STRUCTURED · SELF-PACED · EVIDENCE-BASED
Build Production-Ready BDTs for Your Analysis
Six modules. From first principles to a validated, publication-ready classifier. The structured path that scattered GitHub notebooks can't give you.
Your supervisor assigned a BDT-based analysis. You've spent weeks across 15 GitHub repos, three TMVA tutorials, and a 200-page ROOT manual. Your ROC curve still looks wrong. The review committee meets in four months.
You don't need another lecture recording. You need a structured method — one that takes you from 'what is a decision tree' to 'here is my validated, production-ready classifier' in a clear sequence.
What People Are Saying
"SAMPLE — I went from zero BDT experience to passing my analysis review in 8 weeks. The validation module alone saved me a month of debugging."
— Sample testimonial — PhD Student, LHCb
"SAMPLE — Finally, a resource that teaches ML the way physicists actually use it — not Kaggle-style, but with proper systematic uncertainty handling."
— Sample testimonial — Postdoc, ATLAS
The Research
BDTs outperform rectangular cuts on multivariate classification tasks in HEP, often improving signal efficiency by 30–50% at equal background rejection.
Roe et al., NIM A 543 (2005) 577–584
Gradient boosting (XGBoost/AdaBoost) consistently ranks among top-performing classifiers in HEP benchmarks including the HiggsML challenge.
Chen & Guestrin, KDD 2016; Adam-Bourdarios et al., JMLR W&CP 42 (2015)
What You Get
From First Principles
Module 1 builds your ML vocabulary. No assumed knowledge beyond undergraduate statistics and Python.
Real HEP Data
Work with actual collision data and Monte Carlo samples — not toy datasets.
Publication-Ready
Module 6 covers validation, overtraining checks, and systematic uncertainties — what review committees actually scrutinize.
Inside the Program
Module 1: Introduction to Machine Learning for Physicists
ML vocabulary mapped to physics concepts you already know. Bias-variance trade-off, training vs testing, and why your chi-square fit is already a model.
Module 2: Multivariate Analysis in High Energy Physics
Why rectangular cuts hit a wall in high dimensions. Feature engineering from detector observables. The geometry of signal vs background separation.
Module 3: Decision Trees — Principles and Construction
Recursive binary partitioning from scratch. Gini impurity, information gain, pruning. Build a working tree classifier on HEP data.
Module 4: Boosted Decision Trees — Ensemble Methods
AdaBoost, gradient boosting, XGBoost. Why combining hundreds of weak learners beats any single tree. Hyperparameter tuning that matters.
Module 5: Applying BDTs in LHCb Analyses
The end-to-end pipeline: simulation corrections, sPlot, PID reweighting, EvtGen quirks, and getting from a trained model to a ROC curve your convener trusts.
Module 6: Validation, Systematics, and Production-Ready BDTs
Overtraining diagnostics, k-fold cross-validation, data/MC agreement checks, cut optimization, and systematic uncertainty propagation. What makes a BDT analysis publishable.
Before & After
| Before | After |
|---|---|
| 40 hours across scattered GitHub notebooks | 6 structured modules with clear progression |
| ROC curve looks wrong, no idea why | Validated classifier with proper diagnostic checks |
| 'Figure it out' with the TMVA manual | Step-by-step from principles to production |
30-Day Guarantee
If within 30 days you feel the course hasn't improved your ability to build and validate BDTs for physics analysis, email us for a full refund. No hoops, no drama.