Machine Learning · Analysis

# Part A: Summary view

# Subject card

# Key insights

1. Units 1, 2 and 3 carry most of the paper. Unit 2 (Unsupervised) 128 marks (29%), Unit 1 (Supervised) 111 (25%) and Unit 3 (Evaluation & Ensembles) 106 (24%) across the four papers. Units 4, 5 and the intro block share the rest. Unit 6 (Recent trends) has never been asked.
2. Unit weight swings hard, except Unit 3 (L15). Unit 2 ranged 1% → 46% of a paper (almost absent in 2023-24), Unit 1 9% → 41%, Unit 0 1% → 25%. Unit 3 is the only steady block (21% → 29%) and shows up big every year.
3. Rotation is strong in this subject (L25): 14 of 16 calls correct (88%). Next-exam outlook:
   • Unit 1 (Supervised) ↑ likely UP: it fell to 9% last paper vs a 25% average, the biggest expected riser.
   • Unit 0 (Intro) ↑ likely UP (4% vs 8%). Unit 6 ↑ only because it was absent; it has never been asked, so treat that call as weak.
   • Unit 2 ↓ likely DOWN (46% last paper vs 29%) and Unit 5 ↓.
   • Units 3 and 4 → steady.
4. The latest paper recycled an older one. 2025-26 re-asked 14 of 32 questions (50% of its marks) from earlier papers, mostly straight from 2022-23 (generative mixture model, parametric vs non-parametric, feature selection, curse of dimensionality, matrix factorization, classification vs regression). 23 families have now appeared in 2+ MAKAUT papers: they're the cheapest marks in the subject.
5. It's a theory paper. "Theory" questions carry 58% of the marks and differences 11%. Numericals are only 4% (Euclidean distance, accuracy from a confusion matrix, a Naive Bayes table). Group A one-liners (9%) keep testing the same facts: RL examples, ensemble/boosting/bagging definitions, PCA, deep learning.
6. A compact core covers a lot. 12 topics → 50%, 20 → 70%, 24 → 80% of all marks (out of 50). The intro block and recommender systems (PYQ+) carry 52 marks (12%), so they can't be skipped either.
7. GTU asks some topics far more than MAKAUT does. SVM (1.11: 13 GTU questions, B here), Bayesian learning/MAP (1.6: 9 GTU, never asked by MAKAUT), deep-learning architectures (4.5: 7 GTU) and decision trees (1.4: 14 GTU). MAKAUT can pick these up at any time, and Unit 1 is the likely riser.

# Attack plan (priority order, never a skip list)

1. Tier S first (16 topics), rising units first (L25 tilt, strong back-test):
   • ↑ Unit 1: 1.3 KNN · 1.8 Linear regression · 1.9 Logistic regression
   • ↑ Unit 0: 0.4 Parametric vs non-parametric · 0.2 Types of ML
   • → Unit 3: 3.1 Confusion matrix & metrics · 3.3 Overfitting & bias–variance · 3.7 Bagging & bootstrapping · 3.8 Boosting / bagging vs boosting · 3.6 Ensembles
   • → Unit 4: 4.4 Deep learning (ML vs DL) · 4.3 Neural-network basics
   • ↓ Unit 2: 2.2 K-means · 2.4 Dimensionality reduction & curse · 2.6 PCA
   • ↓ Unit 5: 5.4 Reinforcement learning
2. Every repeated family + the ⚡ facts. Learn the canonical answers in question_bank.md (the 23 MAKAUT-repeated families first), and memorise the Group A facts that keep returning (topics 0.2, 1.8, 1.9, 2.1, 2.6, 3.3, 3.6, 4.3, 4.4, 5.2, 5.4).
3. Tier A, same tilt:
   • ↑ U1: 1.7 Naive Bayes · 1.4 Decision trees/ID3 · 1.1 Classification vs regression · 1.2 Distance measures
   • ↑ U0: 0.1 What is ML, AI vs ML vs DL
   • ↑ U6: 6.1 Recent trends (floor↑)
   • → U3: 3.4 Model selection & cross-validation · 3.9 Random forest
   • → U4: 4.1 Sparse modeling
   • ↓ U2: 2.9 Matrix factorization · 2.10 Generative mixture models · 2.5 Feature selection · 2.8 Recommenders
   • ↓ U5: 5.3 Active learning (floor↑)
4. Per-unit floor check (L15): every syllabus unit already has its must-dos above. U1: 1.3, 1.8 · U2: 2.2, 2.4 · U3: 3.1, 3.3 · U4: 4.4, 4.1 · U5: 5.4, 5.3 · U6: 6.1.
5. Then tier B, Unit 1 first because it's the likely riser: 1.11 SVM and 1.12 kernels (the GTU favourites) · 1.5 Bayes theorem → then 0.3, 2.1, 2.7, 3.5, 4.2, 4.6, 5.1, 5.2, 5.5.
6. Then tier C (never asked by MAKAUT yet), Unit 1 first: 1.6 Bayesian learning/MAP · 1.10 GLM · 1.13 Multi-class & ranking → 2.3, 2.11, 3.2, 4.5, 6.2. These are prepared to the same standard from the syllabus and GTU angles (L4).
7. Full-marks standard (L23): the order sets priority, not a cut-off. Every topic in the map is prepared so that any question the paper offers, including all choices, can be answered.

