Medium Machine Learning Interview Questions

Show answer

Bias is error from overly simple assumptions, which causes underfitting. Variance is error from sensitivity to the particular training sample, which causes overfitting. Increasing model complexity usually lowers bias but raises variance. The goal is the complexity that minimises total test error. For example, a linear model on curved data has high bias, while a very deep unpruned tree has high variance.

What is a loss function? Name common ones for regression and classification.

Show answer

A loss function measures how wrong a prediction is for one example, and training minimises its average over the data. Regression commonly uses mean squared error, mean absolute error or Huber loss, which is robust to outliers. Classification uses log loss, also called cross-entropy, and hinge loss for SVMs. The choice should reflect the business cost of errors.

Explain gradient descent and the role of the learning rate.

Show answer

Gradient descent minimises a loss by repeatedly moving parameters in the direction of the negative gradient: w = w - learning_rate * gradient. The learning rate controls step size. Too high and the loss oscillates or diverges; too low and training is very slow or gets stuck. Variants include batch, stochastic and mini-batch gradient descent, plus adaptive optimisers like Adam.

What is the difference between batch, stochastic and mini-batch gradient descent?

Show answer

Batch gradient descent uses the full dataset for each update, giving stable but slow and memory-heavy steps. Stochastic gradient descent updates after every single example, which is fast and noisy and can escape shallow minima. Mini-batch uses small groups such as 32 to 512 rows, balancing stability with speed and allowing vectorised computation. Mini-batch is the default in practice.

What is the curse of dimensionality?

Show answer

As the number of features grows, data becomes sparse in the feature space, distances between points become similar, and the amount of data needed to generalise grows rapidly. Distance-based methods like k-NN and k-means suffer most. Remedies include feature selection, dimensionality reduction such as PCA, regularisation and collecting more data. Text and genomics data often face this problem.

Your model shows 99 percent training accuracy but 72 percent validation accuracy. What will you do?

Show answer

This is overfitting. I would first check for leakage or a split mismatch, then simplify the model: add regularisation, reduce depth or features, use early stopping, or prune trees. I would gather more data or use augmentation if possible and use cross-validation to confirm the fix. Plotting learning curves shows whether more data would actually help.

A business stakeholder asks you to use deep learning for a 5,000-row tabular dataset. How do you respond?

Show answer

I would explain that for small tabular data, gradient-boosted trees or regularised linear models usually match or beat neural networks, train faster and are easier to explain. I would propose building a simple baseline, then a boosting model, and only test a neural network if there is a clear reason. I would show the comparison on cross-validated metrics so the decision is data-driven.

You are asked to predict customer churn. How would you frame the ML problem from scratch?

Show answer

First define churn precisely, for example no purchase in 60 days, and the prediction window. It becomes binary classification. Build features from data available before the cut-off date only, choose a time-based split, and pick metrics like recall at a fixed precision or lift in the top decile. Start with logistic regression as a baseline, then try boosting, and agree on how retention teams will use the scores.

Leadership wants a model in one week, but data quality is poor. How do you handle it?

Show answer

I would agree on a narrow first version: a simple, explainable baseline built on the cleanest available fields, with clear notes on data gaps and expected accuracy. In parallel I would log specific data issues and their impact. Delivering a working baseline with an honest limitations section builds trust and gives a reference point for improvement in later iterations.

How do you implement simple gradient descent for linear regression in NumPy?

Show answer

Compute predictions, the gradient of MSE, and update weights in a loop. import numpy as np w = np.zeros(X.shape[1]); b = 0.0; lr = 0.01 for _ in range(1000): err = X @ w + b - y w -= lr * (2 / len(y)) * (X.T @ err) b -= lr * (2 / len(y)) * err.sum() Scale features first or the loop converges slowly.

How do you make an ML experiment reproducible?

Show answer

Fix random seeds in NumPy, Python and the model, pin library versions in requirements.txt, version the data or record its hash, and save the exact code commit and hyperparameters. Tools like MLflow log parameters, metrics and artifacts per run. Use pipelines so preprocessing is identical each time, and keep the train-test split stored rather than recreated randomly.

Tell me about a time your model did not work as expected. What did you do?

Show answer

Situation: my first churn model scored 0.95 AUC in validation but failed in a pilot. Task: find out why. Action: I traced features and found one field updated after customers left, which leaked the label. I removed it, rebuilt features with a strict cut-off date and moved to a time-based split. Result: AUC dropped to 0.78 but held steady in production, and I added a leakage checklist to my workflow.

What is the difference between label encoding and ordinal encoding?

Show answer

Both map categories to integers. In scikit-learn, LabelEncoder is meant for the target variable, while OrdinalEncoder is for input features. Integer encoding implies an order, so it is fine for truly ordered values like low, medium, high, or for tree models, but it can mislead linear models on nominal data like colours. For nominal features with linear models, prefer one-hot encoding.

What is data leakage? Give two examples.

