The
IBM Machine Learning Professional Certificate on Coursera is a six-course programme: Exploratory Data Analysis, Supervised Learning: Regression, Supervised Learning: Classification, Unsupervised Learning, Deep Learning and Reinforcement Learning, and a recommender-systems capstone. Each course has graded module quizzes. The 90+ solved MCQs below follow the certificate course by course, with the correct option marked and a short explanation under every question. For the deep-learning and generative-AI track, see the
IBM AI Engineering Professional Certificate answers; for a no-code start, see
IBM AI Fundamentals. A
free PDF of the full set is at the end of the page.
🔍
Course 1 — Exploratory Data Analysis for Machine Learning (Questions 1–16)
Data cleaning, outliers, scaling, encoding, transformations and hypothesis testing
Question 01
What is the main goal of Exploratory Data Analysis (EDA)?
ATo deploy the model
BTo understand the data’s structure, patterns and problems before modelling ✅
CTo write the final report
DTo tune hyperparameters
💡 ExplanationEDA uses summaries and visualisations to guide cleaning and feature choices.
Question 02
Which pandas method shows summary statistics for numeric columns?
A.info()
B.describe() ✅
C.head()
D.shape
💡 Explanationdescribe() returns count, mean, std, min, quartiles and max.
Question 03
How can missing values be handled?
AAlways delete the column
BRemove rows, impute with mean/median/mode, or mask and model them ✅
CIgnore them
DReplace with zero only
💡 ExplanationThe right choice depends on how much is missing and why.
Question 04
What is an outlier?
AA missing value
BAn observation far from the rest of the data ✅
CA duplicated row
DA categorical variable
💡 ExplanationDetect with box plots, z-scores or the IQR rule.
Question 05
The IQR rule flags a point as an outlier if it lies:
AWithin Q1 and Q3
BMore than 1.5 x IQR below Q1 or above Q3 ✅
CAbove the mean
DBelow the median
💡 ExplanationIQR = Q3 – Q1.
Question 06
What is feature scaling?
AAdding features
BTransforming features to a comparable range or distribution ✅
CRemoving features
DEncoding labels
💡 ExplanationNeeded for distance-based and gradient-based models.
Question 07
Standard scaling transforms a feature to have:
AMin 0 and max 1
BMean 0 and standard deviation 1 ✅
CMedian 0
DValues between -1 and 1 only
💡 ExplanationStandardScaler subtracts the mean and divides by the standard deviation.
Question 08
Min-max scaling transforms a feature to:
AMean 0, std 1
BA fixed range, usually 0 to 1 ✅
CA normal distribution
DIntegers
💡 ExplanationIt is sensitive to outliers.
Question 09
One-hot encoding is used to:
AScale numeric features
BConvert a categorical variable into binary columns ✅
CRemove outliers
DReduce dimensions
💡 ExplanationEach category becomes its own 0/1 column.
Question 10
Ordinal encoding is appropriate when categories:
AHave no order
BHave a natural order, such as low/medium/high ✅
CAre numeric
DAre missing
💡 ExplanationIt maps ordered categories to integers.
Question 11
A log transformation is commonly applied to:
ANormally distributed data
BRight-skewed data to make it more symmetric ✅
CCategorical data
DBinary data
💡 ExplanationIt compresses large values.
Question 12
What does a correlation coefficient of -0.9 indicate?
ANo relationship
BA strong negative linear relationship ✅
CA weak positive relationship
DCausation
💡 ExplanationCorrelation ranges from -1 to 1 and does not imply causation.
Question 13
What is a null hypothesis?
AThe hypothesis you want to prove
BThe default assumption of no effect or no difference ✅
CA proven fact
DA model parameter
💡 ExplanationTests assess evidence against it.
Question 14
A p-value of 0.01 with a significance level of 0.05 means:
AFail to reject the null
BReject the null hypothesis ✅
CThe null is proven true
DThe test is invalid
💡 ExplanationThe result is statistically significant.
Question 15
A Type I error is:
AFailing to reject a false null
BRejecting a true null hypothesis (false positive) ✅
CA data entry mistake
DOverfitting
💡 ExplanationIts probability is the significance level alpha.
Question 16
Which plot is best for viewing the distribution of a single numeric variable?
AScatter plot
BHistogram ✅
CPair plot
DHeat map
💡 ExplanationBox plots and KDE plots also work.
