Interactive experiments

A closer look at the work.

Explore web experimentation, model selection, and terrain. The A/B example uses illustrative numbers; the thesis figures use project data.

JavaScript is needed for the interactive controls. Static figures and explanations are available in the research statement and project notes.

From a visit to a conversion.

Follow visitors through two checkout designs. Watch completed orders collect, and see the results take shape.

This illustrative experiment sends 100 visitors to each checkout. A produces 52 completed orders and B produces 68. Enable JavaScript to watch visitors move through the pages and build the 3D result columns.

About this illustration

Visitors split evenly between an original checkout and a variation with express payment. Each dot represents four visitors taking the same illustrated path. Dots that complete an order collect in a column; hollow dots fade out when visitors leave without ordering.

The fixed sequence ends with 100 visitors per version: 52 orders for A and 68 for B. The two columns use the same 0–100 scale. At the end, those counts also give the completion rates: 52% and 68%, a difference of 16 percentage points.

These are invented outcomes to illustrate the flow of an experiment. They are not client results or live visitor data, and do not establish that express payment caused an improvement. A measured test would need a sampling plan and an uncertainty analysis before drawing conclusions.

The shape of a selection rule

A bounded surface over accuracy and error. Drag the figure, or focus it and use the arrow keys, to rotate.

TOS(z_acc, z_err) = ½ ( tanh z_acc − tanh z_err )

drag or use arrow keys

TOS is bounded between −1 and 1 and is antisymmetric about the diagonal. The seven markers locate the thesis models at their standardized accuracy and error. The shape encodes a trade-off; whether that trade-off fits the application remains a separate question.

What should a missed landslide cost?

Change the cost ratio and watch the seven candidates trade places.

one missed landslide is worth 1.0 false alarms

Confusion counts are reconstructed from cross-validated accuracy and recall, with 196 positives and 196 negatives. Gradient boosting remains the winner on this data; the other rankings and the winning margin change. View the source data.

The pool is part of the metric

Add the baseline, or remove a candidate. Each score is recalculated from the selected pool.

Because TOS standardizes across candidates, an alternative that never wins can still change the ordering of other models. A relative score needs to be reported with its candidate pool. Read about the limitation.

A walk through Rangamati Hill Tracts(a district within Chattogram division, Bangladesh)

Explore the thesis susceptibility grid as a terrain model.

Walk the susceptibility map

Explore 5,067 cells using SRTM elevation and the model’s susceptibility scores. In this game, scores of 0.60 to 0.80 trigger shaking and scores of 0.80 or higher trigger collapse. These are illustrative game rules.

WASD or the arrow keys · Esc to pauseTouch the terrain and slide to walk. Hold to keep moving; lift to stop.

The 5,067 cells use SRTM elevations and the model’s susceptibility scores, with vertical scale exaggerated 7×. The collapse behavior is an illustration of score bands, not a physical landslide simulation. Scores rank sites relative to the training data; they are not calibrated probabilities that a hillside will fail.