Show answer

Data leakage is when information not available at prediction time gets into training, giving falsely high validation scores. Example one: fitting a scaler or imputer on the full dataset before splitting. Example two: using a feature created after the event, such as 'refund issued' when predicting fraud. Prevent it with pipelines, time-aware splits and checking each feature's timestamp.

How do you detect and treat outliers?

Show answer

Detect with box plots, the IQR rule (below Q1 - 1.5 IQR or above Q3 + 1.5 IQR), z-scores above 3, or models like Isolation Forest. Treatment depends on cause: fix data entry errors, cap values (winsorise), apply log transforms for skewed data, or use robust models and losses. Genuine extreme values, such as large fraud amounts, may be the most important signal and should not be removed blindly.

Compare filter, wrapper and embedded feature selection methods.

Show answer

Filter methods score features independently of a model using statistics such as correlation, chi-square or mutual information; they are fast but ignore interactions. Wrapper methods like Recursive Feature Elimination train models on feature subsets; they are accurate but slow. Embedded methods select during training, such as L1 regularisation or tree feature importances. Embedded methods are a common practical middle ground.

What is multicollinearity and how do you detect it?

Show answer

Multicollinearity is high correlation between input features, which makes linear model coefficients unstable and hard to interpret, though predictions may remain fine. Detect it with a correlation matrix or the Variance Inflation Factor, where VIF above 5 to 10 signals a problem. Fix it by dropping one of the correlated features, combining them, using PCA, or applying Ridge regularisation.

Why do skewed features often need a log or power transform?

Show answer

Heavily right-skewed features like income or transaction amount have long tails that can dominate linear models and distance metrics. A log1p, Box-Cox or Yeo-Johnson transform compresses large values and makes the distribution closer to normal, which often improves linear and distance-based models. Yeo-Johnson works with zero and negative values. Tree models usually do not need these transforms.

Your dataset has a city column with 2,000 unique values. How will you encode it?

Show answer

One-hot would create 2,000 sparse columns, so I would first group rare cities below a frequency threshold into 'Other'. Then options are frequency encoding, out-of-fold target encoding, or mapping cities to useful attributes such as tier, state or population. For boosting models like CatBoost or LightGBM, native categorical handling also works well. I would compare options with cross-validation.

Validation AUC is 0.99 on a loan default model. What do you suspect?

Show answer

Such a high score on a hard problem usually signals leakage. I would check feature importances for a suspiciously dominant feature, look for fields recorded after the loan outcome such as collection status, verify preprocessing was fit only on training data, and check for duplicate customers across train and test. Then re-run with a time-based split to confirm a realistic score.

40 percent of values in an important income column are missing. What do you do?

Show answer

I would study the missingness pattern first, for example whether self-employed applicants skip it. I would add a binary was_missing feature, impute with the median within relevant groups or use IterativeImputer, and compare against models that handle missing values natively, like HistGradientBoosting or XGBoost. I would also raise the issue with the data owner so collection can improve.

How do you build a scikit-learn pipeline for mixed numeric and categorical data?

Show answer

Use ColumnTransformer inside a Pipeline so preprocessing is fit only on training folds. from sklearn.compose import ColumnTransformer from sklearn.pipeline import Pipeline from sklearn.impute import SimpleImputer from sklearn.preprocessing import StandardScaler, OneHotEncoder num = Pipeline([('imp', SimpleImputer(strategy='median')), ('sc', StandardScaler())]) pre = ColumnTransformer([ ('num', num, num_cols), ('cat', OneHotEncoder(handle_unknown='ignore'), cat_cols) ]) model = Pipeline([('pre', pre), ('clf', LogisticRegression(max_iter=1000))])

How do you calculate VIF for features in Python?

Show answer

Use statsmodels on the numeric feature matrix with a constant added. import pandas as pd from statsmodels.stats.outliers_influence import variance_inflation_factor from statsmodels.tools import add_constant Xc = add_constant(X_num) vif = pd.Series([variance_inflation_factor(Xc.values, i) for i in range(Xc.shape[1])], index=Xc.columns) print(vif.drop('const')) Values above 5 to 10 indicate strong multicollinearity.

How do you select the top k features using mutual information in scikit-learn?

Show answer

SelectKBest with mutual_info_classif captures non-linear relationships, unlike plain correlation. from sklearn.feature_selection import SelectKBest, mutual_info_classif sel = SelectKBest(score_func=mutual_info_classif, k=20) X_new = sel.fit_transform(X_train, y_train) print(X_train.columns[sel.get_support()]) Put the selector inside a pipeline so selection is redone in each CV fold.

Describe a time you found a serious data quality issue late in a project.

Show answer

Situation: two days before a demo, I noticed sales for one region were doubled. Task: fix it without missing the deadline. Action: I traced it to a join that duplicated rows on a non-unique key, fixed the join, added row-count assertions to the pipeline and retrained. I informed my lead immediately with the impact. Result: the demo went ahead with correct numbers, and the assertions caught a similar issue later.

