developers.google.com
Machine Learning Crash Course
Google's own intro course with interactive exercises on loss, gradient descent, generalization and overfitting -- the vocabulary every ML interview opens with. — Google for Developers
Course · intermediate · 240 min · kaggle.com
Intermediate Machine Learning
Covers missing data, categorical encoding, pipelines, cross-validation, XGBoost and leakage -- exactly the practical gaps freshers get caught on in interviews. — Kaggle Learn
Tools for practice
Choose a tool for the job
Start with one tool for each part of your project. You don’t need to learn them all.
5 tools to explore
pandas
Data
Clean tabular data, join datasets and produce reproducible summaries in Python.
NumPy
Data
Work with numerical arrays and vectorised operations for analysis and modelling.
XGBoost
Data
Train a boosted-tree baseline and compare its performance with simpler models.
Jupyter
Data · Code
Keep code, results and explanations together in an exploratory notebook.
Practice
Credit default risk model with a leak-free scikit-learn pipeline
Predict which borrowers default using the UCI 'Default of Credit Card Clients' dataset — 30,000 real consumer credit records where roughly one in five accounts defaults. Build a single scikit-learn Pipeline that does preprocessing and the model together, so nothing from a test fold leaks into fitting, and compare logistic regression against XGBoost with cross-validated scores. Address the class imbalance explicitly and show the before/after. Report ROC-AUC and a cost-weighted metric — a missed default and a wrongly refused loan are not the same size mistake — instead of raw accuracy.
Start from
Milestones
- Load and profile the data: missing values, categorical encoding, class balance · ~4.5h
- Build the leak-free Pipeline and a logistic-regression baseline with 5-fold CV · ~6.5h
- Add XGBoost, tune it, and test imbalance handling (class weights vs SMOTE vs threshold tuning) · ~6h
- Pick the operating threshold from a cost matrix and write up the metrics table · ~6h
Done when
- Pipeline object (not manual steps) handles preprocessing + model, no leakage from test fold
- At least 2 models compared with cross-validated metrics reported in a table
- Class imbalance explicitly addressed (class weights, SMOTE, or threshold tuning) with before/after metrics
- README explains which metric was optimized for and why, with a confusion matrix on the held-out test set
Prove it
Evidence a recruiter can check
- A cross-validated metrics table (ROC-AUC, precision/recall at the chosen threshold) for logistic regression vs XGBoost, with the imbalance treatment as a third dimension
- The cost matrix you chose and the decision threshold it implies, defended in the README against the default 0.5
- A SHAP summary plot with a paragraph translating the top three features into plain lending language
- The Pipeline definition itself, one object covering preprocessing and model, so a reviewer can see the test fold never touches the fit
Signal it
Built a credit-default risk model on 30k labelled consumer loan records — leak-free scikit-learn pipeline, XGBoost benchmarked against logistic regression with cross-validation, and a decision threshold chosen from a cost matrix rather than accuracy.
Interview
Questions you'll get asked
- Walk me through your ML pipeline for [X] end to end -- what would you change if precision mattered more than recall?
- How do you detect and prevent data leakage in a training pipeline?
- Explain bias-variance tradeoff with an example from a project you built
- When would you choose logistic regression over a random forest, and vice versa?
- How do you handle a severely imbalanced dataset (e.g. 2% fraud rate)?
- What's the difference between L1 and L2 regularization, and when do you use each?
- How would you validate a model before it goes to production?