A list of papers read in 2024 as part of a paper reading club. The papers cover data science, recommendation engines, continuous optimization, operations research, information theory and statistics.

We attempt to solve the puzzle game “Former” (a variant of SameGame) using graph search algorithms: beam search, A* search, heuristic search and Monte Carlo tree search.

Presentation slides and video of a talk I gave about notions of averages (one-parameter models), loss functions and simple machine learning problems.

Implementing the Norwegian income tax is a pain. It requires keeping track of various deductions, if-then-else rules and around 20 constants. In this article we introduce a smooth tax function, simplifying the computation to a single formula with two constants.

We shrink Dostoevsky’s “Crime and Punishment” to its limit using arithmetic coding. This article showcases my Python implementation of the algorithm.

(Norwegian) Dårlige vaktplaner kan være en kilde til frustrasjon, og gode vaktplaner har stor samfunnsøkonomisk gevinst. Matematisk optimering kan finne gode turnuser på noen få sekunder.

The Huber loss function behaves like squared loss for small values and like absolute loss for large values. In this article we reverse the pseudo-Huber, creating a smooth function that behaves like absolute loss for small values and like squared loss for large values.

(Norwegian) En analyse av resultatet av kode24s spørreundersøkelse om lønn.

Tags are used to classify and filter items in information management systems. We use optimization to find the most informative subset of tags in two real-world datasets.

Tags are used to classify and filter items in information management systems. In this article we investigate tag structures that lead to efficient information retrieval.

(Norwegian) Regjeringens nye innleieregler kan føre til at tusenvis av små, spesialiserte teknologibedrifter må legge ned.

(Norwegian) Medier rapporterer gjerne om tidsserier: dødsfall i trafikken, strømpriser, arbeidsledighet, EU-oppslutning, fødselstall, valutakurser, renter, og så videre. Ofte blir det meldt om nye rekorder, men hvor uvanlig er det egentlig at nye rekorder blir satt?

We study a dataset consisting of software developers and their skills. The data is a large binary matrix, which we factorize to predict the probability that each developer has each skill. From the model we obtain an embedding, which we use to visualize skills in a two-dimensional space.

A complete notebook with Python code, showing how to write advanced scikit-learn code, using custom Transformers, custom metrics, cross validators, etc.

Suppose you have a target variable that’s bounded, such as a score between 0 and 1. How can Ridge regression be generalized to predict values in this interval? We create some optimization routines and present full Python code.

Here are the ten commandments of data science, encapsulating what I’ve learned in my career so far.

Throughout 2023 I’ve been using math to make bets on football matches. I created one statistical model to predict outcome probabilities and one optimization model to size bets.

Given a set of football (soccer) matches and three outcomes per match to bet on, how should we place our money for maximal long-term growth? We generalize John Kelly’s idea from 1956 and solve it with modern optimization methods.

(Norwegian) Hvilke faktorer bestemmer lønna til en utvikler? Vi analyserer et datasett fra kode24. Utviklerne med høyest lønn har lang erfaring, høy utdannelse, er konsulenter og jobber i Oslo.

In powerlifting, formulas are used to assess and compare lifters with different body weights. In this article we evaluate two such formulas: Wilks and IPF GL. Then we generalize them and propose an alternative. We also create a formula that accounts for age and training experience.