Machine learning on GATE DA is not a Kaggle competition. It is a computer-based exam where you are expected to reason about model assumptions, compute decision boundaries by hand, trace k-means iterations on five data points, and decompose error into bias² + variance + noise. The syllabus is bounded — every model you need is named — and the biggest mistake aspirants make is over-studying deep learning and ensemble methods that are not on the list.
Below: the official syllabus, a model-comparison decision matrix (print it), the recommended book (Hands-On ML by Géron, supplemented with theory), and a study plan that starts with the prerequisite maths you need before touching any ML topic.
Verify: Confirm the syllabus on the official GATE 2027 syllabus page (IIT Madras). The ML section is unchanged from GATE DA 2026.
Official Syllabus
Supervised Learning
Begin with regression and classification problems, then study the named models and evaluation methods:
- Regression — simple linear, multiple linear, ridge
- Logistic regression
- K-nearest neighbour, naive Bayes classifier
- Linear discriminant analysis (LDA)
- Support vector machine (SVM)
- Decision trees
- Bias-variance trade-off
- Cross-validation — leave-one-out (LOO), k-folds
- Multi-layer perceptron (neural networks), feed-forward
Unsupervised Learning
- Clustering — k-means, k-medoid, hierarchical (top-down, bottom-up); single and multiple linkage
- Dimensionality reduction — PCA
Every model you need is on this list. Random forests, gradient boosting, transformers, reinforcement learning, and generative models are not. The explicit mention of bias-variance trade-off and cross-validation means these are direct question targets, not background reading.
The Model Decision Matrix
Use this to practise model selection and reasoning about assumptions. The examples unpack models named in the official syllabus; supporting derivations are not separate syllabus headings.
Official weightage: The official pattern assigns 15 marks to General Aptitude and 85 to DA technical subjects combined, with no fixed allocation for ML or neural networks. Prerequisites, priorities and study times here are pedagogical guidance, not mark forecasts.
| Model | When to use | Practice focus |
|---|---|---|
| Linear regression | Continuous target, linear relationship | Closed-form solution, residuals, OLS assumptions |
| Ridge regression | Linear regression + L2 regularisation | Effect of λ on coefficients, bias-variance |
| Logistic regression | Binary classification, linear boundary | Sigmoid, log-loss, boundary computation |
| KNN | Non-parametric classification/regression | Effect of k, distance metric, dimensionality curse |
| Naive Bayes | Categorical features, fast probabilistic | Bayes + conditional independence; compute posterior |
| LDA | Classification under Gaussian-class assumptions; also dimensionality reduction | Within/between-class scatter; LDA vs PCA |
| SVM | Maximum-margin classifier; kernel for non-linear | Identify support vectors, effect of C |
| Decision tree | Interpretable classification/regression | Splits via Gini/entropy; pruning |
| K-means | Unsupervised clustering, k known | Trace iterations; initialisation sensitivity |
| K-medoid | Clustering using representative observed data points | Compare medoid and centroid; compute dissimilarity costs |
| Hierarchical clustering | Top-down (divisive) or bottom-up (agglomerative) | Single and multiple linkage, including complete/average linkage examples |
| PCA | Unsupervised dimensionality reduction (variance-preserving) | Principal components via covariance; SVD link |
| Neural network (MLP) | Non-linear function approximation | Forward/backprop trace; activation functions |
Book and Resources
| Resource | Role | Use for | Skip |
|---|---|---|---|
| Aurélien Géron — Hands-On Machine Learning | Primary book | Practice, intuition, code, model walk-throughs | Random forests, boosting, CNNs/RNNs, deployment, distributed training, RL |
| Goodfellow et al. — Deep Learning | Supplementary lookup | Perceptrons, backprop, activation functions | CNNs/RNNs, generative models and other material beyond MLP/feed-forward foundations |
| The ML Hub mentor notes | Theory layer | SVM derivations, LDA vs PCA, bias-variance algebra | — |
| Official GATE DA PYQs (2024 onwards) | Practice | Model-identification, conceptual MSQs, trace questions | — |
Full book list: GATE DA books and resources guide.
