Machine Learning Engineer
Machine learning questions
133 questions on Machine learning for machine learning engineers, easiest first.
Easy 43
- Big gap between train and validation accuracyeasy
- Which model needs feature scalingeasy
- Encoding a city featureeasy
- Bias and variance as complexity growseasy
- Accuracy on imbalanced dataeasy
- What logistic regression outputseasy
- k-nearest neighbours voteeasy
- What k-means minimiseseasy
- Variance kept by two componentseasy
- Gini impurity of a nodeeasy
- Why random forests average treeseasy
- Dropout after eval()easy
- Standardising a single valueeasy
- When early stopping triggerseasy
- Adding weight decay to an optimisereasy
- Dropout set to 0.9easy
- Predicting a flat price in rupeeseasy
- One gradient descent stepeasy
- Sigmoid at zeroeasy
- Logistic regression with five classeseasy
- The naive part of Naive Bayeseasy
- Entropy of a 50/50 nodeeasy
- Tree with no depth limiteasy
- Prediction after one boosting roundeasy
- Which points shape an SVMeasy
- Label -1 in DBSCAN outputeasy
- Imputing missing incomeeasy
- Scaling a feature with big outlierseasy
- Encoding education leveleasy
- Spot the bug in training codeeasy
- Parameter or hyperparametereasy
- Degree-15 curve on 16 pointseasy
- What bag-of-words throws awayeasy
- Learning from rewardseasy
- 5,000 features, 300 rowseasy
- A movie that is several genreseasy
- Logistic regression on XOReasy
- Baseline for a tabular dataseteasy
- How KNNImputer fills a gapeasy
- What the intercept meanseasy
- A column with one valueeasy
- Batch, stochastic or mini-batcheasy
- Score versus tree deptheasy
Medium 55
- Scaling before splittingmedium
- Handling class imbalancemedium
- Which regularisation gives sparse weightsmedium
- Millions of product IDs as a featuremedium
- Reading a logistic regression weightmedium
- Precision at 5 for a ranked listmedium
- Random search versus grid searchmedium
- Raising C in a soft-margin SVMmedium
- Target encoding without leakagemedium
- Training loss higher than validationmedium
- Closed-form least squaresmedium
- Ridge penalty pushed very highmedium
- When elastic net beats lassomedium
- Loss for logistic regressionmedium
- Shape of a logistic regression boundarymedium
- Why add-one smoothing is usedmedium
- kNN with k = 1medium
- Feature sampling at each splitmedium
- Rows left out of a bootstrapmedium
- What each boosting tree learnsmedium
- Missing values in XGBoostmedium
- What bagging and boosting each reducemedium
- Why k-means++ is usedmedium
- Clusters shaped like ringsmedium
- Soft clusters for customersmedium
- What PCA components aremedium
- Encoding hour of daymedium
- Helping a linear model with a ratiomedium
- How recursive feature elimination worksmedium
- Preprocessing inside a pipelinemedium
- Tuning with Optuna-style searchmedium
- Loss that tolerates a few wild labelsmedium
- Predicting daily order countsmedium
- Changing the 0.5 cut-offmedium
- Soft voting in an ensemblemedium
- How Extra Trees differ from a forestmedium
- IDF of a fairly rare wordmedium
- Using text embeddings in a tabular modelmedium
- Building a lag featuremedium
- Using a pile of unlabelled datamedium
- Choosing what to label nextmedium
- Training on data that never fits in RAMmedium
- Same inputs, different meaningmedium
- How permutation importance worksmedium
- Reading a partial dependence plotmedium
- A silhouette score near zeromedium
- Different seeds, same logistic modelmedium
- Labels that are 10% wrongmedium
- All one-hot columns plus an interceptmedium
- ROC-AUC looks great, alerts are uselessmedium
- Weighted F1 on imbalanced classesmedium
- Hundreds of runs on one validation setmedium
- Negative scores from cross_val_scoremedium
- What best_score_ meansmedium
- Balanced accuracymedium
Hard 35
- L2 penalty versus AdamWhard
- A feature that is too goodhard
- Retraining on your own recommendationshard
- Random ID tops feature importancehard
- SMOTE applied before cross-validationhard
- Scaling inside dropouthard
- Target after label smoothinghard
- Parameters often excluded from weight decayhard
- Rank 256 on 800 exampleshard
- Training loss of 0.02hard
- Why least squares is maximum likelihoodhard
- Spam probability from one wordhard
- kNN with 2,000 featureshard
- Information gain by handhard
- Overfitting in leaf-wise treeshard
- RBF kernel with a huge gammahard
- Margin width of a linear SVMhard
- One long straggly clusterhard
- Reading a t-SNE plothard
- Both errors high and close togetherhard
- How Isolation Forest scores anomalieshard
- New user in collaborative filteringhard
- Weight given by class_weight balancedhard
- Coefficient on a log-price modelhard
- Promising a delivery time to customershard
- Training a stacking meta-modelhard
- Forest forecast that goes flathard
- How CatBoost encodes categorieshard
- A rolling mean that peeks aheadhard
- Adding up SHAP valueshard
- Dropping gender is not enoughhard
- Training a search ranking modelhard
- Kernel SVM on ten million rowshard
- When nested cross-validation is neededhard
- A metric using all four cellshard