← back to the showcase

Tree-model reports: the same measured outputs, less elapsed time

These result cards compare complete analysis calls on fixed models and inputs with their original libraries. Each card defines its measured workload and output contract. A factor is original elapsed wall time divided by TAKT elapsed wall time for that same unit; it is not a prediction for another model. These measurement cards do not announce a new accelerator package or a deployment offer.

Permutation importance and PDP/ICE

The five results below use three interleaved pairs on a shared Linux x86-64 host. The factor is the ratio of the original and TAKT medians. Both sides start with the fitted model and analysis data loaded; imports, loading and training are outside the windows. Preparation for each analysis call, its complete returned result and cleanup are included. Output checks are outside the timed windows.

Complete measured unit Original / TAKT elapsed seconds, medians Factor Fixed model and prediction threads
XGBoost permutation importance: 90 features × 5 repeats 44.187 / 6.549 6.75× 128 trees, depth 6; four threads
XGBoost PDP + ICE: a suite of 12 requests 20.612 / 4.833 4.26× The same 128-tree model; four threads
RandomForestRegressor permutation importance: 90 features × 5 repeats 568.666 / 175.484 3.24× 200 trees, depth at most 12; one thread
GradientBoostingRegressor permutation importance: 90 features × 5 repeats 264.996 / 26.563 9.98× 300 trees, depth at most 4; one thread
CatBoostRegressor permutation importance: 90 features × 5 repeats 76.729 / 15.111 5.08× 500 symmetric trees, depth 6, numeric features only; one thread

Measured inputs. All five use the same finite, dense, C-contiguous float32 matrix of 100,000 rows and 90 features, with float64 targets, from UCI YearPredictionMSD. The first 50,000 source rows train each frozen model; the following 100,000 form the analysis set. This is the stated experimental split, rather than the dataset's official held-out test split. The comparisons use XGBoost CPU 3.0.5, scikit-learn 1.7.2 and NumPy 2.2.6; the CatBoost card uses CatBoost 1.2.10. Each scikit-learn analysis has one worker. These are separate model comparisons, not a performance ranking of the libraries.

PI output contract. Each PI call uses R², five repeats, integer random seed 20261027 and the full analysis sample. Both sides use the same permutations. All 450 importances, 90 means and 90 standard deviations matched bit for bit. Separate complete checks compared 45.1 million predictions for each model: float32 for XGBoost and float64 for RF, GBR and CatBoost, with zero differences. The input matrix was restored; targets and the frozen model remained unchanged. The three float64 PI result arrays retain the original shapes, strides and layout flags. The measured models are scalar regressors with numeric inputs; other models, missing/sparse inputs, classifiers and other prediction modes are outside these cards.

PDP/ICE output contract. The 4.26× unit is the entire twelve-request suite, including its per-suite preparation. It uses scikit-learn method='brute', kind='both', 32 grid points and percentiles 0.05–0.95, for features 0, 8, 16, …, 88. All 38,400,000 float32 ICE values, 384 float32 averages and 384 float64 grid values matched bit for bit. Shapes, dtypes, strides and array-layout flags matched; the input matrix remained unchanged. It is not a factor for a single prediction or a SHAP report.

Prepared scenario books

These two cards compare complete responses to already prepared scenario requests using native XGBoost 3.0.5 CPU prediction as the original baseline, with one prediction thread. The factor is the sum of three original wall times divided by the sum of three TAKT wall times. The displayed seconds are those three-call totals, not per-call medians. Three interleaved pairs were recorded on a shared Linux x86-64 host.

Complete measured unit Original / TAKT elapsed seconds, three-call totals Factor Fixed input and model
Credit-data sensitivity report: 19 one-column scenarios 200.140 / 89.160 2.24× 1,345,310 rows × 15 numeric columns; 384 trees, depth 4
Motor-claim frequency what-if report: 17 BonusMalus values 76.780 / 39.979 1.92× 678,013 rows × 9 numeric columns; 320 trees, depth 4

Measured boundary. Imports, data/model loading, training, original encoding and construction of the requested scenario values are outside both windows. Each timed call includes runtime preparation, all scenarios over the complete book, the returned response and cleanup. Input data and models remain unchanged. These fixed numerical comparisons do not establish a general supported-model or API range.

Credit-data contract. The input comes from the public LendingClub dataset collection. The frozen model is a squared-error regression of a binary loan outcome, with scores clipped to [0, 1]. The 19 requests comprise 17 debt-to-income values, one verification-status flip and one debt-to-income replacement by its training median. The complete float32 response contains all 19 × 1,345,310 scores and 19 means and sums. All response bytes, shapes, dtypes, strides and layout flags matched. This card covers those finite requests; a separate NaN request was outside the accelerated comparison.

Motor-claim contract. The input is freMTPL2freq, OpenML 41214. The model is an exposure-weighted squared-error regression of ClaimNb/Exposure, with predicted frequency clipped below at zero. The complete float32 response contains all 17 × 678,013 scenario values and 17 means and sums; all bytes, shapes, dtypes, strides and layout flags matched. Encoding and scenario preparation are fixed before timing. These cards measure computation, not predictive accuracy or regulatory acceptance.

Existing LightGBM results

The existing LightGBM showcase has its own named workloads, thread settings, output checks and conditions. Its scope and results remain separate from the fixed-model cards above.

Measurement record fingerprints

These SHA-256 fingerprints identify the seven retained measurement reports. The records are kept for private review; this page supplies no code, models, datasets or accelerator packages.

Measurement record SHA-256
XGBoost PI 8747e34af843db1c823560fac811548547c05694df613a5e73ef17d82f6baa90
XGBoost PDP/ICE 0532a1cd094898c400b52a5041a1a1464120638c944a84e0815f4c0209b2bab2
Random forest PI f40edc03836c3a26e067699363deabe8667c8e5f330b4288a8ac8d4607a2d3b8
Gradient boosting PI 05f25e7b84eef16c6ac4ea64bb32b80a6587b33ca1f16c62748e1c6ba063dcc9
CatBoost PI a6bf9ddfc69478546944ddabd1339fb98a5cc9dcbeb01b39272c8afb4507dc26
Credit-data sensitivity df5b4a074bfbd43e82adeacc2feb313b0bbddaf3789b1c038d6e3926aff70d3d
Motor-claim what-if 9a34285d59c0603236baed29aeb20175c98ae2e438961536e1494f7132ada876

Software and data provenance

XGBoost and CatBoost use Apache-2.0; scikit-learn uses BSD-3-Clause. UCI lists YearPredictionMSD under CC BY 4.0, and OpenML's metadata lists freMTPL2freq as CC0. The LendingClub link identifies the source collection used for the measurement. This page contains result cards; it includes no source dataset records. Upstream software and data retain their respective licence conditions.

Measure the report you actually run

Describe the library and version, model and matrix sizes, report or scenario settings, required output, CPU allocation and run frequency. The useful comparison is the complete requested unit with its own declared output contract on the same system.

Describe a tree-analysis workload.

Project names identify the measured software; no endorsement is claimed. The factors describe the recorded builds and inputs, without a promised gain for another workload.

Source text: README.md

Telegram