Topic Walkthrough
Regression — Linear, Multiple, Ridge
OLS estimator (closed-form), residuals, R², assumptions. Ridge adds L2 penalty; understand how λ shrinks coefficients and trades bias for variance. Be able to compute the closed-form ridge solution.
Logistic Regression
Sigmoid activation, binary cross-entropy loss, linear decision boundary. Maximum likelihood interpretation. The decision boundary shape is the same hyperplane as a linear SVM — different optimisation objective, different boundary position.
KNN and Naive Bayes
KNN is non-parametric; effect of k and distance metric matter; curse of dimensionality is a common MSQ target. Naive Bayes assumes conditional independence given the class; this is applied Bayes' theorem.
LDA — and Why PCA Is Not the Same Thing
LDA is listed under supervised learning. Study classification with Gaussian class-conditional distributions and a shared covariance matrix, which produces linear decision boundaries. Its related discriminant-projection view maximises between-class relative to within-class scatter. PCA instead maximises retained variance without class labels. Comparing these objectives helps distinguish supervised classification from unsupervised dimensionality reduction.
SVM
Maximum-margin classifier. Hard-margin vs soft-margin (C controls the trade-off). In the hard-margin setting, support vectors lie on the margin; with a soft margin they can also lie inside it or be misclassified. They are the training points with non-zero dual weights that determine the boundary. Kernel trick (polynomial, RBF) for non-linear boundaries — conceptual exposure, not derivation depth.
Decision Trees
Splits via information gain (entropy reduction) or Gini impurity. Pruning controls overfitting. Be able to compute the next split given a small dataset.
Bias-Variance Trade-off
Error = bias² + variance + irreducible. Underfitting = high bias; overfitting = high variance. Regularisation (ridge), early stopping, and reduced complexity trade variance for bias. PYQs ask you to identify which is responsible for a train/test gap.
Cross-Validation
LOO: uses n − 1 for training, 1 for validation, repeated n times — high variance, expensive. K-fold: splits into k partitions — lower variance. Stratified k-fold preserves class proportions.
Clustering
K-means: minimise within-cluster sum of squares; iterate assignment/centroid update; sensitive to initialisation and k. K-medoid: choose a representative observed point for each cluster and minimise total dissimilarity to the medoids; compare its updates and sensitivity to outliers with k-means. Hierarchical clustering includes both agglomerative (bottom-up) merging and divisive (top-down) splitting. Cover the official single and multiple linkage wording using single, complete and average linkage examples, and practise reading dendrograms.
PCA
Eigen-decomposition of the covariance matrix; principal components = eigenvectors of largest eigenvalues. PCA = SVD applied to centred data — this is where linear algebra meets ML.
Neural Networks
The official topics are multi-layer perceptrons and feed-forward neural networks. Perceptrons, activation functions (sigmoid, tanh, ReLU), forward passes and chain-rule/backpropagation calculations are useful ways to understand those models. Trace a small network and its loss derivatives by hand. Gradient behaviour can provide context, but no neural-network mark allocation is guaranteed.
What to Skip
- Random forests, gradient boosting, XGBoost — supplementary ensemble methods, not part of the mandatory model list
- CNNs and RNNs — not the listed MLP/feed-forward network topics
- Reinforcement learning — Q-learning, policy gradients
- Transformers and attention mechanisms
- Generative models — GANs, VAEs, diffusion
- Production ML / MLOps
- Specific DL frameworks beyond conceptual MLP
- Statistical learning theory proofs (VC dimension, PAC bounds) beyond conceptual awareness
Keep supplementary architectures separate from the required models. The reason is syllabus scope, not an assumed distribution of marks. Loss differentiation, likelihood interpretations and other foundational methods remain useful when tied to a listed model.
Practice Patterns for Official Topics
Use these exercise categories alongside official PYQs; they do not represent a verified frequency or weightage analysis:
- Identify the right model for a described problem
- Compute the next iteration of k-means or a decision-tree split
- Bias-variance attribution given train/test errors
- PCA computation on a small covariance matrix
- Bayes-style probabilistic classification
- SVM support-vector identification
Solve every ML question from official GATE DA 2024 and 2025 papers. The decision matrix above tells you what to look for; PYQs train the pattern recognition.