Rendered by tools/analyze.py. Basis: 4 counted papers (CORE + R-OLD) · mode STANDARD · marks = printed paper marks incl. all choices (448 in total, 112 · 112 · 112 · 112 per paper).

# Unit importance (historical evidence, NOT a fixed weightage: L15)

# Unit volatility: marks per unit in each paper (% of that paper)

# Unit rotation check (L25): does a heavy unit drop next year, and a light unit rise?

Rule tested: a unit more than 3 points above its average share → expect it to go DOWN in the next paper; more than 3 below → expect UP. ✓ = it happened, ✗ = it didn't.

Back-test: 14 of 16 calls correct (88%): the rotation pattern holds strongly in this subject's papers. Caveat: the average includes the years being tested, so with few papers this flatters the hit rate a little.

# Next-exam unit outlook (L25: from the latest paper, 2025-26, vs each unit's average)

Outlook = priority tilt only. Every unit is still fully prepared (L15, L23): MAKAUT's split stays partly random, and confidence in the tilt follows the back-test above (88% on 16 calls).

# Tier S and A topics

Tier counts: S 16 · A 14 · B 12 · C 8

# Top repeated question families

# Repeat rate: how much each paper re-asks earlier papers

# Coverage curve (topics ranked by marks)

12 topics → 50% of all marks · 20 topics → 70% of all marks · 24 topics → 80% of all marks · 30 topics → 90% of all marks (out of 50 topics). Historical only; every unit's floor topics still apply (L15).

# Question-type mix

# Syllabus vs extra (by tag)

# Part B: Detailed view

Rendered by tools/analyze.py. Columns per evidence class are kept separate (L9); – = no papers of that class.

Flags: floor↑ promoted by the per-unit floor (L15) · ⚡facts asked as 1–2-mark facts in ≥2 papers · recent asked in the latest paper · +sec also appears as a secondary part of another topic's question.

