Module 1What "unusual" means
4 lessonsWhat "unusual" means. Meet the payment that did not fit, name the four kinds of anomaly, see why labels are rare, and load the Larkfield data and the comparison table that every module fills.
Find the few events that matter among millions that do not. Compare rules, robust statistics, control charts, isolation, time-series and deep methods on the same data, choose by the cost of a miss, and run the detector as a daily job that people trust.
Read every lesson as a guest, or sign in to enroll and track your progress.
Enroll to save the lessons that you complete. It is free.
You are enrolled. From now on, the lessons that you complete are saved.
This course teaches how to find the unusual: framing an anomaly problem, robust baselines and rules, control charts and drift tests, distance, density and isolation methods, time-series detectors and change points, autoencoders and embeddings, evaluation with few labels under an alert budget, learning from reviewers, and a detection job in production. It does not teach supervised models in general, forecasting in depth, or how to operate a service.
Take another course if you want to:
By the end of this course, you will be able to:
Larkfield, an online shop for home and garden products, wants to find unusual events before they cost money. A customer who buys a small basket every week suddenly makes eight payments in a foreign city in ten minutes. One support agent’s refund rate rises a little and stays high. The ticket API slows down a little more every day. A data feed sends zeros for a whole day.
In this course, you follow Priya, a data scientist, as she builds and compares detectors for these problems. Sofia’s team reviews the payment alerts and can check 50 a day. Grace reads the control charts for refunds. Amira owns the ticket API, and Mei owns the data feeds.
Every method runs on the same Larkfield data and goes into one comparison table: alerts per day, the planted anomalies found within the daily budget, the cost per week and the time to explain one alert. You see a three-line rule beat a neural network on one problem and lose on another, and you learn why.
All data is synthetic and has no real people. Every practice anomaly is one we planted on purpose, and the course says so each time. Everything runs in notebooks on a laptop CPU or in free Colab, with no paid service.
You completed Machine Learning, From Problem to Reliable Model, or you have the same skills (features, splits, scikit-learn pipelines, precision and recall, thresholds). You can read and change Python and pandas code in a notebook. The mathematics, forecasting and MLOps courses help but are not required; lessons that use their ideas explain them briefly and link to them. You need a free Google or Kaggle account for the notebooks, or Python 3.12 on your computer. You need no paid service.
Courses to take first: Machine Learning: From Problem to Reliable Model.
Data scientists, ML engineers, analysts and backend or operations engineers who must find rare, unusual events (fraud, faults, bad data, sudden changes) and turn them into alerts that people act on.
Experienced AI practitioners designed this course. AI helped us draft the content; people chose the teaching approach and the examples, and reviewed every lesson. The outputs that you see are the actual results of running the code. How we make our courses
What "unusual" means. Meet the payment that did not fit, name the four kinds of anomaly, see why labels are rare, and load the Larkfield data and the comparison table that every module fills.
Simple baselines that are hard to beat. Score values with z-scores, see how one extreme value fools the mean, use the median, MAD and IQR per customer, and write the rules that a fraud team already uses.
Statistical process control and distribution drift. Watch the refund rate with I-MR, p-, CUSUM and EWMA charts, measure the price of each run rule, test distributions with KS, PSI and Wasserstein, and use Benford's law and other business checks.
Distance, density and isolation. Score payments with kNN distance, LOF, Isolation Forest and a One-Class SVM, use PyOD to run many detectors with one interface, and see that better features beat a better algorithm.
Anomalies in time. Build seasonal baselines, detect on STL residuals, forecast and compare with a prediction interval, find change points, and group alerts from many metrics so that one incident gives one alert.
Deep and embedding methods. Train a tiny autoencoder and compare it honestly, find new log messages with templates and embeddings, use an LLM only to summarise an alert, and know when not to use machine learning.
Evaluate with few labels. See why accuracy lies, choose the threshold from a daily alert budget, compute the cost per week, evaluate without a full answer key, and check the alerts by slice.
Learn from the reviewers. Re-rank alerts with the reviewers' labels, avoid the selection-bias trap with a random sample, and explain each alert so that a reviewer can decide in one minute.
Run a detector in production. Build a scheduled detection job that follows the budget and never sends an alert twice, handle a product launch that moves normal, and keep alerts that people still read.
Four end-to-end case studies on real open data. Card fraud with a review budget, faults in machine or server metrics, unusual readings in public open data, and control charts on a real production line. Complete at least one for the certificate; each one is a portfolio piece.
Complete at least one case study for the certificate. The others are optional. What each one covers
Use every module in one practice project. You collect evidence, and check it against the course outcomes.
Realistic projects from start to finish, on real data, with every output explained. Each one makes a portfolio piece. Complete at least one for the certificate; the others are optional.
Card fraud with a review budget2 hr 30 min
Anonymised card transactions: compare a rule, Isolation Forest, an autoencoder and a supervised model on the few labels, and choose by precision at a daily review budget.
Faults in machine or server metrics2 hr
Labelled time series of real systems: compare seasonal baselines, STL residuals, change points and a multivariate detector, scored by time to detection and alerts per day.
Unusual readings in public open data2 hr 30 min
A public open-data series with no labels: plant anomalies, separate real events from data faults, and write the alert policy.
Control charts on a real production line2 hr
Sensor readings and pass/fail results from a manufacturing process: control charts and drift checks on the key signals, compared with Isolation Forest.
Loading the reviews…
Took this course? Sign in to review it. Learners can write a review after they complete 25% of the lessons.