Free benchmark
The ML chapter in our free GATE DA demo course includes mentor-led lectures on linear regression, SVM, PCA and neural networks plus a topic-test.
Study Plan
Prerequisites (do first)
Probability and statistics (naive Bayes, bias-variance, evaluation) and linear algebra (PCA, ridge, SVM, neural networks). Do both before ML.
Weeks 1–4: Supervised Learning
- Linear, multiple, ridge regression — Géron + theory notes.
- Logistic regression — sigmoid, log-loss, boundary.
- KNN, naive Bayes.
- LDA — and the PCA comparison.
- SVM — hard/soft margin, support vectors, kernel intuition.
- Decision trees — Gini, entropy, information gain.
- Bias-variance + cross-validation.
Weeks 5–6: Unsupervised + Neural Networks
- K-means and k-medoid; hierarchical clustering, top-down and bottom-up, with single and multiple linkage — trace on small datasets.
- PCA — covariance eigen-decomposition, SVD connection.
- Neural networks — MLP, activations, forward/backprop trace.
Weeks 7–8: PYQs and Revision
- Every ML PYQ from GATE DA 2024 and 2025.
- 2–3 topic-wise tests from The ML Hub GATE DA test series.
- Finalise the decision matrix and PCA-vs-LDA reference.
Mistakes That Cost Marks
- Expanding the neural-network scope. Cover the listed MLP/feed-forward models. CNNs, RNNs, RL, generative models and transformers are supplementary, not mandatory syllabus topics.
- Confusing PCA and LDA. PCA = unsupervised, maximises variance. LDA = supervised, maximises class separation.
- Treating Hands-On ML as complete. Géron is practice-leaning; supplement theory for SVM, bias-variance algebra, LDA vs PCA.
- Skipping bias-variance. Explicitly in the syllabus, common PYQ target.
- Forgetting PCA = SVD on centred data. Reinforces the LA connection.
- Studying ensembles "just in case". Not in the syllabus. Allocate that time to in-syllabus models.
ML in The ML Hub's Course
The ML block in The ML Hub's GATE DA course layers mentor-led theory on top of Géron's practical material. Every in-syllabus model is covered, plus the bias-variance decomposition and cross-validation methods the syllabus names explicitly. Topic-wise tests include dedicated ML packs aligned to PYQ patterns. See ranker journeys for how AIR 9 and AIR 6 candidates used this material.
The subject that matters beyond GATE
ML is the subject most directly tied to the careers GATE DA leads into — and the one where structured prep beats scattered tutorials most clearly.
- Mentor-led lectures on every in-syllabus model, with the theory Géron underplays
- Model decision matrix and PCA-vs-LDA disambiguation from GATE DA rankers
- Topic-wise tests on regression, classification, clustering, PCA in the test series
FAQs
Which book for GATE DA machine learning?
Aurélien Géron's Hands-On Machine Learning. It covers nearly every algorithm in the syllabus. Pair with mentor notes for theory (SVM, bias-variance, LDA vs PCA).
Is deep learning in GATE DA?
The syllabus explicitly lists multi-layer perceptrons and feed-forward neural networks. Activation functions and backpropagation are supporting concepts for those models, not separately named syllabus entries. CNNs, RNNs, generative models, transformers and RL are not listed, and no marks allocation for neural networks is published.
Is SVM in the GATE DA syllabus?
Yes. Cover the maximum-margin formulation, hard/soft margin (C), support vectors, and kernel trick at conceptual level.
What is bias-variance trade-off in GATE DA?
Explicitly in the syllabus. Error = bias² + variance + irreducible. Underfitting = high bias; overfitting = high variance.
Related Guides
ML's natural neighbours: Artificial Intelligence (overlaps on probabilistic reasoning), Probability & Statistics and Linear Algebra (both prerequisites). Full subject map: GATE DA books and resources · GATE DA syllabus 2027 guide. For a structured month-by-month plan, see How to Prepare for GATE DA in 8 Months.