# Unit 0: Introduction to Machine Learning (PYQ+ block)

# Unit 1: Supervised Learning (Regression/Classification)

# Unit 2: Unsupervised Learning

# Unit 3: Evaluating Machine Learning algorithms and Model Selection

# Unit 4: Sparse Modeling and Estimation

# Unit 5: Scalable Machine Learning

# Unit 6: Recent trends in various learning techniques of machine learning and classification methods

# Heatmap: marks per topic per year (· = appears only as a secondary part)

# All repeated families

• F-5.4-02 (5.4) · Identify reinforcement learning from an application / give a practical example of RL. · 3/4 papers
  • 2022-23 A·1(IX) [1]: Imagine a Newly-Born starts to learn walking. It will try to find a suitable policy to learn walking after repeated falling and getting up. Specify what type of machine learning is best suited?
  • 2023-24 A·1(IV) [1]: Real-Time decisions, Game AI, Learning Tasks, Skill Aquisition, and Robot Navigation are applications in _____.
  • 2025-26 A·1(III) [1]: Give an practical example of reinforcement learning.
• F-2.5-01 (2.5) · Explain feature selection: its importance/benefits and its techniques. · 2/4 papers
  • 2022-23 C·10(a) [6]: Explain the three techniques under supervised feature Selection
  • 2022-23 C·10(b) [4]: Explain the benefits of using feature selection in machine learning
  • 2025-26 C·7(a) [6]: Explain the importance of feature selection in supervised machine learning. What techniques can be used for feature selection and what factors should be considered when choosing features?
• F-2.9-01 (2.9) · What is matrix factorization, and how does it work in (movie) recommendation systems? · 2/4 papers
  • 2022-23 B·2 [5]: Explain Matrix Factorization and where it is uesd.
  • 2025-26 C·10(a) [7]: Matrix factorization plays a significant role in collaborative filtering-based recommendation systems. Explain how matrix factorization works in the context of recommendation systems.
• F-0.4-01 (0.4) · What are parametric and non-parametric models? How do they differ? · 2/4 papers
  • 2022-23 C·7(b) [5]: What are parametric and non-parametric model?
  • 2025-26 C·9(c) [5]: What are parametric and non-parametric model.
• F-2.10-02 (2.10) · With a proper diagram, explain the steps of a generative mixture model. · 2/4 papers
  • 2022-23 C·8(b) [5]: With a proper diagram explain the steps of a genarative mixture model
  • 2025-26 C·9(b) [5]: With a proper diagram explain the steps of a generative mixture model.
• F-2.4-01 (2.4) · Explain the curse of dimensionality. · 2/4 papers
  • 2022-23 C·10(c) [5]: Explain the curse of dimensionality
  • 2025-26 C·8(b) [5]: Discuss problems associated with high-dimensional data, highlighting the concept of the "curse of dimensionality".
• F-3.1-02 (3.1) · Explain TP, TN, FP and FN, and the significance of FP vs FN, with an example. · 2/4 papers
  • 2022-23 C·9(c) [5]: Explain True Positive, True Negative, False Positive, and False Negative in Confusion Matrix with an example.
  • 2023-24 C·7(a) [5]: What Is a False Positive and False Negative and How are they significant with an example?
• F-3.3-01 (3.3) · How can overfitting (and underfitting) be tackled or avoided? · 2/4 papers
  • 2023-24 B·2 [5]: How to Tackle Overfitting and Underfitting?
  • 2025-26 B·6 [5]: Explain the concept of overfitting in machine learning. How can techniques like cross-validation help in mitigating it?
• F-4.1-01 (4.1) · Explain sparse modeling and estimation: its goal, benefits, and how it differs from traditional modeling. · 2/4 papers
  • 2023-24 C·8(a) [5]: Explain: Sparse Modelling and estimation.
  • 2025-26 B·4 [5]: What is the primary goal of sparse modeling and how does it differ from traditional modeling approaches?
• F-4.4-02 (4.4) · How is deep learning different from machine learning? · 2/4 papers