Next: Supervised Machine Learning: Regression →
📈
Course 2 — Supervised Machine Learning: Regression (Questions 17–32)
Linear and polynomial regression, Ridge, Lasso, Elastic Net, cross-validation and bias-variance
Question 17
In linear regression, the cost function usually minimised is:
ACross-entropy
BMean squared error ✅
CHinge loss
DGini impurity
💡 ExplanationOLS minimises the sum of squared residuals.
Question 18
Polynomial regression is used to:
AReduce features
BModel non-linear relationships by adding powers of features ✅
CClassify data
DCluster data
💡 ExplanationIt is still linear in the coefficients.
Question 19
What is regularisation?
ARemoving data
BAdding a penalty on coefficient size to reduce overfitting ✅
CScaling features
DShuffling data
💡 ExplanationIt trades a little bias for lower variance.
Question 20
Ridge regression uses which penalty?
AL1 (absolute values)
BL2 (squared values) ✅
CNo penalty
DElastic
💡 ExplanationRidge shrinks coefficients toward zero but rarely to exactly zero.
Question 21
Lasso regression uses which penalty?
💡 ExplanationLasso can set coefficients to exactly zero, performing feature selection.
Question 22
Elastic Net combines:
ARidge and decision trees
BL1 and L2 penalties ✅
CLasso and KNN
DPCA and Ridge
💡 ExplanationA mixing parameter balances the two.
Question 23
What is the bias-variance trade-off?
AMore data always helps
BSimple models have high bias; complex models have high variance; the goal is the balance with lowest total error ✅
CBias equals variance
DVariance is always bad
💡 ExplanationUnderfitting is high bias; overfitting is high variance.
Question 24
Why must features be scaled before Ridge or Lasso?
ATo speed up pandas
BBecause the penalty depends on coefficient magnitude, which depends on feature scale ✅
CIt is not needed
DTo remove outliers
💡 ExplanationUnscaled features are penalised unevenly.
Question 25
K-fold cross-validation with K=5 trains the model:
AOnce
BFive times, each time holding out a different fold ✅
CTen times
DOn the test set
💡 ExplanationScores are averaged across folds.
Question 26
What does a learning curve show?
ALoss vs epoch only
BTraining and validation error as training set size increases ✅
CFeature importance
DCorrelation
💡 ExplanationIt helps diagnose bias and variance.
Question 27
What is the purpose of a validation set in hyperparameter tuning?
AFinal evaluation
BChoosing hyperparameters without touching the test set ✅
CTraining
DScaling
💡 ExplanationThe test set is used only once at the end.
Question 28
In scikit-learn, GridSearchCV is used to:
AScale data
BSearch over hyperparameter combinations with cross-validation ✅
CPlot results
DEncode categories
💡 ExplanationIt returns the best estimator and parameters.
Question 29
R-squared can be misleading because:
AIt is always 1
BIt never decreases when more features are added ✅
CIt is negative
DIt ignores the target
💡 ExplanationAdjusted R-squared penalises extra features.
Question 30
What does a residual plot help check?
AClass balance
BWhether errors are randomly scattered (linearity and constant variance) ✅
CCluster shape
DMissing values
💡 ExplanationPatterns suggest a mis-specified model.
Question 31
What does ‘alpha’ control in Ridge and Lasso?
ALearning rate
BStrength of the regularisation penalty ✅
CNumber of folds
DPolynomial degree
💡 ExplanationHigher alpha means stronger shrinkage.
Question 32
Which is an example of a regression error metric robust to outliers?
AMSE
BMean Absolute Error (MAE) ✅
CR-squared
DRMSE
💡 ExplanationMAE does not square the errors.
Next: Supervised Machine Learning: Classification →
🎯
Course 3 — Supervised Machine Learning: Classification (Questions 33–50)
Logistic regression, KNN, SVM, trees, bagging, boosting, stacking and imbalanced classes
Question 33
Logistic regression predicts:
AA continuous value
BThe probability of class membership ✅
CCluster labels
DPrincipal components
💡 ExplanationA threshold (usually 0.5) converts it to a class.
Question 34
What does the ROC curve plot?
APrecision vs recall
BTrue positive rate vs false positive rate at different thresholds ✅
CLoss vs epochs
DAccuracy vs depth
💡 ExplanationAUC summarises it; 0.5 is random, 1 is perfect.
Question 35
The F1 score is the:
AArithmetic mean of precision and recall
BHarmonic mean of precision and recall ✅
CProduct of precision and recall
DDifference between them
💡 ExplanationIt balances the two when classes are imbalanced.
