Learn Feature Engineering with Phoebe

Two tracks into the inputs that decide your model

A column is what the database gives you. A feature is what the model can learn from - and almost all of the winnable performance in applied machine learning lives in the gap between them. A leader track teaches why feature work beats model swaps, how a suspiciously good score is a red flag rather than a win, and what feature governance actually costs. A practitioner track rebuilds Lumen's checkout data column by column: scaling, encoding, datetime decomposition, interactions, missing-data signal, selection, and an honest bridge to embeddings. A live workbench trains a real model in your browser - watch held-out AUC climb 0.500 to 0.813 with the algorithm never changing, then press the leakage lever and watch a training score of 0.967 collapse to 0.789 the moment the column that knew the future goes missing.

2tracks
16sessions
813held-out AUC x1000
45min per session

The leader track 🤝 · for data leads, PMs & DS managers · 6 x 45 min · no code

Where the ROI actually is: why feature work outperforms model shopping, the leakage disaster explained without maths, data quality as a budget line, feature governance and stores, when to stop hand-crafting - and the review questions that make a DS team's feature work legible.

💡 Leader 1 · easy

Why features decide ROI before models do

The unglamorous truth: practitioners spend far more time on data than models, and that is where the return is. Where to put the effort.

▶ Leaders start here 🕳️ Leader 2 · easy

The leakage disaster, in plain English

The one failure that announces itself with a better score - and sklearn's own experiment where leakage produced 0.76 accuracy on pure noise.

▶ The trap demo 🧹 Leader 3 · medium

Data quality is a feature budget

Missingness, cardinality, drift-prone columns, bad labels - what to fund upstream versus patch downstream, and why a few bad rows spoil a large dataset.

The spend decision 🏛️ Leader 4 · medium

Feature governance + feature stores

Training-serving skew, Google's Rule 29, point-in-time correctness, reuse across teams - and an honest answer on when a feature store pays for itself.

Build-vs-buy framing 🧬 Leader 5 · deciding

Hand-crafted vs learned features

The embeddings question for decision-makers: when to stop engineering columns and let the model represent - with the cost and interpretability tradeoffs stated plainly.

The stop-engineering line 📋 Leader 6 · deciding

Questions to ask your DS team

The feature review checklist, what a feature spec must contain, and the four questions that surface leakage before a launch rather than after.

The review ritual

The practitioner track 🧪 · for DS, analysts & ML engineers · 10 x 45 min · Python + the live workbench

One running case, rebuilt column by column: Lumen Skincare's checkout sessions. The held-out ladder climbs 0.500 → 0.696 → 0.785 → 0.806 → 0.812 → 0.813 with the algorithm untouched, and b5 exists to make sure none of it was a lie. Ends with a written feature spec you can hand to a colleague.

🔤 Practitioner 1 · easy

What a feature actually is

Column versus feature, the model's-eye view of your table, the five feature groups - and your first climb up the live ladder.

▶ Practitioners start here 📐 Practitioner 2 · easy

Numeric transforms

The scaler family and when each one is right, skew and log, clipping, binning strategies, and splines versus polynomials for curvature.

Shape the numbers 🏷️ Practitioner 3 · medium

Categorical encoding

One-hot and its parameters, rare-category grouping, the high-cardinality problem - and target encoding done safely with internal cross-fitting.

▶ The quiet leak 🕰️ Practitioner 4 · medium

Datetime + cyclical features

Decomposition, cyclical encoding for hours and months, tenure and recency - everything a timestamp was hiding in plain sight.

+0.021 on the ladder 🕳️ Practitioner 5 · the heart

Leakage: the session

THE core session: the four families, sklearn's three rules, the Pipeline as structural defence - and two live demos where good work and leaked work look identical until they do not.

▶ 0.967 → 0.789 ✖️ Practitioner 6 · hands-on

Interactions + aggregations

Ratios, feature crosses and their dimensionality bill, group-level aggregates and the entity-history pattern that most teams never build.

Relationships, not columns 🕵️ Practitioner 7 · hands-on

Missing data as signal

Imputation strategies and their assumptions, missingness indicators, NaN as its own category - and when absence predicts better than presence.

The empty cell talks 🎚️ Practitioner 8 · hands-on

Selection + importance

Variance, univariate, RFE and model-based selection, mutual information and its caveats - plus why impurity importance lies and permutation importance does not.

Retiring session_hour 🧬 Practitioner 9 · hardest

Embeddings as features

The bridge: a free-text column becomes a vector becomes a tabular feature. Dimensionality and cost tradeoffs, and when learned beats hand-crafted.

The honest bridge 🏆 Practitioner 10 · hardest

Capstone: the feature spec

Assemble the full pipeline on Lumen and write the spec that makes it survivable: definition, source, write-time, leakage check, owner, refresh.

Ship the spec
easy - everyone medium hands-on / deciding hardest

Choose your path 🗺️

Two doors into one conviction: the model is rarely the bottleneck. Leaders learn where to spend and what to ask; practitioners learn to build the columns and prove they are real.

🤝 Leader (fund it, review it) a1→ a2→ a3→ a4→ a5→ a6
🧪 Practitioner (build the columns) b1→ b2→ b3→ b4→ b5→ b6→ b7→ b8→ b9→ b10
🕳️ Leakage fast lane b1→ b5→ b3→ b10
🚀 Full journey (fund AND build) a1-a6→ b1-b10

The knowledge map 🧠

Both tracks at a glance - hover a session to spotlight its concepts, click any node to jump in.