  • 2023-24 B·4 [5]: Differentiate Between Machine Learning and Deep Learning.
  • 2024-25 B·3 [5]: Discuss working of Nueral network. State two differences between Deep learning and Machine learning.
• F-1.1-01 (1.1) · Differentiate classification and regression, with examples. · 2/4 papers
  • 2022-23 B·4 [5]: Explain the Difference Between Classification and Regression?
  • 2025-26 C·7(b) [4]: Compare and contrast classification and regression tasks in supervised machine learning. Provide examples of each.
• F-2.10-01 (2.10) · Explain the generative mixture model / Gaussian Mixture Model. · 2/4 papers
  • 2022-23 C·8(a) [4]: Explain Generative Mixture model
  • 2025-26 C·9(a) [5]: Explain Generative Mixture model.
• F-3.1-01 (3.1) · What is a confusion matrix, and why do we need it? · 2/4 papers
  • 2022-23 C·9(a) [5]: Explain the Confusion Matrix with Respect to Machine Learning Algorithms with an suitable example
  • 2023-24 C·7(b) [4]: Why do you need Confusion matrix?
• F-1.7-02 (1.7) · Why is Naive Bayes called "naive"? (class-conditional independence) · 2/4 papers
  • 2023-24 B·6 [5]: Explain ‘Naive’ in a Naive Bayes Algorithm.
  • 2024-25 C·11(b) [3]: Why it is called Naïve Bayes ?
• F-1.7-03 (1.7) · Apply Naive Bayes to a given dataset and predict the class of a new instance. · 2/4 papers
  • 2023-24 C·7(c) [6]: For the given dataset, apply Naive Baye's Algorithm and predict the outcome for a car= {Red, Domestic, SUV}. [figure: no car dataset is printed. The image under this question instead shows X(z) = (1 + ½z⁻¹) / (1 − ½z⁻¹) with the line "Determine the x(n) using time −shifting property." (a Z-transform item, unnumbered, no marks)]
  • 2024-25 C·p2 [0]: Using these probabilities estimate the probability values for the new instance (Color=Green,legs =2,Height=Tall,and Smelly=No). [data below]
• F-3.4-02 (3.4) · Explain cross-validation (k-fold) and its use. · 2/4 papers
  • 2023-24 C·11(b) [5]: Explain Cross validation with an example.
  • 2025-26 A·1(X) [1]: What is the use of cross validation.
• F-3.7-01 (3.7) · How does bagging reduce variance and overfitting? (Which ensemble reduces variance?) · 2/4 papers
  • 2023-24 A·1(X) [1]: Which ensemble model helps in reducing variance?
  • 2025-26 C·11(a) [5]: Define what bagging is and explain its core principle. Describe how it reduces overfitting and variance in machine learning models.
• F-4.4-01 (4.4) · What is deep learning? Explain it with real-world examples. · 2/4 papers
  • 2022-23 C·11(b) [4]: What is Deep Learning, and give some of its example that is used in real-world?
  • 2024-25 A·1(III) [1]: Define deep learning.
  • 2024-25 A·1(XII) [1]: Give one real life example of deep learning.
• F-0.2-03 (0.2) · What is unsupervised learning, and what is its goal? · 2/4 papers
  • 2023-24 A·1(III) [1]: The prior goal of unsupervised learning model is to determine the ________ .
  • 2024-25 A·1(I) [1]: What is unsupervised learning?
• F-1.9-03 (1.9) · Logistic regression is a classification algorithm: for which type of classification (binary)? · 2/4 papers
  • 2022-23 A·1(I) [1]: ______ is a classification algorithm used to assign observations to a discrete set of classes.
  • 2025-26 A·1(IV) [1]: Logistic regression is generally used for which type of classification?
• F-2.6-02 (2.6) · What is PCA (what does PCA stand for)? · 2/4 papers
  • 2024-25 A·1(VI) [1]: What is PCA?
  • 2025-26 A·1(V) [1]: What does PCA stands for?
• F-3.6-01 (3.6) · What is ensemble learning (an ensemble model)? · 2/4 papers
  • 2022-23 A·1(V) [1]: A collection of individual models that learn to predict a target by combining their strengths and avoiding the weaknesses of each is called ________
  • 2024-25 A·1(II) [1]: What is an ensemble model in machine learning?