Question 36
Why is accuracy misleading for imbalanced classes?
AIt is hard to compute
BA model can score high by always predicting the majority class ✅
CIt ignores true positives
DIt needs probabilities
💡 ExplanationUse precision, recall, F1 or AUC instead.
Question 37
What does SMOTE do?
ARemoves outliers
BCreates synthetic minority-class samples to balance classes ✅
CScales features
DSelects features
💡 ExplanationSynthetic Minority Over-sampling Technique.
Question 38
Which techniques address class imbalance?
AOversampling, undersampling and class weights ✅
BScaling only
CPCA only
DDropping the target
💡 ExplanationStratified sampling keeps class ratios in splits.
Question 39
In KNN, the ‘K’ is:
AThe number of clusters
BThe number of nearest neighbours considered ✅
CThe number of features
DThe number of trees
💡 ExplanationChosen by cross-validation.
Question 40
In SVM, the regularisation parameter C controls:
AKernel type
BThe trade-off between a wide margin and misclassification ✅
CNumber of support vectors directly
DLearning rate
💡 ExplanationSmall C gives a wider margin and more tolerance of errors.
Question 41
A decision tree splits to maximise:
AEntropy
BInformation gain / reduction in impurity ✅
CDepth
DNumber of leaves
💡 ExplanationGini impurity and entropy are common criteria.
Question 42
How do you reduce overfitting in a decision tree?
AGrow it deeper
BPrune it or limit max depth and minimum samples per leaf ✅
CAdd more features
DRemove the root
💡 ExplanationUnrestricted trees memorise the training data.
Question 43
What is bagging?
ATraining models sequentially on errors
BTraining many models on bootstrap samples and averaging their predictions ✅
CUsing one large tree
DScaling data
💡 ExplanationBootstrap aggregating reduces variance.
Question 44
A random forest improves on bagged trees by:
AUsing deeper trees
BAlso choosing a random subset of features at each split ✅
CUsing fewer trees
DRemoving bootstrapping
💡 ExplanationThis decorrelates the trees.
Question 45
What is boosting?
ATraining trees independently
BTraining models sequentially, each correcting the errors of the previous ones ✅
CAveraging random trees
DRemoving weak learners
💡 ExplanationExamples: AdaBoost, Gradient Boosting, XGBoost.
Question 46
In AdaBoost, misclassified samples are:
ARemoved
BGiven higher weight in the next round ✅
CIgnored
DDuplicated in the test set
💡 ExplanationLater learners focus on hard cases.
Question 47
What is stacking?
ADeep trees
BCombining different models by training a meta-model on their predictions ✅
CBagging with one model
DFeature scaling
💡 ExplanationThe meta-learner learns how to weigh the base models.
Question 48
What does ‘out-of-bag’ error estimate in a random forest?
ATraining error
BGeneralisation error using samples not in each tree’s bootstrap ✅
CFeature count
DTree depth
💡 ExplanationAbout one-third of samples are out-of-bag for each tree.
Question 49
What is model interpretability important for?
ASpeed
BTrust, debugging and regulatory compliance ✅
CStorage
DScaling
💡 ExplanationTools include feature importance, partial dependence, LIME and SHAP.
Question 50
Which model is generally the most interpretable?
AGradient boosting
BA shallow decision tree or logistic regression ✅
CDeep neural network
DRandom forest of 500 trees
💡 ExplanationSimple models are easier to explain.
Next: Unsupervised Machine Learning →
🧩
Course 4 — Unsupervised Machine Learning (Questions 51–66)
K-means, hierarchical clustering, DBSCAN, Mean Shift, PCA, Kernel PCA and NMF
Question 51
K-means requires you to specify:
AThe labels
BThe number of clusters K ✅
CThe tree depth
DThe learning rate
💡 ExplanationK is chosen with the elbow method or silhouette score.
Question 52
What is ‘inertia’ in K-means?
AThe number of iterations
BThe sum of squared distances from points to their cluster centroid ✅
CThe number of clusters
DThe data size
💡 ExplanationLower inertia means tighter clusters.
Question 53
The elbow method helps choose:
AThe learning rate
BThe number of clusters where adding more gives little improvement ✅
CThe kernel
DThe depth
💡 ExplanationPlot inertia against K and look for the bend.
Question 54
K-means++ improves K-means by:
AUsing more clusters
BChoosing initial centroids that are far apart ✅
CRemoving outliers
DScaling data
💡 ExplanationIt leads to faster and better convergence.
