1 MONTH | LIVE ONLINE | INTERMEDIATE
A disciplined applied ML workflow: problem framing, leakage-safe preprocessing, evaluation and compact delivery
Learn to frame prediction problems, build leakage-safe pipelines, compare supervised and unsupervised models, select metrics that reflect the real decision and deliver a reproducible prediction service — with Python.
Share your details and an academic counsellor will help with the curriculum, fees, batch schedule and eligibility.
By submitting, you agree to our privacy policy. We use your information to contact you about this course.
The Machine Learning Course focuses on applied classical ML for learners who already have Python and data-handling foundations. Its syllabus connects preparation, training and evaluation so you can explain why a result is trustworthy. Review the project scope and prerequisite check before choosing this focused pathway. Learners needing broader analytical foundations can compare Data Science with Machine Learning, while neural-network study belongs in the Deep Learning specialisation.
| You learn | You build | You leave with |
|---|---|---|
| Leakage-safe preprocessing | Reproducible pipeline | Prevent data leakage in a model pipeline |
| Supised and unsupervised modelling | Model card and segment playbook | Build classification and clustering models |
| Metric selection and evaluation | Evaluation report with baseline comparison | Choose and report appropriate metrics |
| Model interpretation | Feature attribution and error analysis | Explain model behavior to stakeholders |
| Compact deployment | Tested API and endpoint contract | Serve a model with consistent preprocessing |
A machine-learning model is not trustworthy because it achieved a high score. It is trustworthy when the problem is valid, the data is leakage-safe, the evaluation reflects reality, its errors are understood and its result can be reproduced.
Every section of this course is organised around producing that evidence. The loop below is the workflow you practise in every lab, assignment and project.
01
What decision are we trying to support?
Target definition, prediction unit, business objective, cost of errors, time horizon.
02
What information is actually available when prediction happens?
Prediction-time availability, future information, operational timing, leakage risk.
03
What does genuinely unseen data mean for this problem?
Random, group or temporal split; train, validation and test roles.
04
What simple approach must machine learning beat?
Majority class, mean prediction, a simple heuristic or a logistic-regression baseline.
05
Can preprocessing be reproduced safely?
Imputation, encoding, scaling and transformations fitted on training data only.
06
Which model families are worth comparing?
Do not start with complexity. Compare plausible candidate models.
07
Which errors matter?
Precision, recall, PR-AUC, ROC-AUC, MAE/RMSE and calibration — depending on the problem.
08
Where does the model fail?
Subgroup errors, residual patterns, outliers, data slices, temporal failure.
09
What can we responsibly say about model behavior?
Feature effects, interpretability, limitations, correlation versus causation.
10
Can another application use the exact workflow we evaluated?
Pipeline, artifact, prediction interface, FastAPI, validation, tests.
A score without this evidence is not enough to establish that a machine-learning model is reliable.
Goal: predict on June 1 which customers are likely to cancel during June. The feature set decides whether the resulting score means anything.
A model trained on the invalid fields may appear highly accurate — but it cannot reproduce that performance at prediction time, because on June 1 those fields do not exist yet.
Information created by or after the target event — the closure date exists only because the churn already happened.
A scaler or imputer fitted on the entire dataset before the train/test split lets validation data influence training. Always fit transformations on training data (or training folds) only.
The same customer, patient, device or account appears in both training and validation, so the model can memorise entity-specific patterns instead of learning generalisable ones.
Future information enters historical training or validation data — unrealistic for any problem where predictions are made forward in time.
A dataset has 10,000 customers: 500 churn, 9,500 stay. A model that predicts "everyone stays" scores 95% accuracy — and is almost useless for the business, because it never identifies a single customer who needs outreach.
| Metric | What it measures | When it matters |
|---|---|---|
| Accuracy | Share of predictions that are correct. | Misleading on imbalanced problems — the 95%-do-nothing model above. |
| Precision | Among predicted positives, how many were actually positive? | When false positives are expensive — e.g. wasted retention offers. |
| Recall | Among actual positives, how many did we identify? | When missing a positive is expensive — e.g. churners lost without contact. |
| F1 | Harmonic mean of precision and recall. | When both error types matter and you need one comparable number. |
| PR-AUC | Area under the precision-recall curve. | Ranking quality in strongly imbalanced positive-event problems. |
| ROC-AUC | Probability the model ranks a random positive above a random negative. | Useful, but should not automatically be the only reported metric. |
| Calibration | Do predicted probabilities behave like probabilities? | Among customers assigned ~0.70 churn probability, roughly 70% should churn. |
| MAE | Average absolute numeric error. | Regression: every error unit counts equally. |
| RMSE | Root mean squared error. | Regression: penalises larger errors more strongly. |
Choose the metric by asking what type of mistake matters to the real decision — not by choosing whichever metric produces the highest-looking number.