• F-5.2-01 (5.2) · What is semi-supervised learning, and how does it differ from supervised and unsupervised learning? · 2/4 papers
  • 2022-23 A·1(VI) [1]: Semi-supervised learning algorithm deals with which types of data
  • 2024-25 A·1(IX) [1]: What is semisupervised learning in machine learning?
• F-1.4-01 (1.4) · Explain how a decision tree is built with ID3, using entropy and information gain (with an example). · 1/4 papers
  • 2023-24 C·9(a) [4]: Explain the ID3 algorithm for decision tree learning.
  • 2023-24 C·9(b) [5]: What is the procedure of building Decision tree using ID3 with Gain and Entropy.
  • 2023-24 C·9(c) [6]: What do you mean by Information gain and Entropy and its relation with ID3.
• F-1.3-01 (1.3) · Explain the KNN algorithm with an example. · 1/4 papers
  • 2024-25 C·11(c) [8]: Explain KNN algorithm with suitable example.
• F-0.2-01 (0.2) · Explain the types of machine learning (supervised, unsupervised, reinforcement) and how they differ. · 1/4 papers
  • 2022-23 C·7(a) [6]: Discuss the different types of Machine Learning?
• F-1.2-01 (1.2) · What are proximity (distance) measures? Explain them with formulas, e.g. Euclidean distance. · 1/4 papers
  • 2023-24 C·10(c) [6]: What is Euclidean distance in terms of machine learning? Using Euclidean distance find the distance between points P(3, 2) and Q(4, 1).
• F-2.6-01 (2.6) · Explain PCA and its steps for dimensionality reduction. · 1/4 papers
  • 2022-23 C·8(c) [6]: Write down the steps of PCA (Principal Component Analysis)
• F-3.8-01 (3.8) · Compare bagging and boosting. · 1/4 papers
  • 2024-25 C·9(a) [6]: Compare between Bagging & Boosting technique.
• F-1.3-02 (1.3) · How do you choose K in KNN (odd K, larger K for noisy data)? · 1/4 papers
  • 2022-23 B·6 [5]: How do we decide the value of "K" in KNN algorithm? Why is the odd value of “K” preferable in KNN algorithm?
• F-1.5-01 (1.5) · Explain Bayes theorem with an example. · 1/4 papers
  • 2024-25 B·4 [5]: Define conditional probability. Discuss Bayes theorem with suitable example.
• F-1.8-01 (1.8) · Explain linear regression with an example (best-fit line, prediction). · 1/4 papers
  • 2025-26 C·7(c) [5]: Describe linear regression as a supervised learning algorithm. How is the best-fit line determined and how are predictions made with it?
• F-1.9-01 (1.9) · Differentiate linear regression and logistic regression, with examples. · 1/4 papers
  • 2024-25 B·5 [5]: Compare linear regression with logistic regression. Write two applications of logistic regression.
• F-2.1-02 (2.1) · Differentiate clustering and classification. · 1/4 papers
  • 2024-25 C·8(b) [5]: How does the clustering differ from classification in machine learning?
• F-2.2-01 (2.2) · Explain the K-means clustering algorithm (with an example). · 1/4 papers
  • 2025-26 B·2 [5]: Describe the K-means clustering algorithm.
• F-2.2-02 (2.2) · What are the weaknesses of K-means (outliers, non-linear clusters, fixed k)? · 1/4 papers
  • 2024-25 C·8(c) [5]: Describe the strengths and weaknesses of the K-Means clustering algorithm. When is it suitable for use?
• F-2.4-02 (2.4) · What is dimensionality reduction, and why is it important (benefits)? · 1/4 papers
  • 2025-26 C·8(a) [5]: Define dimensionality reduction and explain its significance in data processing.
• F-2.7-01 (2.7) · Explain kernel PCA and how it overcomes PCA's weakness on non-linear data. · 1/4 papers
  • 2025-26 C·8(c) [5]: Explain the relationship between the variance of the data and the principal components derived using PCA. Discuss a limitation of PCA when it comes to capturing non-linear relationships in the data.
• F-2.9-02 (2.9) · Explain matrix completion (and its challenges with sparse data). · 1/4 papers
  • 2025-26 C·10(b) [5]: Discuss the challenges of matrix completion when dealing with sparse user-item interaction data.