Question 55
Hierarchical agglomerative clustering starts with:
AOne big cluster
BEach point as its own cluster and merges the closest pairs ✅
CK centroids
DA density threshold
💡 ExplanationThe result is shown as a dendrogram.
Question 56
Which linkage uses the minimum distance between points of two clusters?
AComplete
BSingle ✅
CAverage
DWard
💡 ExplanationWard linkage minimises the increase in within-cluster variance.
Question 57
DBSCAN groups points based on:
ADistance to centroids
BDensity, using epsilon and minimum samples ✅
CA tree
DProbabilities
💡 ExplanationIt finds arbitrary shapes and labels noise points.
Question 58
An advantage of DBSCAN over K-means is that it:
ANeeds K specified
BCan find non-spherical clusters and identify outliers ✅
CIs always faster
DWorks without parameters
💡 ExplanationIt does not require the number of clusters.
Question 59
Mean Shift clustering finds clusters by:
ASplitting trees
BShifting points toward the densest region (mode) ✅
CRandom assignment
DLinkage
💡 ExplanationThe bandwidth parameter controls the result.
Question 60
What is the ‘curse of dimensionality’?
AToo little data
BAs dimensions grow, data becomes sparse and distances less meaningful ✅
CToo many rows
DSlow computers
💡 ExplanationIt motivates dimensionality reduction.
Question 61
PCA reduces dimensions by:
ARemoving rows
BProjecting data onto directions of maximum variance ✅
CClustering features
DEncoding categories
💡 ExplanationPrincipal components are orthogonal.
Question 62
In PCA, the ‘explained variance ratio’ tells you:
AThe number of clusters
BHow much of the total variance each component captures ✅
CThe error rate
DThe feature names
💡 ExplanationKeep enough components to explain, say, 95%.
Question 63
Why scale data before PCA?
AIt is optional
BFeatures with larger scales would dominate the components ✅
CTo add variance
DTo cluster
💡 ExplanationPCA is variance-based.
Question 64
Kernel PCA is used when:
AData is linear
BThe structure in the data is non-linear ✅
CThere are no features
DData is categorical
💡 ExplanationIt applies the kernel trick before PCA.
Question 65
Non-negative Matrix Factorization (NMF) requires that:
AData be normalised to mean 0
BAll data values be non-negative ✅
CK equals 2
DData be categorical
💡 ExplanationUsed for topic modelling and image decomposition.
Question 66
Which distance metric suits text data represented as word counts?
AEuclidean
BCosine distance ✅
CManhattan
DChebyshev
💡 ExplanationIt compares direction rather than magnitude.
Next: Deep Learning and Reinforcement Learning →
🧠
Course 5 — Deep Learning and Reinforcement Learning (Questions 67–82)
Neural networks, CNNs, RNNs, autoencoders, GANs and Q-learning
Question 67
What is a perceptron?
AA clustering model
BThe simplest neural unit: weighted sum plus a step activation ✅
CAn optimiser
DA loss
💡 ExplanationMultilayer perceptrons stack them with non-linear activations.
Question 68
What does the learning rate control?
ANumber of layers
BThe size of each weight update step ✅
CBatch size
DNumber of epochs
💡 ExplanationToo high diverges; too low is slow.
Question 69
Stochastic gradient descent updates weights using:
AThe whole dataset
BOne sample (or a small batch) at a time ✅
CNo data
DThe test set
💡 ExplanationIt is noisier but faster per update.
Question 70
Which is a regularisation technique for neural networks?
AAdding layers
BDropout, weight decay and early stopping ✅
CLarger learning rate
DMore epochs
💡 ExplanationThey reduce overfitting.
Question 71
A CNN’s convolution layer learns:
AClass labels
BFilters that detect local patterns such as edges ✅
CWord order
DRewards
💡 ExplanationWeights are shared across the image.
Question 72
Transfer learning for images typically reuses:
AThe output layer
BEarly convolutional layers trained on a large dataset ✅
CThe labels
DThe optimiser
💡 ExplanationLater layers are retrained for the new task.
Question 73
An RNN suffers from which problem on long sequences?
AToo many filters
BVanishing or exploding gradients ✅
CToo few parameters
DNo memory
💡 ExplanationLSTM and GRU cells address it.
