Machine Learning interview questions and answers

1. Which model needs feature scaling

One feature is "salary" (range 10,000 to 5,000,000) and another is "age" (range 18 to 70). Which model is most affected if you do NOT scale these features?

1. A.Decision tree
2. B.Random forest
3. C.k-nearest neighbours
4. D.Gradient boosted trees

2. k-nearest neighbours vote

A k-NN classifier with k = 3 finds these nearest training points, sorted by distance: 0.5 (cat), 0.7 (dog), 0.9 (dog), 1.1 (cat), 1.2 (cat). What does it predict?

1. A.cat
2. B.dog
3. C.It is a tie, so no prediction
4. D.cat, since cat is most common among all five

3. Dropout after eval()

What does this PyTorch code print?

import torch
drop = torch.nn.Dropout(p=0.5)
drop.eval()
x = torch.ones(4)
print(drop(x))

1. A.tensor([0.5000, 0.5000, 0.5000, 0.5000])
2. B.A random mix of 0. and 2. values
3. C.tensor([0., 0., 0., 0.])
4. D.tensor([1., 1., 1., 1.])

4. Parameter or hyperparameter

Which of these is a hyperparameter rather than a learned parameter?

1. A.The maximum depth of a decision tree
2. B.The weights of a linear regression
3. C.The split thresholds inside a fitted tree
4. D.The centroids found by k-means

5. Degree-15 curve on 16 points

You fit a degree-15 polynomial to 16 noisy data points. What is the most likely outcome?

1. A.It underfits and gives an almost straight line
2. B.It hits nearly every point and swings wildly between them
3. C.It gives the same result as a degree-1 fit
4. D.It cannot be fit at all with only 16 points

6. What bag-of-words throws away

With a bag-of-words representation, which two reviews would look exactly the same?

1. A."good not bad" and "bad not good"
2. B."good" and "very good"
3. C."bad" and "not bad at all"
4. D."good food" and "good service"

7. A movie that is several genres

Each movie can be tagged with any number of genres: action, comedy, drama and so on. What kind of problem is this?

1. A.Multiclass classification
2. B.Binary classification
3. C.Ordinal regression
4. D.Multi-label classification

8. Why stratify the folds

Only 3% of loan applications in your data are fraudulent. Why use StratifiedKFold instead of plain KFold?

1. A.It makes training faster on each fold
2. B.Each fold keeps about the same 3% fraud rate
3. C.It removes duplicate rows between folds
4. D.It balances classes to 50/50 in each fold

9. MAPE by hand

Forecasts were 110 (actual 100) and 180 (actual 200). What is the MAPE?

1. A.15%
2. B.5%
3. C.10%
4. D.20%

10. R-squared by hand

On a validation set, the residual sum of squares is 20 and the total sum of squares around the mean is 100. What is R²?

1. A.0.2
2. B.0.8
3. C.5
4. D.0.25

11. Top-5 accuracy

An image model scores 72% top-1 accuracy and 91% top-5 accuracy. What does top-5 accuracy measure?

1. A.Accuracy measured on the 5 most common classes
2. B.Average accuracy over 5 random training seeds
3. C.Accuracy on the 5% of images with top confidence
4. D.How often the true label is in the 5 top-scored classes

12. First number to compare against

A new churn model gets 92% accuracy. 91% of customers do not churn. What baseline should you report alongside it?

1. A.A random model: 50% accuracy
2. B.The training accuracy of the same model
3. C.The accuracy of a perfect model: 100%
4. D.Always predicting "no churn": 91% accuracy

13. Score versus tree depth

As max_depth increases, training score keeps rising but validation score peaks at depth 6 and then falls. Which depth should you pick?

1. A.The largest depth tried
2. B.Depth 1
3. C.Any depth; they are equivalent
4. D.About 6

14. Reporting cross-validation results

Five CV folds give AUC 0.81, 0.83, 0.74, 0.82 and 0.80. What is the most informative way to report this?

1. A.Only the best fold, which is 0.83
2. B.Mean and standard deviation over folds
3. C.Only the worst fold, which is 0.74
4. D.Only the first fold, since folds are random

15. Checking a 90% prediction interval

A model gives 90% prediction intervals for delivery time. On 1,000 test orders, 720 actual times fall inside their intervals. What does this show?

1. A.The intervals are too wide
2. B.The intervals are well calibrated
3. C.The point predictions are biased upward
4. D.The intervals are too narrow

16. Precision from a confusion matrix

A classifier gives TP = 40, FP = 10, FN = 20, TN = 130. What is its precision?