How does ordinary least squares find the regression coefficients?

Show answer

OLS chooses coefficients that minimise the sum of squared residuals between actual and predicted values. It has a closed-form solution, beta = (X^T X)^-1 X^T y, when X^T X is invertible. For very large or ill-conditioned data, gradient descent or a pseudo-inverse is used instead. Squaring errors penalises large mistakes heavily, which makes OLS sensitive to outliers.

What is the difference between Ridge, Lasso and Elastic Net regression?

Show answer

All add a penalty on coefficient size to reduce overfitting. Ridge uses an L2 penalty (sum of squares), shrinking coefficients toward zero but rarely to zero, and handles multicollinearity well. Lasso uses L1 (sum of absolute values) and can set coefficients exactly to zero, performing feature selection. Elastic Net mixes both, which is useful when there are groups of correlated features.

What is polynomial regression and what risk does it carry?

Show answer

Polynomial regression adds powers and interaction terms of features, such as x squared and x1 times x2, so a linear model can fit curves. It is still linear in its coefficients. The main risk is overfitting at high degrees, plus wild behaviour outside the training range. Use cross-validation to choose the degree, scale features, and combine it with Ridge regularisation.

What is heteroscedasticity and why does it matter?

Show answer

Heteroscedasticity means the variance of residuals changes with the level of the prediction, such as errors growing for expensive houses. Predictions remain unbiased, but standard errors and p-values become unreliable. Detect it with a residuals versus fitted plot or the Breusch-Pagan test. Fixes include log-transforming the target, weighted least squares, or robust standard errors.

How do you approach a time series forecasting problem differently from normal regression?

Show answer

Time series data is ordered and autocorrelated, so random splits leak future information. I use time-based splits or TimeSeriesSplit, create lag and rolling-window features from past values only, and handle trend and seasonality. Baselines like last value or seasonal naive come first. Models range from ARIMA and Prophet to gradient boosting on lag features.

Your house price model's residuals grow larger for expensive houses. What do you do?

Show answer

That is heteroscedasticity, typical for price data that is right-skewed. I would model log(price) instead of price, which turns absolute errors into roughly percentage errors and stabilises variance. Then I would convert predictions back with exp and check residual plots again. Evaluating with RMSLE or MAPE also fits better when relative error matters more than absolute error.

A demand forecast is consistently under-predicting during festivals. How would you fix it?

Show answer

The model lacks festival information. I would add features for festival dates such as Diwali or regional holidays, days-to-festival, and past festival uplift for the same item. If the business cost of stock-outs is higher than over-stock, I would use a quantile loss for a higher percentile. Then validate specifically on past festival periods, not just the average error.

Two models have RMSE 120 and 125, but the second has much lower MAE. Which do you choose?

Show answer

It depends on error cost. The first model makes fewer big errors, while the second is better on typical cases but has some large misses. If large errors are expensive, such as stock-outs of costly items, choose the first. If typical accuracy matters more, choose the second. I would also check error distributions by segment and discuss with stakeholders before deciding.

How do you tune the alpha of Ridge regression using cross-validation?

Show answer

RidgeCV tests many alphas efficiently; scale features first. from sklearn.linear_model import RidgeCV from sklearn.pipeline import make_pipeline from sklearn.preprocessing import StandardScaler import numpy as np m = make_pipeline(StandardScaler(), RidgeCV(alphas=np.logspace(-3, 3, 30))) m.fit(X_train, y_train) print(m[-1].alpha_) LassoCV and ElasticNetCV work the same way.

How do you create polynomial features with regularisation in a pipeline?

Show answer

from sklearn.preprocessing import PolynomialFeatures, StandardScaler from sklearn.linear_model import Ridge from sklearn.pipeline import make_pipeline model = make_pipeline( PolynomialFeatures(degree=2, include_bias=False), StandardScaler(), Ridge(alpha=1.0) ) model.fit(X_train, y_train) Tune degree and alpha together with GridSearchCV to avoid overfitting.

Use statsmodels to get p-values for regression coefficients.

Show answer

import statsmodels.api as sm Xc = sm.add_constant(X_train) ols = sm.OLS(y_train, Xc).fit() print(ols.summary()) The summary shows coefficients, standard errors, p-values, confidence intervals, R-squared and diagnostics. scikit-learn focuses on prediction and does not report p-values, so statsmodels is preferred for inference.

Tell me about a time you had to explain a model's error to a business team.

Show answer

Situation: a sales forecast missed by 18 percent in one month. Task: explain why to the category team. Action: I broke the error down by region and found most came from a sudden promotion not in our data. I showed a simple chart and proposed adding a promotion calendar feed. Result: the team shared promotion plans in advance, and the next quarter's error fell below 8 percent.

How do you handle disagreement with a senior colleague about model choice?