Question 74
A GRU differs from an LSTM in that it:
AHas more gates
BHas fewer gates and no separate cell state ✅
CCannot handle sequences
DIs not recurrent
💡 ExplanationIt is simpler and often trains faster.
Question 75
An autoencoder’s bottleneck layer:
AIs the output
BHolds the compressed representation of the input ✅
CHolds labels
DIs unused
💡 ExplanationThe decoder reconstructs the input from it.
Question 76
A variational autoencoder (VAE) differs from a standard autoencoder by:
AHaving no decoder
BLearning a probability distribution in the latent space so it can generate new samples ✅
CUsing labels
DBeing a classifier
💡 ExplanationVAEs are generative models.
Question 77
In a GAN, the generator’s goal is to:
AClassify images
BProduce samples the discriminator cannot tell from real data ✅
CReduce dimensions
DLabel data
💡 ExplanationThe two networks are trained adversarially.
Question 78
In reinforcement learning, the ‘policy’ is:
AThe reward
BThe agent’s strategy mapping states to actions ✅
CThe environment
DThe dataset
💡 ExplanationThe aim is to find the policy that maximises expected return.
Question 79
What is the ‘exploration vs exploitation’ trade-off?
ATraining vs testing
BTrying new actions vs using actions known to give reward ✅
CBias vs variance
DSpeed vs memory
💡 ExplanationEpsilon-greedy is a simple strategy.
Question 80
Q-learning learns:
AA classifier
BThe value of taking each action in each state ✅
CA cluster map
DAn embedding
💡 ExplanationIt is model-free and off-policy.
Question 81
The discount factor gamma in reinforcement learning determines:
ALearning rate
BHow much future rewards are valued relative to immediate ones ✅
CThe number of states
DExploration rate
💡 ExplanationGamma near 0 is short-sighted; near 1 is far-sighted.
Question 82
Which library does the certificate use for deep learning labs?
Ascikit-learn only
BKeras (TensorFlow) ✅
COpenCV
DNLTK
💡 ExplanationModels are built with the Keras API.
Next: Machine Learning Capstone (Recommender Systems) →
🏁
Course 6 — Machine Learning Capstone (Recommender Systems) (Questions 83–92)
Content-based and collaborative filtering, matrix factorisation, embeddings and Streamlit
Question 83
What is a recommender system?
AA classifier of images
BA system that suggests items to users based on data about preferences ✅
CA clustering tool
DA search engine
💡 ExplanationThe capstone builds course recommenders.
Question 84
Content-based filtering recommends items:
ALiked by similar users
BSimilar to items the user liked, using item features ✅
CAt random
DMost popular overall
💡 ExplanationIt uses item profiles such as genres or keywords.
Question 85
Collaborative filtering recommends items based on:
AItem descriptions
BBehaviour of similar users or similar items ✅
CRandom choice
DPrice
💡 ExplanationIt needs interaction data, not item features.
Question 86
What is the ‘cold start’ problem?
ASlow servers
BDifficulty recommending for new users or items with no history ✅
CToo much data
DOverfitting
💡 ExplanationContent-based methods help with new items.
Question 87
A user-item interaction matrix is usually:
ADense
BSparse, because each user rates few items ✅
CSquare
DBinary only
💡 ExplanationMatrix factorisation fills in the gaps.
Question 88
Matrix factorisation for recommendations decomposes the interaction matrix into:
AClusters
BUser and item latent-factor matrices ✅
CTrees
DPrincipal labels
💡 ExplanationNMF and SVD-based methods are common.
Question 89
Which similarity measure is common in collaborative filtering?
AHamming
BCosine similarity or Pearson correlation ✅
CEdit distance
DJaccard only
💡 ExplanationThey compare user or item rating vectors.
Question 90
How can a neural network be used for recommendation?
AIt cannot
BBy learning user and item embeddings and predicting ratings from them ✅
COnly for images
DOnly for clustering
💡 ExplanationThe capstone trains an embedding-based model in Keras.
Question 91
Which metric evaluates predicted ratings?
AAccuracy
BRMSE ✅
CSilhouette
DInertia
💡 ExplanationLower RMSE means closer predictions.
Question 92
Which tool does the capstone use to build an interactive app for the recommender?
AExcel
BStreamlit ✅
CPowerPoint
DTableau
💡 ExplanationStreamlit turns Python scripts into web apps.
📥 Download all 90+ Machine Learning certificate MCQs as PDF (free)
🎯 More IBM and AI certification MCQs
Question banks updated regularly to match the latest course content
Leave a Comment