1. A.0.67
2. B.0.80
3. C.0.85
4. D.0.20

17. High accuracy on rare events

Only 1% of transactions are fraud. A model that predicts "not fraud" for every transaction scores 99% accuracy. What is the main problem?

1. A.It catches every fraud case
2. B.It has zero recall on the fraud class
3. C.It is overfitting the training data
4. D.Its precision on fraud is 99%

18. High train score, low validation score

A model reaches 99% accuracy on training data but only 70% on validation data. What is most likely happening?

1. A.Underfitting
2. B.Good generalisation
3. C.Overfitting
4. D.Too much regularisation

19. Data used to tune hyperparameters

Which dataset should be used to choose hyperparameters such as learning rate or tree depth?

1. A.The validation set
2. B.The test set
3. C.The training set only
4. D.Live production traffic

20. What training-serving skew is

What is training-serving skew?

1. A.The model is trained on more data than it serves
2. B.Serving uses a bigger machine than training
3. C.Features are computed differently in training than in production, so the model sees different inputs
4. D.The model is served in a different region

21. Handling class imbalance

You train a churn classifier where only 5% of users churn. Which approach is a sound first step?

1. A.Delete all churned users so the model is not confused
2. B.Use class weights (or resampling on the training set only) and tune the decision threshold on validation data
3. C.Oversample the minority class before splitting into train and test
4. D.Report accuracy and stop there

22. Which regularisation gives sparse weights

You want a linear model that automatically sets many feature weights to exactly zero. Which regularisation should you use?

1. A.L2 (ridge) regularisation
2. B.Dropout
3. C.Batch normalisation
4. D.L1 (lasso) regularisation

23. F1 score from a confusion matrix

A fraud model has TP = 40, FP = 10, FN = 20 and TN = 130. What is its F1 score?

1. A.0.67
2. B.0.73
3. C.0.80
4. D.0.85

24. NDCG for a three-item ranking

A search returns three results with relevance 1, 0, 1 in that order. Using gain = relevance and discount 1 / log2(position + 1), what is NDCG@3?

1. A.0.50
2. B.0.67
3. C.0.81
4. D.0.92

25. Loss for logistic regression

Why is logistic regression trained with log loss (cross-entropy) rather than mean squared error on the sigmoid output?

1. A.MSE cannot be computed for values between 0 and 1
2. B.Log loss needs no gradient to be minimised
3. C.MSE always gives the same weights but slower
4. D.Log loss with a sigmoid is convex in the weights; MSE is not

26. Rows left out of a bootstrap

Each tree in a random forest is trained on a bootstrap sample of n rows drawn with replacement from n rows. For large n, roughly what fraction of the original rows are left out of a given tree (out-of-bag)?

1. A.About 50%
2. B.About 10%
3. C.About 37%
4. D.About 63%

27. What bagging and boosting each reduce

Broadly, which statement about bagging and boosting is right?

1. A.Bagging mainly cuts bias; boosting mainly cuts variance
2. B.Both only reduce training time
3. C.Bagging mainly cuts variance; boosting mainly cuts bias
4. D.Both need models that are trained in sequence

28. Why k-means++ is used

Running k-means twice with random starting centres gives quite different clusters. What does k-means++ initialisation do to help?

1. A.Picks spread-out starting centres
2. B.Finds the best value of k automatically
3. C.Guarantees the global optimum every time
4. D.Removes outliers before the clustering starts

29. Encoding hour of day

A demand model uses hour of day as a number 0–23, so 23 and 0 look far apart even though they are one hour apart. What is a common fix for a linear or distance-based model?

1. A.Divide the hour by 24 to scale it
2. B.Take the logarithm of the hour value
3. C.Sort the training rows by hour first
4. D.Add sin(2π·h/24) and cos(2π·h/24) features

30. MAPE on low-selling products

You report MAPE for a demand forecast. Many products sell 0 or 1 unit on some days. What goes wrong?

1. A.MAPE ignores errors on the low sellers
2. B.MAPE turns negative for small values
3. C.MAPE gives the same value as MAE
4. D.Errors explode or are undefined at tiny actuals

31. Which class gets confused

Rows are true labels and columns are predictions. What is the recall for class "refund"?

            billing  refund  other
billing        80      15      5
refund         30      60     10
other           5       5     90

1. A.0.75
2. B.0.80
3. C.0.60
4. D.0.30

32. MRR across three queries

For three queries, the first relevant result appears at ranks 1, 2 and 4. What is the mean reciprocal rank?

1. A.About 0.43
2. B.About 2.33
3. C.About 0.75
4. D.About 0.58

33. Better offline, flat online

A new recommender raises offline NDCG by 4%, but an A/B test shows no change in orders. What is the most reasonable reading?

