Pablo Ibieta

Notes on data science, machine learning, physics, information theory, econometrics, and causal inference.

Should I retrain my score, ensemble it, or both?

information-theory

machine-learning

cmi-framework

You already have a score in production and you’ve retrained it. Should you replace the old one, or keep both and build an ensemble? At first glance those sound like the same question. They are not. On Lending Club, they give opposite answers.

Is my model’s score portable to a different target?

information-theory

machine-learning

causal-inference

cmi-framework

A conditional-mutual-information diagnostic for the question every data scientist that trains a binary classifier eventually faces: does the score I trained on one target still know something about a related target or is the apparent cross-target performance just an artifact of label overlap?

Conditional mutual information is just log-loss gain

information-theory

machine-learning

simulation

cmi-framework

Two models, a subtraction, and you’ve estimated an information-theoretic quantity. The identity that makes CMI cheap to compute — and the first post in a short series on the CMI framework.

Welcome

meta

What this blog is for, and how it’s put together.

Mutual information sees what correlation can’t

information-theory

statistics

simulation

A short note on why mutual information is the right generalization of Pearson correlation — and a simulation that shows the gap.

---
title: "Pablo Ibieta"
subtitle: "Notes on data science, machine learning, physics, information theory, econometrics, and causal inference."
listing:
  contents: posts
  sort: "date desc"
  type: default
  categories: true
  sort-ui: [date, title, reading-time]
  filter-ui: true
  feed: true
  fields: [date, title, reading-time, description, categories]
  page-size: 15
page-layout: full
title-block-banner: false
---