DYSARTISANIDEAS, FORGED INTO FORM.Search the episode library ↗

THE INNOVATION FORGE / Workbench Workthroughs

Feature Engineering Act 1: Counts, flags and distance in Slate

Build inspectable Slate features with event counts, existence flags, click-rate bands and distance, while keeping dates, scope and missing data visible.

A brass instrument in progress with components and diagrams on a maker's bench.
Make the feature useful before making it automatic.

THE IDEA, IN A MINUTE

Give every feature a definition you can check.

Feature Engineering Act 1 turns a conference overview into a hands-on Configurable Joins walkthrough. The build moves from application scope and decision-date cutoffs to behavior counts, flags, click-rate bands and school or permanent-address distance.

01 / THE RECORDING

Watch the episode.
Or find your moment.

Start with the question behind the number

An event count looks simple until someone asks what it includes. Registrations or attendance? Interviews as well as tours? Everything on the person's record, or only activity connected with this application? Feature engineering starts with those questions. A useful export gives a behavior a definition that another person can understand and test.

Feature Engineering Act 1 works through that craft in Slate Configurable Joins. The query begins at the application level, returning a row for each application, then reaches into decisions, people, form responses, messages and addresses. Keep that starting point visible. A person can have several applications, and a person-level behavior can appear against more than one application unless the joins and filters narrow it appropriately.

In the recording · 07:51 ↗

Put a deliberate end on the observation window

The first substantial build compares an event's start date with a deposit date, falling back to a commitment deadline when no deposit date exists. Coalesce selects the first available value. The fallback year comes from the round name, illustrating how an export can adapt across cycles instead of carrying a forgotten hard-coded year.

That convenience depends on a reliable naming convention. Check which part of the round contains the year, whether that year represents the intended enrollment cycle, and whether the constructed date is valid. Also decide whether the cutoff includes the boundary day. The early event example establishes an upper bound; a complete post-admission window needs an appropriate lower bound too.

For historical analysis, use information available at the point the feature is meant to describe. Events after a deposit may reflect a different stage of the journey. Including them in a pre-deposit feature can make a retrospective result look more informative than it would have been at the time.

In the recording · 11:33 ↗

Keep counts, flags and points distinct

The event count adds registration status, template, folder and form-type filters. Attendance and a future registration answer different questions. Form responses hold the individual's status, while the form supplies context such as event date and template. If an interaction stands in for an event, its date should represent when the interaction happened, rather than when someone uploaded it.

A count preserves frequency. A flag records whether any qualifying activity exists. The walkthrough demonstrates both a count-based comparison and an existence export, then shows how either can contribute to a scoring rubric. A first-event bonus plus points per event can express diminishing returns without hiding the component counts.

Keep those components available for review. Illustrative points are not established predictive weights. When comparing two flag implementations, match their scope, attendance criteria and date windows first. The existence example switches to application-scoped responses, so its output should not automatically be treated as equivalent to the earlier person-scoped count.

In the recording · 17:00 ↗

Give a rate an honest denominator

The email example selects message method, type, status and delivery dates before dividing clicked messages by the chosen sent-message population. These choices define the rate. A denominator based on selected sent, delivered, opened and clicked statuses is a particular operational measure, so name it clearly and keep the numerator inside that same population and window.

Use a numeric type that preserves decimals. Guard a zero denominator with NULLIF so division produces a missing result rather than a divide-by-zero error. Zero clicks out of ten eligible messages is a valid zero rate; no eligible messages gives no rate to measure. NULLIF addresses the specified comparison, not every possible formula error.

The recording then uses ordered CASE conditions to create bands. Its 100-percent example is a prompt to inspect tracking quality, not proof that an individual is disengaged. Privacy features and automated scanning can distort email measurements. Review the raw counts, sample size and local evidence before deciding how any band should affect outreach.

In the recording · 35:42 ↗

Preserve the meaning of a distance

Distance begins with a ranked school, its matched organization and an address, then adds the person's ranked permanent address. The more elaborate example prefers a domestic school location, then a domestic permanent address. That ordering answers a specific question; it does not establish where someone currently lives or how easily they can travel.

Check the campus coordinates, address selection and output units before applying distance bands. The demonstration supplies a large static fallback when neither domestic condition qualifies. In an operational export, keep that fallback distinguishable from measured mileage and separate missing data from country information. Retain the selected source alongside the result so a surprising band can be investigated.

Finish with a small test set: a boundary value, an empty input and a record with competing sources. Keep the raw value, derived feature and scoring choice separately visible. That makes the work easier to revise when a new recruitment cycle, question or data issue arrives.

In the recording · 42:14 ↗

03 / A SMALL EXPERIMENT

Test a feature before giving it points

Use these three fictional cases to review a feature's scope, denominator and fallback. Write the expected result before checking the export.

One person, two applications

A fictional person has two applications and one attended interview linked only to the first application. A person-scoped count returns one on both rows. What should an application-specific interview flag return?

Compare your reasoning

It should qualify only the application with the linked interview, assuming that is the agreed definition. Match the join scope, attendance filter and observation window before comparing an existence flag with a count-based flag.

Revisit this moment · 26:24 ↗
No messages to measure

Fictional Record A has zero clicks from ten eligible messages. Record B has no eligible messages. Should both receive a zero click rate?

Compare your reasoning

A has a measured zero rate. B has an undefined rate because its denominator is zero. Guard division and preserve that absence through the banding step instead of silently equating it with no engagement.

Revisit this moment · 35:42 ↗
No usable address

A fictional record has an unmatched school organization and no usable permanent-address coordinates. A fallback returns 3,001. Should the record enter the same group as a measured distance of 3,001 miles?

Compare your reasoning

Keep the fallback status visible and handle unknown distance explicitly. A placeholder is not measured mileage. Review source selection, country conditions and coordinates separately before assigning a meaningful distance band.

Revisit this moment · 51:06 ↗
Build a test card to take with you

Notes stay in this tab. Reloading clears them.

04 / INSIDE THE STUDIO

The seams are part
of the story.

Behind the episode

While setting the email observation window, David catches a selection of the deposit decision where he wanted the admission decision. He corrects it and explicitly leaves the moment in the recording. It makes a practical review point visible: copying an export saves time, but its decision join and date comparison still need inspection.

Revisit the studio moment · 33:34 ↗

Art from the workshop

Workbench Workthroughs brings the build onto the bench, where each export and comparison can be inspected.

Open the artwork ↗

KEEP THE IDEA MOVING

Where would you take it next?

CONTINUE EXPLORING