• F-3.1-03 (3.1) · Compute accuracy and other metrics from a given confusion matrix (numerical). · 1/4 papers
  • 2022-23 C·9(b) [5]: Calculate the accuracy percentage for the given Confusion Matrix. [data below]
• F-3.1-04 (3.1) · Explain the evaluation parameters of a classifier: precision, recall, F1 and others. · 1/4 papers
  • 2024-25 B·2 [5]: Explain Precision, Recall and F1 Score.
• F-3.3-03 (3.3) · Explain bias vs variance and the bias–variance trade-off. · 1/4 papers
  • 2024-25 C·9(b) [5]: State the Difference Between Bias and variance in ML.
• F-3.4-01 (3.4) · What is model selection? Explain the procedure/techniques of model selection and evaluation. · 1/4 papers
  • 2023-24 C·11(a) [5]: What is model selection? Write down the types of data and based on that which model is used?
• F-3.5-01 (3.5) · Explain statistical learning theory, in particular the PAC learning model. · 1/4 papers
  • 2023-24 C·11(c) [5]: Explain Statistical Learning theory.
• F-3.6-02 (3.6) · Why are ensemble methods used and how do they improve accuracy? Explain the principle and types. · 1/4 papers
  • 2022-23 B·3 [5]: Why ensemble learning is used? What is the general principle of an ensemble method and what is bagging and boosting in ensemble method?
• F-3.7-02 (3.7) · What is bagging? Explain its core principle and steps. · 1/4 papers
  • 2025-26 C·11(a) [5]: Define what bagging is and explain its core principle. Describe how it reduces overfitting and variance in machine learning models.
• F-3.7-03 (3.7) · Explain bootstrapping and how it is used in bagging. · 1/4 papers
  • 2025-26 C·11(c) [5]: Discuss the concept of bootstrapping in bagging and explain how it contributes to model diversity.
• F-3.9-02 (3.9) · What is the benefit of a random forest over a single decision tree? · 1/4 papers
  • 2025-26 B·3 [5]: What is the benefit of Random forest classifier over individual decision trees.
• F-4.2-01 (4.2) · Explain time-series data, time-series analysis/forecasting, and the models used. · 1/4 papers
  • 2023-24 C·8(b) [5]: Explain: Time series data in machine learning
• F-4.6-01 (4.6) · Explain feature representation learning and how it relates to deep learning. · 1/4 papers
  • 2023-24 C·8(c) [5]: Explain how deep learning and feature representation learning related.
• F-5.5-01 (5.5) · Explain the Bayesian (belief) network with an example. · 1/4 papers
  • 2025-26 B·5 [5]: Define a Bayesian network and explain how it can be used to represent probabilistic relationships between a set of variables.
• F-1.12-01 (1.12) · Why does a linear SVM fail on non-linear data, and how do kernels (non-linear SVM) fix it? · 1/4 papers
  • 2024-25 C·7(b) [4]: What is the difference between a linear and non-linear SVM?
• F-1.7-01 (1.7) · Explain the Naive Bayes classifier with an example. · 1/4 papers
  • 2024-25 C·11(a) [4]: Discuss Naïve Bayes Classifier algorithm.
• F-3.9-01 (3.9) · Explain the random forest algorithm. · 1/4 papers
  • 2024-25 C·9(c) [4]: Write short note on Random Forest algorithm.
• F-1.3-03 (1.3) · Why is KNN non-parametric / instance-based (lazy) learning? · 1/4 papers
  • 2023-24 C·10(a) [3]: Why KNN is non-parametric?
• F-2.1-01 (2.1) · What is clustering? · 1/4 papers
  • 2024-25 A·1(X) [1]: What do you mean by Clustering?
• F-3.3-02 (3.3) · Differentiate overfitting and underfitting. · 1/4 papers
  • 2024-25 A·1(VIII) [1]: Write down the difference between under fitting and over fitting.
• F-3.8-02 (3.8) · What is the advantage of boosting over bagging? · 1/4 papers
  • 2025-26 A·1(XII) [1]: In ensemble methods, what is the main advantage of boosting over bagging?
• F-3.8-04 (3.8) · What is boosting (models trained in sequence, each correcting the previous one's errors)? · 1/4 papers
