Learn Unsupervised with Phoebe

Finding structure with no answer key

Mango Lane has 12,000 customers and no labels. Ten builder sessions turn that pile into named, fundable segments - real k-means, DBSCAN, PCA, and basket mining running live in your browser - while teaching the skepticism that keeps you from naming noise. The unlabeled half of the machine learning zoo, with the traps left armed so you can spring them safely.

10sessions
4real algorithms in the lab
300customers to segment
45min per session
🥭 Session 1 · easiest

The pile with no labels

What changes when the answer column disappears, the RFM portrait of every customer, and the framing that anchors the whole course: clusters are imposed, then validated - never simply discovered.

▶ Start hereFoundations ⚖️ Session 2 · easiest

What "similar" means

Distance is the whole game - and the billing export's cents column has been playing it against you. The scaling lever: 70% to 100% truth match in one preprocessing line, live in the lab.

The lab opens 🎯 Session 3 · building

k-means from the inside

Step Lloyd's loop by hand: assign, update, repeat until still. Inertia, k-means++ starts, the unlucky initialization that locks into a worse answer - and the n_init default old tutorials still get wrong.

The workhorse 🎛 Session 4 · building

Picking k: the question with no answer key

Silhouette read fluently, the elbow method named and demoted, the actionability test - and the "More segments!" trap button that turns a 0.68 into a 0.34 while looking more sophisticated.

✨ The k-trapInteractive 🌙 Session 5 · building

Beyond k-means

Two interleaved moons k-means cannot see (83%) and DBSCAN traces perfectly (100%), dendrograms, GMM's soft maybe-both answers - and how to choose between them like the sklearn comparison table does.

The rivals 🗜 Session 6 · building

Compressing the customer

Real PCA: three features become two axes carrying 86% of the variance. Plus the honesty split the field keeps fumbling - PCA compresses, t-SNE and UMAP visualize, and embedding distances lie.

Dimensionality 🏷 Session 7 · building

Naming and shipping segments

From cluster ids to personas with owners: profiling tables, stability checks, the segment sheet artifact - and why naming before validating is the most enjoyable mistake in data science.

The deliverable 🧺 Session 8 · building

What sells together

Association rules mined live from 500 Mango Lane baskets: support, confidence, and lift - why lift is the honest column, and the sneakers-to-socks 3.0x rule your cross-sell email writes itself from.

Basket mining 🎭 Session 9 · deepening

When clustering lies

Run 5-means on pure noise and watch it answer confidently - then learn the tells: the silhouette gap, sklearn's own assumptions demo, drift, and the professional judgment of when NOT to cluster.

🎭 Noise labThe skeptic session 🏁 Session 10 · capstone

Capstone: the segmentation project

The whole pipeline in one sitting: scale, validate k, name, action, monitor - the four-press ladder from 70% wallet bands to five named segments with campaigns, plus your hand-sort from session 1, graded.

Ship it
easiest - foundations building - the craft deepening - the skeptic capstone - ship it
🥭 The running case: Mango Lane's 12,000 customers. One brief from marketing - "give us segments we can campaign to" - carries all ten sessions. The lab's 300-customer sample has five planted archetypes; you recover them badly (raw cents), then perfectly (scaled), validate that five is real, name them, mine their baskets, and ship the segment sheet. The same twelve customers you hand-sort in session 1 come back graded in the capstone.
The honesty rail. The lab data is simulated - that is the only reason a "truth match" score can exist, and every page says so. Everything else is real: real k-means++ (best of five restarts), real DBSCAN, real spherical-GMM EM, real Jacobi PCA, real silhouette, real rule mining - all computed in your browser on every click. And where a topic has its own course on this shelf (anomaly detection, recommenders, feature engineering), we point instead of re-teaching.

Choose your focus 🧭

Everyone starts at Session 1 for the mindset and the case. After that, follow what you need - or walk all ten in order.

🎯 I need to segment customers now 2→ 3→ 4→ 7
🕵️ Make me the skeptic in the room 1→ 4→ 9
🛤 The full journey 1→ 2→ 3→ 4→ 5→ 6→ 7→ 8→ 9→ 10

The knowledge map 🧠

The whole course at a glance - hover a session to spotlight its concepts, click any node to jump in.