Show answer

I keep it about evidence. In one project a senior preferred a complex model while I proposed regularised regression. I suggested we compare both on the same cross-validation folds and also measure inference time and explainability. The complex model was only 1 percent better, so we chose the simpler one. I made sure to credit his idea of the extra features, which we kept.

How does a support vector machine work?

Show answer

An SVM finds the hyperplane that separates classes with the maximum margin, and the points on the margin are the support vectors. The soft-margin version uses C to trade off margin width against misclassification. Kernels like RBF or polynomial let SVMs learn non-linear boundaries by implicitly mapping data to higher dimensions. SVMs need scaled features and become slow beyond roughly a hundred thousand rows.

What is the ROC curve and AUC? When is PR-AUC better?

Show answer

The ROC curve plots true positive rate against false positive rate at every threshold, and AUC is the probability that a random positive is ranked above a random negative. It is threshold-independent. On heavily imbalanced data, like 0.5 percent fraud, ROC-AUC can look high even when precision is poor, so the precision-recall curve and PR-AUC give a more honest picture.

How do you handle class imbalance?

Show answer

Start with the right metrics, such as PR-AUC, recall at fixed precision or F1, not accuracy. Then use class_weight='balanced', resampling like random undersampling or SMOTE applied only on training folds, and threshold tuning based on business cost. Boosting models with scale_pos_weight also help. Stratified splits keep class ratios consistent across folds.

How do one-vs-rest and multinomial approaches differ for multiclass classification?

Show answer

One-vs-rest trains one binary classifier per class against all others and picks the highest score; it is simple but probabilities may not sum to one. Multinomial (softmax) logistic regression models all classes jointly with a single loss, giving consistent probabilities. For metrics, use macro averaging to treat classes equally or weighted averaging to account for class frequency.

The fraud team can only review 500 alerts a day. How do you set the threshold?

Show answer

Instead of 0.5, rank transactions by predicted probability and pick the threshold that yields about 500 alerts per day on recent data. Then report precision and recall at that cut-off, or precision at k. I would monitor daily alert volume because drift can change it, and recompute the threshold periodically. If cost data exists, I would choose the threshold that maximises expected savings.

How do you choose a classification threshold that maximises F1?

Show answer

from sklearn.metrics import precision_recall_curve import numpy as np proba = clf.predict_proba(X_val)[:, 1] p, r, t = precision_recall_curve(y_val, proba) f1 = 2 * p * r / (p + r + 1e-9) best = t[np.argmax(f1[:-1])] print(best) Choose it on validation data, not the test set. scikit-learn 1.5+ also offers TunedThresholdClassifierCV.

How do you build a text classifier with TF-IDF and logistic regression?

Show answer

from sklearn.feature_extraction.text import TfidfVectorizer from sklearn.linear_model import LogisticRegression from sklearn.pipeline import make_pipeline clf = make_pipeline( TfidfVectorizer(ngram_range=(1, 2), min_df=2, sublinear_tf=True), LogisticRegression(max_iter=1000, class_weight='balanced') ) clf.fit(train_texts, train_labels) print(clf.score(test_texts, test_labels))

Tell me about a time you balanced model accuracy against a business constraint.

Show answer

Situation: our lead-scoring model had the best AUC with a large ensemble, but the CRM needed scores within 50 ms. Task: meet latency without losing much quality. Action: I compared a smaller LightGBM model with fewer features and measured both AUC and latency. Result: we lost 0.01 AUC but met the 50 ms limit, and sales teams adopted it because scores appeared instantly.

What is boosting and how is it different from bagging?

Show answer

Boosting trains models sequentially, each one focusing on the errors of the combined previous models, which mainly reduces bias. Bagging trains models independently in parallel and averages them, which mainly reduces variance. Boosting uses shallow trees and a learning rate, and can overfit if run too long, so early stopping is important. Examples are AdaBoost, Gradient Boosting, XGBoost, LightGBM and CatBoost.

How does gradient boosting work?

Show answer

Gradient boosting starts with a simple prediction, like the mean, then repeatedly fits a small tree to the negative gradient of the loss, which for MSE is just the residuals. Each tree's output is scaled by a learning rate and added to the ensemble. A lower learning rate with more trees usually generalises better. Any differentiable loss can be used, which makes it flexible.

What makes XGBoost, LightGBM and CatBoost different?

Show answer

XGBoost adds regularisation to the objective, uses second-order gradients and handles missing values natively. LightGBM grows trees leaf-wise and uses histogram binning, making it very fast on large data, though it can overfit small data. CatBoost handles categorical features with ordered target statistics and uses symmetric trees, giving strong defaults with less tuning. All three support GPU training.

Your random forest model file is 2 GB and too slow for real-time scoring. What can you do?

Show answer

Reduce size by limiting max_depth and min_samples_leaf, using fewer trees since gains flatten after a few hundred, and dropping weak features. Alternatively switch to a gradient boosting model like LightGBM, which usually gives similar accuracy with far smaller, faster models. Converting to ONNX or using a compiled predictor can also speed up inference. Measure accuracy and latency after each change.