Evidence from a real decision context, not a leaderboard, decides which model ships:
| Evidence | Logistic Regression | Gradient Boosting |
|---|---|---|
| PR-AUC | 0.54 | 0.57 |
| Latency | Very low | Higher |
| Explainability | High | Medium |
| Maintenance | Simple | More complex |
| Improvement | Baseline | +0.03 |
Sometimes yes. Sometimes no. The decision depends on business value, latency, interpretability, maintenance, data size and the cost of failure. This judgment — not algorithm memorisation — is what the course teaches.
You should already be comfortable writing basic Python, working with functions and manipulating tabular data with pandas. Advanced software engineering is not required before joining.
Expected abilities
You do not need advanced mathematics before joining, but you should understand basic algebra, descriptive statistics, averages, variation and basic probability.
Required foundation
Algebra · Mean and median · Variance and standard deviation · Percentages · Basic probability
Useful during the course
Vectors · Matrices · Distributions · Conditional probability · Gradient intuition
Not required before joining
Proof-heavy calculus · Advanced linear algebra · Graduate-level statistics
Can you tick most of these?
Mostly yes
This focused Machine Learning pathway may be appropriate for you.
Several no
Build your foundations first — build your Python, SQL and statistics foundations.
Required knowledge
Basic Python, pandas, descriptive statistics and comfort with algebra are required.
Optional prework
A scored readiness check is available to confirm your starting level before enrolling.
A serious machine learning project should answer: what problem is being solved? What baseline exists? How is the split designed? What leakage risks exist? Which metric matters? Where does the model fail? How stable are the results? What limitations remain? What artifact proves the work?
| Artifact | Why it matters |
|---|---|
| Problem statement | Shows the actual decision |
| Baseline | Proves ML adds value |
| Split rationale | Shows validation is credible |
| Reproducible pipeline | Prevents inconsistent preprocessing |
| Evaluation report | Shows actual model evidence |
| Error analysis | Shows failure awareness |
| Model card | Records intended use and limitations |
| Decision note | Connects model to action |
| API | Demonstrates delivery beyond notebook |
If you need Python, SQL, statistics and EDA from the beginning, start with the broader foundation program instead.
Data Science Course →Compact model delivery is included. Production lifecycle automation and operations belong in the specialist program.
MLOps Course →LLM and RAG applications belong in the generative AI specialization.
Generative AI Course →Neural-network depth belongs in the dedicated deep learning pathway.
Deep Learning specialization →Not always. If you already understand basic Python, pandas, descriptive statistics and tabular data, you can enter a focused machine learning course directly. If those foundations are missing, learn them first through a broader data science pathway.
Machine learning focuses on building and validating models. MLOps becomes important when the challenge shifts to repeatable deployment, model versions, production monitoring, retraining and lifecycle automation.
Classical machine learning focuses largely on predictive tasks such as classification, regression, clustering and forecasting. Generative AI uses large foundation models to generate, retrieve or transform content through applications such as LLM systems and RAG.
| Aspect | Machine Learning | Generative AI |
|---|---|---|
| Model type | Predictive models | Generative / foundation models |
| Typical applications | Classification | LLM applications |
| Numeric outcomes | Regression | RAG |
| Grouping | Clustering | Embeddings / retrieval |
| Working style | Feature / model workflow | Prompt / context / retrieval workflow |
| Evaluation | Classical evaluation | Generative / LLM evaluation |
The curriculum is organised into 4 modules. Each module includes concepts, a guided lab, an independent assignment, and exit evidence.
Attend scheduled live online classes with an instructor.
Work through guided labs with support during and between sessions.
Complete assignments and projects that produce reviewable artifacts.
Receive feedback on your work and revise before presenting.
Speak with an academic counsellor for the current fee, GST status and available payment options before you enrol.
Ask admissions which payment options are available and whether eligibility conditions or additional charges apply.
Last curriculum review: 2026-09-09
Machine Learning
CurrentData Science with Machine Learning
Full Stack Data Science
Ask an academic counsellor about the curriculum, prerequisites and next available cohort.