  • 2025-26 A·1(VI) [1]: Which ensemble method involves training multiple models in such a way that each model corrects the errors of the previous model?
• F-5.1-01 (5.1) · Why does scalability matter, and how does distributed learning make ML scalable? · 1/4 papers
  • 2025-26 A·1(VIII) [1]: How does distributed learning allow for scalability in training large datasets?
• F-5.4-01 (5.4) · What is reinforcement learning? Explain it (short note). · 1/4 papers
  • 2024-25 A·1(IV) [1]: Define Reinforcement learning.
• F-0.1-01 (0.1) · What is machine learning? Explain it with examples and applications. · 0/4 papers
• F-0.1-02 (0.1) · What are the steps to develop a machine learning application? · 0/4 papers
• F-0.1-03 (0.1) · What are the objectives, design issues and approaches of machine learning? · 0/4 papers
• F-0.2-02 (0.2) · Differentiate supervised and unsupervised learning, with examples. · 0/4 papers
• F-1.1-02 (1.1) · List the common classifiers, and explain KNN and SVM. · 0/4 papers
• F-1.10-01 (1.10) · Explain Generalized Linear Models (GLM). · 0/4 papers
• F-1.11-01 (1.11) · Explain how an SVM works, with a diagram (advantages and disadvantages). · 0/4 papers
• F-1.11-02 (1.11) · Explain the maximum-margin (large-margin) principle / optimal hyperplane of an SVM. · 0/4 papers
• F-1.13-01 (1.13) · Differentiate binary and multi-class classification, and explain how multi-class problems are handled. · 0/4 papers
• F-1.4-02 (1.4) · What are the issues (limitations) of decision tree learning, and how are they overcome? · 0/4 papers
• F-1.4-03 (1.4) · Explain CART with an example. · 0/4 papers
• F-1.6-01 (1.6) · What is the MAP hypothesis? Explain the brute-force MAP learner (and the MDL principle). · 0/4 papers
• F-1.6-02 (1.6) · Explain Bayesian learning and inference with an example. · 0/4 papers
• F-1.8-02 (1.8) · Explain multivariate (multiple) linear regression. · 0/4 papers
• F-1.9-02 (1.9) · Explain logistic regression in detail (with an example, assumptions). · 0/4 papers
• F-2.1-03 (2.1) · Differentiate hard and soft clustering. Which suits real-world data better? · 0/4 papers
• F-2.10-03 (2.10) · Differentiate generative and discriminative (predictive) models. · 0/4 papers
• F-2.2-03 (2.2) · Cluster the given points with K-means (numerical). · 0/4 papers
• F-2.3-01 (2.3) · Explain kernel K-means and its advantages over K-means. · 0/4 papers
• F-3.2-01 (3.2) · Explain the ROC curve and AUC. · 0/4 papers
• F-3.8-03 (3.8) · Explain the AdaBoost algorithm with an example (advantages, disadvantages). · 0/4 papers
• F-4.3-01 (4.3) · Explain the single perceptron. Can it solve the XOR problem? Justify. · 0/4 papers
• F-4.3-02 (4.3) · Explain the different types of neural networks. · 0/4 papers
• F-4.3-03 (4.3) · What is backpropagation? How does a neural network reduce error with it? · 0/4 papers
• F-4.3-04 (4.3) · What are activation functions? Explain the common ones and their role. · 0/4 papers
• F-4.4-03 (4.4) · Deep learning extracts features by itself, so is it a "black box"? Justify. · 0/4 papers
• F-4.5-01 (4.5) · Explain the Deep Belief Network, with equations and features. · 0/4 papers
• F-4.5-02 (4.5) · Explain any one deep learning model in detail (e.g. CNN). · 0/4 papers
• F-5.2-02 (5.2) · Explain (and differentiate) semi-supervised learning, active learning and reinforcement learning. · 0/4 papers
• F-5.3-01 (5.3) · Explain active learning (and online learning) with an example. · 0/4 papers
• F-5.4-03 (5.4) · How does an RL agent learn, moving between states with the help of rewards? · 0/4 papers
• F-5.4-04 (5.4) · How does reinforcement learning differ from supervised learning? · 0/4 papers
• F-6.1-01 (6.1) · Discuss recent trends in machine learning (incl. IoT applications, LLMs and generative AI). · 0/4 papers