XGBoost validation loss starts rising after 300 rounds while training loss keeps falling. What do you do?

Show answer

That is overfitting beyond about 300 rounds. I would use early stopping with a patience such as 50 rounds on a validation set so training stops at the best iteration, then lower the learning rate, reduce max_depth, add subsample and colsample_bytree below 1, and increase min_child_weight or lambda. The final model is refit using the best iteration count found.

Random forest and LightGBM give the same AUC. Which would you deploy?

Show answer

I would compare beyond AUC: model size, inference latency, training time, calibration and stability across time periods. LightGBM is usually smaller and faster to score, while random forest needs less tuning and is less prone to overfitting. If both are equal, I would choose the one that is cheaper to run and easier for the team to maintain and retrain.

Train LightGBM with early stopping.

Show answer

import lightgbm as lgb m = lgb.LGBMClassifier(n_estimators=5000, learning_rate=0.05, num_leaves=31, subsample=0.8, subsample_freq=1, colsample_bytree=0.8) m.fit(X_train, y_train, eval_set=[(X_val, y_val)], eval_metric='auc', callbacks=[lgb.early_stopping(100)]) print(m.best_iteration_) The large n_estimators is only an upper limit; early stopping picks the real number.

Compute permutation importance on validation data.

Show answer

from sklearn.inspection import permutation_importance import pandas as pd r = permutation_importance(model, X_val, y_val, n_repeats=10, scoring='roc_auc', random_state=42, n_jobs=-1) imp = pd.Series(r.importances_mean, index=X_val.columns) print(imp.sort_values(ascending=False)) Near-zero or negative values mean the feature adds little on unseen data.

Describe a project where you improved a model's performance significantly.

Show answer

Situation: a product-return prediction model had 0.68 AUC with logistic regression. Task: improve it for a warehouse pilot. Action: I engineered features like past return rate per customer and category, moved to LightGBM with early stopping, and tuned it with Optuna on time-based folds. Result: AUC rose to 0.81, and the pilot reduced avoidable returns in two categories. I documented each step's gain.

How do you decide when a model is good enough to stop tuning?

Show answer

I agree on a target metric and business threshold before starting, such as recall of 70 percent at 90 percent precision. I stop tuning when gains fall below the noise between CV folds or when further work costs more than the expected value. In one project I stopped at a 0.3 percent gain and spent the time on monitoring instead, which proved more useful.

What is DBSCAN and when is it better than k-means?

Show answer

DBSCAN groups points that lie in dense regions, defined by eps (neighbourhood radius) and min_samples, and labels points in sparse areas as noise. It finds arbitrarily shaped clusters, does not need k, and handles outliers. It struggles when clusters have very different densities and in high dimensions. HDBSCAN improves on it for varying densities.

Explain hierarchical clustering and dendrograms.

Show answer

Agglomerative hierarchical clustering starts with each point as its own cluster and repeatedly merges the two closest clusters. Linkage defines closeness: single, complete, average or Ward, which minimises variance. A dendrogram shows the merge history, and cutting it at a height gives the clusters. It does not need k upfront but scales poorly, roughly O(n squared) memory, so it suits smaller datasets.

What is PCA and how does it work?

Show answer

Principal Component Analysis finds new orthogonal axes, the principal components, that capture maximum variance in the data. It works by computing eigenvectors of the covariance matrix, or via SVD, after centring and usually standardising the data. Keeping the top components that explain, say, 95 percent of variance reduces dimensions, removes correlation and speeds up models, but the components are harder to interpret.

How does Isolation Forest detect anomalies?

Show answer

Isolation Forest builds random trees that split on random features and random thresholds. Anomalies are few and different, so they get isolated in fewer splits, giving shorter average path lengths. The anomaly score comes from that path length. It is fast, works without labels and scales well. The contamination parameter sets the expected share of anomalies and determines the threshold.

What is association rule mining? Explain support, confidence and lift.

Show answer

Association rule mining finds items that occur together, such as in market basket analysis. Support is how often an itemset appears in all transactions. Confidence of A to B is the share of transactions with A that also contain B. Lift is confidence divided by B's support; lift above 1 means A and B occur together more than by chance. Apriori and FP-Growth are common algorithms.

Marketing wants customer segments for campaigns. How do you approach it?

Show answer

I would agree on the purpose first, such as retention or upsell. Then build RFM features (recency, frequency, monetary) plus category preferences, log-transform skewed values and scale them. I would try k-means for k between 3 and 8, use silhouette and business review to pick k, profile each segment and give them clear names. Finally, test segment-based campaigns against a control group.

k-means gives one huge cluster and several tiny ones. What could be wrong?