1. A.The A/B test setup must be broken somehow
2. B.NDCG gains always show up in order counts
3. C.Offline metrics are proxies; trust the test
4. D.Orders are not a valid metric to look at

34. Negative scores from cross_val_score

Why does this print negative numbers?

scores = cross_val_score(reg, X, y, cv=5, scoring='neg_mean_squared_error')
print(scores)  # e.g. [-412.3 -398.1 -455.0 -401.7 -420.9]

1. A.The model is predicting negative values
2. B.Scorers are maximised, so losses are negated
3. C.MSE turns negative when the model is poor
4. D.The folds were not shuffled before splitting

35. Precision and recall at 0.6

Six cases have (score, true label): (0.9, 1), (0.8, 0), (0.7, 1), (0.5, 1), (0.4, 0), (0.2, 0). At threshold 0.6, what are precision and recall?

1. A.Precision 0.67, recall 1.0
2. B.Precision 0.67, recall 0.67
3. C.Precision 1.0, recall 0.67
4. D.Precision 0.5, recall 0.5

36. Reading a MASE of 0.8

A demand forecast has mean absolute scaled error (MASE) of 0.8. What does this mean?

1. A.It is 80% accurate
2. B.80% of forecasts are within 1 unit
3. C.Its MAE is 80% of the naive forecast's MAE
4. D.It is 20% worse than the naive forecast

37. Which prediction hurts log loss most

Each pair is (true label, predicted probability of class 1). Which single prediction adds the most to log loss?

1. A.(1, 0.6)
2. B.(0, 0.3)
3. C.(0, 0.9)
4. D.(1, 0.05)

38. Why a forest beats one tree

Why does a random forest usually generalise better than a single deep decision tree?

1. A.Averaging many decorrelated trees reduces variance
2. B.Each tree in the forest is deeper and more accurate
3. C.Random forests remove bias completely
4. D.Random forests do not need training data

39. Unseen category at predict time

A OneHotEncoder fitted on training cities raises an error when a new city appears in production. What setting avoids the crash?

1. A.sparse_output=False always
2. B.drop="first" on the encoder
3. C.dtype=float on the encoder
4. D.handle_unknown="ignore"

40. Precision moved, AUC did not

With a fixed 0.5 threshold, precision fell sharply this month, but AUC is unchanged. What is a likely explanation?

1. A.Scores shifted, so the fixed cut lands elsewhere
2. B.AUC must be wrong because precision has fallen
3. C.The model has stopped ranking cases correctly
4. D.Precision does not depend on the threshold used

41. ROC-AUC after calibration

You apply isotonic-style monotonic calibration that changes probabilities but keeps the order of scores the same (no new ties). What happens to ROC-AUC?

1. A.It always increases
2. B.It always decreases
3. C.It becomes undefined
4. D.It stays the same, because ROC-AUC depends only on the ranking of scores

42. L2 penalty versus AdamW

Why is AdamW preferred over adding an L2 penalty to the loss when using Adam?

1. A.AdamW does not use momentum, so it is more stable
2. B.With Adam, L2 gradients get rescaled by the adaptive learning rate, so the penalty is uneven; AdamW applies weight decay directly to the weights, decoupled from the gradient
3. C.AdamW removes the need for a learning rate
4. D.L2 penalties are not differentiable

43. A feature that is too good

Your churn model reaches 0.99 AUC. The most important feature is "days_since_last_support_ticket_closed", and many churned users have a final ticket titled "cancel account". What is the most likely issue?

1. A.The model is excellent and ready to ship
2. B.The feature should be one-hot encoded
3. C.The AUC is low because of class imbalance
4. D.Target leakage: the feature contains information created by the churn event itself

44. Target after label smoothing

With nn.CrossEntropyLoss(label_smoothing=0.1) and 5 classes, what target probability does the true class effectively get?

1. A.0.90
2. B.0.98
3. C.0.80
4. D.0.92

45. kNN with 2,000 features

kNN worked well with 10 features but performs badly after you add 2,000 sparse features. What is the main reason?

1. A.In high dimensions distances become similar, so "nearest" means little
2. B.kNN cannot store more than 1,000 features
3. C.Extra features always make the data linearly separable
4. D.k must be at least the number of features

46. RBF kernel with a huge gamma

An RBF-kernel SVM has gamma set very high. Training accuracy is near 100% but validation accuracy is poor. Why?

