📚 Study Hub

Curse of Dimensionality

Machine Learning in Python · Module 1. Why high-dimensional data is "sparse" — probability of points inside a hypercube, the r for 90%, and expected squared distance.

● 4 questions · simulation + GIFs + interactive

Bias-Variance Tradeoff

Machine Learning in Python. Match the three graph lines — Test MSE (U-shape), Squared Bias (falls), Variance (rises) — with a live "fit" game.

● 3 questions · simulation + GIF + interactive fit game

Linear Algebra — Dot Products & Matrices

Linear Algebra · Learning Check 1 of 4. Dot products (xᵀy, yᵀx), matrix rules (when you can multiply/add), and AᵀA — with a dot-product machine and matrix toy.

● 4 questions · interactive widgets

Linear Regression & Coefficient Uncertainty

Linear Regression · Learning Check 1 of 7. Fit the line to 4 points — x̄, β₁, Mₓᵀ, MₓᵀMₓ, its inverse, Mₓᵀy, and what σ² means — with a regression fitter and noise simulator.

● 7 questions · Python simulation + interactive widgets

Hypothesis Testing

Linear Regression · Learning Check 1 of 5. Is a predictor significant? t-stats, β̂₁, β̂₀, RSS, and RSE — with a significance checker and RSS calculator.

● 5 questions · Python simulation + interactive widgets

EDA — College Dataset (Assignment 1)

Exploratory Data Analysis on 777 US universities: head()/info(), duplicates & missing values, tuition private vs public, "large university" mask, applications histogram, and acceptance rate — with 4 Python figures and interactive widgets.

● 15 questions · Python simulation + interactive widgets

The Silent Bug — MLSP merge crash

R package autopsy · Lab A. How aggregate()'s Group.1 cover page + one fatal rename created duplicate LAB_NUM columns and killed merge() — interactive column surgery + R outputs.

● 3 widgets · live autopsy · quiz

PLS Regression — the 1986 Tutorial

Geladi & Kowalski's classic, rebuilt with real numbers: why MLR dies on spectra (b₂ flips to −5.44 after dropping one sample), NIPALS iteration simulator, inner relation, PRESS cross-validation, and a live new-sample predictor.

● paper walkthrough · 6 interactive widgets · numpy-verified

OSSL v1.2 — Interactive Soil Spectra Data Map

All 64,484 vis-NIR scanned soil layers of the Open Soil Spectroscopy Library plotted on a real map: LUCAS (EU), KSSL (US), ICRAF (Africa/Asia), Woodwell. Filter by dataset, label completeness (SOC/pH/clay/sand), and coordinate type. Built from soillab+soilsite+visnir L0 with the corrected 400nm scan detection.

● data map · 64,484 points · dataset/quality filters

Leakage Detective Kit — Illusions of Accuracy

Learn to spot data leakage (Jang et al. 2026) using YOUR own datasets: twin spectra split across folds in CV.98, random vs grouped splits on ICRAF-952, leaky PCA simulation, and a browser Monte Carlo that shows the illusion growing. 5 guess-first cases + self-audit + cheat sheet.

● 5 kasus tebak-dulu · real-data audit · browser Monte Carlo

Predicting Soil NPK from Spectra: 11 Models, Head to Head

TabPFN vs PLS, RF, SVR, XGBoost, LightGBM, CatBoost, Cubist, MLP and a 1-D CNN on real OSSL data — identical splits, 6 preprocessing chains, full metrics (R², RPD, RPIQ). Includes SHAP/LIME profiles, and the cross-library transfer test where every model falls apart. 111 configurations, run on the UKY workstation.

● 11 models · N/K/P · 7 figures · report + PDF

The Five Validation Methods

Holdout, repeated holdout, bootstrap, 3-way holdout, and cross-validation (incl. leave-one-out) — all five run on the SAME 400 real OSSL soil spectra so you can see why they disagree. Includes the optimism problem, the test-set contamination gap, and why random folds hide a 0.52 R² drop when you hold out a whole source library.

● 5 methods · real spectra · notebook + page · 1.05 MB

Before You Read: TabPFN for Faba Bean Yield (Çilesiz et al. 2026, Agronomy)

Everything to know BEFORE reading the paper, from zero: what faba bean traits physically are and why phenological ones barely vary while harvest ones vary 10× more; what TabPFN actually is (and why "training set" is the wrong word for its context, with the paper's own three sentences as evidence); R²/RMSE/MAE/MAPE and what each punishes; SHAP vs permutation vs LOCO and why three methods are needed; Holm correction, Friedman–Nemenyi, and why only "TabPFN vs SVR" survives; plus the ablation that shows the accuracy comes from harvest-stage information. 10 parts, 15-point cheat sheet, 2 interactive toys.

● paper primer · 10 parts · 2 interactive widgets · 1 base64 figure

Wudu Masah — Panduan Praktis & Dalilnya

Masah atas kaus/sepatu: dalil dari Ibn Taymiyyah, Ibn Hazm, al-Qaradawi, AMJA, ECFR dan Dar al-Ifta; syarat sah; cara masah yang benar; batas 24/72 jam; apa yang membatalkan wudu vs membatalkan izin masah; kaus vs sepatu; dan hukum menyentuh lawan jenis per mazhab. Dua alat interaktif: kalkulator masa masah + penuntun keputusan.

● fiqh praktis · 2 widget interaktif · ringkasan + rujukan