Show answer

Likely causes are unscaled features, where one large-range feature dominates distance; heavy skew or outliers pulling centroids; or data that is not naturally spherical. I would log-transform and scale features, remove or cap extreme outliers, and try different k values. If shapes are irregular, DBSCAN, HDBSCAN or a Gaussian Mixture Model may fit better.

Your dataset has 800 correlated features and the model trains slowly. What would you do?

Show answer

I would remove constant and near-duplicate features, drop one from each highly correlated pair, then apply PCA inside the pipeline and keep components explaining about 95 percent of variance. Alternatively, use L1-regularised models or tree importances for supervised selection. I would compare speed and cross-validated scores; often 50 to 100 features give nearly the same accuracy.

Apply PCA and check explained variance.

Show answer

from sklearn.decomposition import PCA import numpy as np pca = PCA().fit(Xs) cum = np.cumsum(pca.explained_variance_ratio_) n = np.argmax(cum >= 0.95) + 1 print(n, 'components explain 95 percent') X_red = PCA(n_components=0.95).fit_transform(Xs) Passing a float to n_components keeps enough components to reach that variance.

Detect anomalies with Isolation Forest.

Show answer

from sklearn.ensemble import IsolationForest iso = IsolationForest(n_estimators=300, contamination=0.01, random_state=42) iso.fit(X_train) scores = -iso.score_samples(X_new) flags = iso.predict(X_new) == -1 Higher scores mean more anomalous. Tune contamination to match how many alerts reviewers can handle.

Build RFM features for customer segmentation in pandas.

Show answer

snap = df['order_date'].max() + pd.Timedelta(days=1) rfm = df.groupby('customer_id').agg( recency=('order_date', lambda d: (snap - d.max()).days), frequency=('order_id', 'nunique'), monetary=('amount', 'sum')) rfm[['frequency', 'monetary']] = np.log1p(rfm[['frequency', 'monetary']]) Scale these columns before clustering.

Tell me about a time your analysis was not used by stakeholders. What did you learn?

Show answer

Situation: I built eight customer segments, but marketing did not use them. Task: understand why. Action: I met the team and learned the segments were hard to act on and lacked clear names. I reduced them to five, added a one-page profile and suggested campaign ideas for each. Result: two segments were used in the next campaign. I learned to involve users from the start.

How do you prioritise when several teams ask you for models at the same time?

Show answer

I estimate each request's business value, effort and urgency, then share a ranked list with my manager so priorities are agreed rather than assumed. For quick wins I offer a simple baseline first. In one quarter this helped us deliver a high-value churn model first while giving another team a dashboard-based interim answer, and both teams knew the timeline.

How would you encode cyclical features like hour of day or month?

Show answer

Map the value onto a circle with sine and cosine so the end connects to the start. import numpy as np df['hour_sin'] = np.sin(2 * np.pi * df['hour'] / 24) df['hour_cos'] = np.cos(2 * np.pi * df['hour'] / 24) Now 23:00 and 00:00 are close, which linear and distance-based models cannot see with raw integers. Tree models often manage with the raw value.

When should you use TimeSeriesSplit or GroupKFold instead of normal k-fold?

Show answer

Use TimeSeriesSplit when data is ordered in time, so each validation fold comes after its training data and no future information leaks back. Use GroupKFold when multiple rows belong to the same entity, such as several visits per patient, so the same entity never appears in both train and validation. Normal k-fold in these cases gives over-optimistic scores.

Compare grid search, random search and Bayesian optimisation.

Show answer

Grid search tries every combination of listed values; it is thorough but expensive and wastes trials on unimportant parameters. Random search samples combinations and usually finds good settings faster in high-dimensional spaces. Bayesian optimisation, as in Optuna or Hyperopt, builds a model of past trials to pick promising settings next, using fewer trials. Successive halving methods add early pruning of weak trials.

What are learning curves and validation curves?

Show answer

A learning curve plots training and validation scores against training set size; if both converge at a low score the model underfits, and if a gap remains, more data may help. A validation curve plots scores against one hyperparameter, such as max_depth, showing where overfitting starts. Both help decide whether to collect data, add features or change complexity.

What is the log loss metric and why is it useful?

Show answer

Log loss, or cross-entropy, is the negative average log of the probability assigned to the true class. It heavily penalises confident wrong predictions, such as predicting 0.99 for an event that does not happen. It evaluates probability quality, not just ranking, so it suits problems where probabilities drive decisions. Lower is better, and a constant base-rate prediction is the baseline to beat.

What is the difference between offline evaluation and online A/B testing of a model?

Show answer

Offline evaluation measures metrics like AUC on historical holdout data; it is fast and cheap but cannot capture user behaviour changes or feedback loops. Online A/B testing exposes a share of real traffic to the new model and measures business outcomes like conversion or loss rate. Offline gains do not always translate online, so important models are validated with A/B tests or shadow deployments.

A new model has slightly better AUC but worse precision in the top 1,000 cases the team acts on. Which do you pick?