1. A.Each point affects only a tiny area, so the boundary hugs points
2. B.High gamma turns the RBF kernel into a linear one
3. C.High gamma removes all of the support vectors
4. D.High gamma makes the SVM ignore the training labels

47. Margin width of a linear SVM

A hard-margin linear SVM has weight vector w = (3, 4). What is the total width of the margin between the two supporting hyperplanes w·x + b = ±1?

1. A.0.2
2. B.0.4
3. C.5
4. D.2

48. One long straggly cluster

Agglomerative clustering with single linkage merges two obvious groups into one long cluster because a thin trail of points connects them. Which linkage is less prone to this?

1. A.Single linkage with more points
2. B.Any linkage, after shuffling rows
3. C.Single linkage with Manhattan distance
4. D.Ward or complete linkage

49. Weight given by class_weight balanced

A training set has 1,000 rows: 900 negatives and 100 positives. scikit-learn's class_weight="balanced" sets weight = n_samples / (n_classes × n_samples_in_class). What weight does each positive row get?

1. A.5.0
2. B.9.0
3. C.0.1
4. D.10.0

50. Macro F1 with a weak minority class

Class A (900 samples) has F1 = 0.90 and class B (100 samples) has F1 = 0.50. What is the macro-averaged F1?

1. A.0.70
2. B.0.86
3. C.0.90
4. D.0.50

51. A metric using all four cells

Why is the Matthews correlation coefficient (MCC) often recommended for imbalanced binary problems?

1. A.It uses all four cells; high only if both classes do well
2. B.It ignores true negatives, like F1 does
3. C.It always lies between 0 and 1, like accuracy
4. D.It equals accuracy whenever data is balanced

52. Calibrating with only 300 rows

You must calibrate a classifier using a held-out set of only 300 rows. Why is Platt (sigmoid) scaling often safer than isotonic regression here?

1. A.Isotonic overfits small data; Platt fits two parameters
2. B.Isotonic regression cannot output probabilities
3. C.Platt scaling changes the ranking and so lifts AUC
4. D.Isotonic needs the full training data again to fit

53. Choosing features before cross-validating

With 10,000 random features and 100 rows, you pick the 50 most correlated with the label using all rows, then run 5-fold CV and get 90% accuracy. Labels are actually random. Why?

1. A.Selection saw validation labels, so CV is optimistic
2. B.Five folds are too few for 100 rows
3. C.Correlation is the wrong selection method
4. D.Random labels always give 90% accuracy

54. Comparing two models on one test set

Models A and B are scored on the same 2,000 test rows. Which test focuses on rows where exactly one model is right?

1. A.An unpaired two-sample t-test
2. B.A chi-square test of independence on labels
3. C.McNemar's test
4. D.Kolmogorov–Smirnov on the features

55. Average prediction versus observed rate

On 10,000 recent loans, a default model's mean predicted probability is 6%, but 3% actually defaulted. AUC is unchanged from launch. What is the most likely issue and fix?

1. A.Base rate shifted; recalibrate the intercept
2. B.Ranking broke; retrain from scratch
3. C.Labels leaked; drop the newest features
4. D.Nothing; mean predictions need not match

56. Using attention maps as explanations

A product manager wants to show attention heat maps as the "reason" a model rejected a loan application. What is the main caution?

1. A.Attention weights show where information moved, not why the decision was made
2. B.Attention maps are always random and contain no information
3. C.Attention can only be shown for the first layer of a model
4. D.Heat maps reveal the model's training data, which is private

57. NDCG at 3 for a ranking

A search result list has relevance grades [0, 3, 2] in ranked order. Using DCG = sum of rel_i / log2(i + 1) for positions i = 1, 2, 3, what is NDCG@3 (to two decimals)?

1. A.0.68
2. B.0.50
3. C.1.00
4. D.0.85

58. SMOTE before cross-validation

A team applies SMOTE to the whole dataset, then runs 5-fold cross-validation. CV recall is far higher than in production. Why?

1. A.SMOTE always lowers recall in production
2. B.Cross-validation cannot be used with SMOTE at all
3. C.Five folds are too few for an imbalanced dataset
4. D.Synthetic points made from test rows leak into training folds

59. Computing ECE

A model has two confidence bins. Bin A: 100 predictions, mean confidence 0.9, accuracy 0.8. Bin B: 300 predictions, mean confidence 0.6, accuracy 0.6. What is the expected calibration error?

1. A.0.05
2. B.0.10
3. C.0.0
4. D.0.025

60. Updating a score for a new base rate

A calibrated model trained where 10% of cases were positive outputs 0.50 for a case. The base rate is now 30%, and P(x|y) is unchanged. What is the adjusted probability?

1. A.About 0.50
2. B.About 0.65
3. C.About 0.90
4. D.About 0.79