Skip to content
datarekha

ML

20 articles in this topic.

ML Jun 10, 2026

AutoML raised the floor, not the ceiling

AutoML can make a strong tabular baseline cheap. AutoGluon's 2025 release notes report a benchmark result under a five-minute budget, but that vendor report is not evidence of a current 2026 leader. The ceiling still depends on trustworthy data, domain features, and problem framing.

9 min read Read
ML Jun 10, 2026

Feature engineering still matters on many tabular problems

On many medium-sized tabular problems, engineered, point-in-time features can matter more than switching among strong algorithms. This is an engineering heuristic, not a universal law; TabPFN's Nature benchmark result makes the raw-versus-engineered comparison worth testing rather than settling it in advance.

9 min read Read
ML Jun 10, 2026

You can't be fair three ways — and the EU AI Act clock is ticking

A concrete lending example separates demographic parity, equalized odds, predictive parity, and score calibration, then turns their trade-offs into an operational fairness and EU AI Act checklist.

9 min read Read
ML Jun 3, 2026

Why your model made that prediction: SHAP in production

SHAP is the most-used model explainability library on earth, and most teams misread it. A field guide to what Shapley values actually compute, where TreeSHAP beats KernelSHAP by orders of magnitude, and the four ways SHAP quietly lies to you.

12 min read Read
ML Jun 2, 2026

Accuracy lies on imbalanced data

A model that never flags a single fraud transaction can score 99% accuracy — and that number is not wrong, it is just completely useless.

8 min read Read
ML Jun 2, 2026

Attention, explained without the matrices

Self-attention is a soft dictionary lookup: every word asks a question, every other word posts a label, and the answer is a weighted average of what each word actually contains.

9 min read Read
ML Jun 2, 2026

Backprop is the chain rule with good bookkeeping

Backpropagation is not a mysterious learning algorithm — it is systematic calculus, applied backward through a network, made cheap by refusing to compute the same number twice.

9 min read Read
ML Jun 2, 2026

Always beat the baseline first

A 0.91 F1 score is impressive until you discover a constant predictor scores 0.90 — which means your model's only real contribution is the gap between those two numbers.

10 min read Read
ML Jun 2, 2026

The bias-variance tradeoff, drawn from scratch

Model error is not one thing — it is the sum of two competing forces, and the art of machine learning is finding the narrow valley between them.

10 min read Read
ML Jun 2, 2026

Cross-validation: a score you didn't overfit to

A single train/test split gives you one lucky or unlucky number; k-fold cross-validation gives you five honest ones — and the variance between them tells you as much as the mean.

9 min read Read
ML Jun 2, 2026

Eigenvectors: the directions a matrix doesn't turn

Most vectors get rotated when a matrix acts on them — eigenvectors are the stubborn exceptions, and that stubbornness is exactly why they run machine learning.

9 min read Read
ML Jun 2, 2026

Feature scaling: the step KNN and gradient descent never forgive

When one feature spans dollars and another spans years, the dollar column dominates every distance calculation and every gradient update — silently wrecking models that otherwise look fine.

8 min read Read
ML Jun 2, 2026

One-hot encoding and the curse of high cardinality

One-hot encoding a zip code column turns 42,000 rows into a sparse desert — thousands of nearly-empty columns that starve your model of signal while drowning it in noise.

9 min read Read
ML Jun 2, 2026

Overfitting is memorizing the answer key

A model that aces its training data but collapses on new data has not learned anything — it has memorized noise, and the gap between its training score and its test score is the confession.

9 min read Read
ML Jun 2, 2026

Regularization is a tax on complexity

Your model will always overfit given enough rope — regularization is the mechanism that prices complexity so the model only grows elaborate when the data actually demands it.

9 min read Read
ML Jun 2, 2026

Transfer learning: standing on a pretrained model's shoulders

You do not train from scratch in 2026 — you inherit a model's years of learned representations and spend your compute budget on the narrow task that actually matters.

9 min read Read
ML Jun 2, 2026

Why neural networks need activation functions

Without a nonlinear activation, stacking a thousand layers is mathematically identical to stacking one — and every straight-line model in history could have told you that.

9 min read Read
ML Jun 2, 2026

Why we normalize: batch norm, intuitively

Internal covariate shift is not a subtle bug — it is the reason deep networks used to need weeks of careful babysitting, and batch norm is why that era quietly ended.

9 min read Read
ML May 5, 2026

Why XGBoost is still winning in 2026

Twelve years after the original paper, and despite a parade of transformer-based challengers, gradient-boosted trees remain the default for tabular data at most enterprises. Here's why the boring answer keeps winning — and what TabPFN, CatBoost, and LightGBM actually changed.

13 min read Read
ML Apr 24, 2026

LLM recommenders: Netflix, Spotify, and the post-collaborative-filtering era

Recommendation systems are absorbing LLMs, but almost never the way Twitter expects. Netflix and Spotify both run hybrid stacks — classical CF and content-based ranking as the workhorse, LLMs at the edges for explanation, cold start, and conversation. Pure LLM-as-recommender doesn't survive contact with YouTube-scale latency or freshness.

13 min read Read