# Never-asked topics (still fully prepared from the syllabus: L4)

• Unit 1: 1.6 Bayesian Learning & Inference · 1.10 Generalized Linear Models · 1.13 Beyond Binary Classification: Multi-class, Structured Outputs & Ranking
• Unit 2: 2.3 Kernel K-means · 2.11 Latent Factor Models
• Unit 3: 3.2 ROC, AUC & Other Evaluation Metrics
• Unit 4: 4.5 Deep Learning Architectures: CNN, RNN, Autoencoders, DBN
• Unit 6: 6.1 Recent Trends in Learning Techniques · 6.2 Recent Trends in Classification Methods

# Excluded questions

• PROXY-GTU-3710216-2019-W#b16: OUT-PROXY
• PROXY-GTU-3710216-2022-S#b17: OUT-PROXY
• PROXY-GTU-3710216-2023-W#b16: OUT-PROXY
• PROXY-GTU-3710216-2023-W#b17: OUT-PROXY
• PROXY-GTU-3735904-2020-W#b13: OUT-PROXY
• PROXY-GTU-3735904-2021-W#b14: OUT-PROXY
• PROXY-GTU-3735904-2021-W#b15: OUT-PROXY
• PROXY-GTU-ME01000631-2025-S#b15: OUT-PROXY
• PROXY-GTU-ME01000631-2025-W#b17: OUT-PROXY
• PROXY: 304 in-syllabus rows from 18 papers: mapped to topics and shown in their own column, never counted (L16).

# Confidence & caveats

• 4 MAKAUT papers (STANDARD). Every count rests on four papers, so a topic asked once shows as 1/4 = 25%. The tiers are a solid guide, not a law. GTU papers add confidence about what can be asked, never about how often MAKAUT asks it.
• Primary-only marks. A question's marks go to its main topic; secondary topics get presence only. This matters for 2.7 Kernel PCA (asked inside 2025-26#8(c), shows 0 marks), 4.4 ML vs DL (the 2024-25#3 half goes to 4.3) and 1.2 distance measures (2024-25#6 goes to KNN).
• Marks are % of the printed 112, including all choices, not of the 70 a student answers.
• 2024-25 p2 (an unnumbered Naive Bayes item with no printed marks) counts as presence with 0 marks. 2023-24#7(c) was printed without its dataset; it still counts as a Naive Bayes numerical.
• Bayes topics count under Unit 1 (founder, CP1): 1.5 and 1.6 sit in Unit 1, and Unit 5's syllabus line points there.
• Rotation back-test: strong (88% on 16 calls), but the average includes the years being tested, and Unit 6's "↑" is only an absence rule on a never-asked unit.
• The 2025-26 recycling of 2022-23 is one observation. It may mean MAKAUT reuses papers from about three years back, which would point the next paper at 2023-24 questions, but four papers can't confirm that.
• No split unknown rows, no ODD questions, no MAKAUT question OUT. 9 GTU IoT-only rows are OUT-PROXY and never counted.

# Questions for CP3

1. S list size. 16 of 50 topics are S. Five of them (0.2, 1.9, 3.6, 4.3, 5.4) are S mostly through Group A one-liners, being descriptive in only 1 paper. Keep the standard rule (recommended: they're asked every year, and P1E sizes their notes by what's asked), or demote those five to A for this subject?
2. GTU favourites. SVM (1.11) is B and Bayesian learning/MAP (1.6) and DL architectures (4.5) are C on MAKAUT evidence, while GTU asks them constantly. Tiers stay MAKAUT-only (your CP0 rule). Should I add a "GTU favourite" flag (no tier change) so students don't under-prepare them?
3. Recycling pattern. Does it match your experience that MAKAUT reuses questions from about three papers back (2025-26 ← 2022-23)? If yes, I'll note in the attack plan that the 2023-24 questions deserve an extra look for the next exam.
4. Unit 6 has never been asked; the floor rule makes 6.1 Recent trends an A. Keep that (recommended, L15), or is Unit 6 effectively never set in ML? → Founder 2026-10-25: tier comes from MAKAUT papers, so 6.1 = C (☆, Not asked yet).

# CP3 decisions (founder, 2026-10-22: "CP3 approved", no edits)