Show answer

I would pick the model that does better where decisions are actually made. If the team only reviews the top 1,000 cases, precision at 1,000 or recall at that cut-off matters more than global AUC. I would make that the primary metric in future comparisons, and confirm the result holds across several time periods before switching.

Grid search with 5 hyperparameters is taking two days. How do you speed it up?

Show answer

Switch to random search or Optuna with pruning, which find good settings with far fewer trials. Narrow ranges using a quick coarse search, tune the most important parameters first, use early stopping for boosting, run trials in parallel with n_jobs, and tune on a stratified sample of the data before a final refit on the full set. HalvingRandomSearchCV is another quick option.

Your manager asks for a single number showing the model's business value. What do you give?

Show answer

I would convert model outcomes into money or time using a cost matrix agreed with the business, for example savings from caught fraud minus the cost of reviewing false alerts. I would compare this against the current process or a rule-based baseline on holdout data and show a range rather than one point. A short note on assumptions keeps the number honest.

Plot a learning curve in scikit-learn.

Show answer

from sklearn.model_selection import LearningCurveDisplay import numpy as np LearningCurveDisplay.from_estimator(model, X, y, cv=5, train_sizes=np.linspace(0.1, 1.0, 6), scoring='roc_auc', n_jobs=-1) plt.show() A persistent gap between curves suggests more data or regularisation; two low curves suggest a richer model or better features.

Check calibration with a reliability curve and Brier score.

Show answer

from sklearn.calibration import CalibrationDisplay from sklearn.metrics import brier_score_loss proba = clf.predict_proba(X_test)[:, 1] print('Brier', brier_score_loss(y_test, proba)) CalibrationDisplay.from_predictions(y_test, proba, n_bins=10) plt.show() Points on the diagonal mean predicted probabilities match observed frequencies.

Tell me about a time you discovered your own mistake after sharing results.

Show answer

Situation: after presenting a 12 percent uplift from a pricing model, I found my evaluation used a random split on time-series data. Task: correct it quickly. Action: I re-ran with a time-based split, which showed 7 percent, and informed my manager the same day with the corrected analysis. Result: the team appreciated the honesty, the rollout plan used the right number, and I added split checks to our review template.

What is data drift and concept drift?

Show answer

Data drift, or covariate shift, means the distribution of input features changes, such as average order value rising after a pricing change. Concept drift means the relationship between features and target changes, such as fraud patterns evolving. Data drift can be detected without labels using PSI or KS tests; concept drift usually needs labels and shows up as falling performance. Both trigger investigation and retraining.

What is a model registry and why is it useful?

Show answer

A model registry is a central store of model versions with metadata: training data version, parameters, metrics, owner and stage such as staging or production. It makes promotions and rollbacks controlled and auditable, and lets serving systems load a model by name and version. MLflow Model Registry, SageMaker Model Registry and Vertex AI Model Registry are common examples.

Compare batch scoring and real-time inference.

Show answer

Batch scoring runs on a schedule, for example nightly churn scores for all customers written to a table; it is simple, cheap and tolerant of slow models. Real-time inference returns predictions on request, such as fraud checks during payment, and needs low latency, high availability, scaling and fresh features. Choose real-time only when the decision must happen at the moment of the event.

What are shadow deployment and canary release for models?

Show answer

In shadow deployment the new model receives a copy of live traffic and its predictions are logged but not used, so you can compare it with the current model safely. In a canary release a small share of real traffic, such as 5 percent, is served by the new model, and the share grows if metrics stay healthy. Both reduce the risk of a bad model affecting all users.

What should you monitor for a model in production?

Show answer

Monitor service health such as latency, error rate and throughput; input data quality such as missing values, schema changes and out-of-range values; feature and prediction distribution drift; and real performance once labels arrive. Also track business KPIs the model influences. Set alerts with thresholds and owners, and keep a dashboard so issues are found before users report them.

A model's accuracy dropped 10 percent over three months in production. How do you investigate?

Show answer

I would first check data pipelines for broken features, nulls or schema changes, then compare current feature distributions with training data using PSI to find drift. I would check whether the label definition or business process changed, and analyse errors by segment. If drift is real, retrain on recent data, possibly with time-weighted samples, and set up automatic drift alerts and scheduled retraining.

A new model version performs badly right after release. How do you handle it?

Show answer

Roll back immediately to the previous version from the model registry, which should be a one-step action. Then investigate: compare the input features served with those used in validation, check for training-serving skew, version mismatches or a wrong artifact. Add the failure case to automated pre-release checks, and use canary or shadow releases next time so issues affect only a small share of traffic.

Your team retrains models manually in notebooks every month. How would you improve it?

Show answer

I would convert the notebook into a versioned training pipeline with separate steps for data extraction, validation, feature building, training and evaluation, scheduled with Airflow or a cloud pipeline. Each run would log to MLflow, compare against the production model on a fixed holdout, and register the new model only if it passes thresholds. Promotion would stay a reviewed step, at least initially.

Serve a scikit-learn model with FastAPI.

Show answer

from fastapi import FastAPI from pydantic import BaseModel import joblib, pandas as pd app = FastAPI() model = joblib.load('model_v1.joblib') class Item(BaseModel): age: int income: float city: str @app.post('/predict') def predict(item: Item): df = pd.DataFrame([item.model_dump()]) return {'score': float(model.predict_proba(df)[0, 1])} Run it with: uvicorn app:app --host 0.0.0.0 --port 8000

Write a simple Dockerfile for an ML API.

Show answer

FROM python:3.12-slim WORKDIR /app COPY requirements.txt . RUN pip install --no-cache-dir -r requirements.txt COPY app.py model_v1.joblib ./ EXPOSE 8000 CMD ["uvicorn", "app:app", "--host", "0.0.0.0", "--port", "8000"] The exec-form CMD avoids a shell wrapper. Pin versions in requirements.txt so training and serving match.

Log a training run to MLflow.

Show answer

import mlflow, mlflow.sklearn mlflow.set_experiment('churn') with mlflow.start_run(): pipe.fit(X_train, y_train) auc = roc_auc_score(y_val, pipe.predict_proba(X_val)[:, 1]) mlflow.log_params({'C': 1.0, 'model': 'logreg'}) mlflow.log_metric('val_auc', auc) mlflow.sklearn.log_model(pipe, 'model', registered_model_name='churn-model')

Write a unit test for a model prediction function.

Show answer

import joblib, pandas as pd def test_predict_shape_and_range(): model = joblib.load('model_v1.joblib') df = pd.DataFrame([{'age': 30, 'income': 50000.0, 'city': 'Pune'}, {'age': 45, 'income': 0.0, 'city': 'UnknownCity'}]) p = model.predict_proba(df)[:, 1] assert p.shape == (2,) assert ((p >= 0) & (p <= 1)).all() Run with pytest in CI so broken artifacts never reach production.

Tell me about a time you took a model from notebook to production.

Show answer

Situation: our demand model ran only in a notebook on my laptop. Task: make it run daily for 300 stores. Action: I refactored it into modules with a scikit-learn pipeline, added tests and input checks, containerised it with Docker and scheduled it with Airflow writing forecasts to the warehouse. Result: forecasts became available by 6 a.m. daily without manual work, and failures triggered alerts instead of silent gaps.

When would you use AdaBoost and how does it re-weight samples?

Show answer

AdaBoost trains weak learners, usually decision stumps, in sequence. After each round it increases the weights of misclassified samples so the next learner focuses on them, and gives each learner a vote weight based on its accuracy. It works well on clean data but is sensitive to noisy labels and outliers, which get ever larger weights. Gradient boosting is a generalisation that has largely replaced it.

Use HistGradientBoosting with native categorical support in scikit-learn.

Show answer

from sklearn.ensemble import HistGradientBoostingClassifier clf = HistGradientBoostingClassifier( categorical_features='from_dtype', max_iter=500, learning_rate=0.05, early_stopping=True, random_state=42) X_train[cat_cols] = X_train[cat_cols].astype('category') clf.fit(X_train, y_train) It handles missing values natively and is much faster than GradientBoostingClassifier on large data.

A product team wants 'similar products' recommendations with no purchase labels. How would you start?

Show answer

I would build product embeddings from content: TF-IDF or sentence embeddings of titles and descriptions, plus category and price features. Then use cosine similarity with nearest-neighbour search, such as scikit-learn NearestNeighbors or an approximate index like FAISS for large catalogues. I would review samples with the product team and later add co-view or co-purchase signals once collected.

Find the most similar items using cosine similarity.

Show answer

from sklearn.feature_extraction.text import TfidfVectorizer from sklearn.neighbors import NearestNeighbors X = TfidfVectorizer(min_df=2).fit_transform(df['title']) nn = NearestNeighbors(n_neighbors=6, metric='cosine').fit(X) dist, idx = nn.kneighbors(X[0]) print(df['title'].iloc[idx[0][1:]].tolist()) The first neighbour is the item itself, so skip it.

What is regularisation in general and why does it reduce overfitting?

Show answer

Regularisation adds a constraint or penalty that discourages overly complex models, such as large weights in linear models, deep trees or too many boosting rounds. It trades a little extra bias for a larger drop in variance, so test error falls. Examples include L1 and L2 penalties, dropout in neural networks, early stopping, pruning and limiting tree depth.

100

What is the softmax function?

Show answer

Softmax turns a vector of raw scores into probabilities that sum to one: p_i = exp(z_i) / sum of exp(z_j). It is used in multinomial logistic regression and the output layer of neural networks for multiclass problems. Larger scores get exponentially larger shares. In practice the maximum score is subtracted before exponentiation to avoid numerical overflow.

101

Tell me about a time a production model failed and how you communicated it.