CatBoost · AIxSci

What it is

A decision tree is a chain of yes-or-no questions about the data that ends in a guess. One tree on its own is crude. Gradient boosting grows a sequence of them: the first makes a rough guess, the next is trained on what the first got wrong, and so on, each tree nudging the running total closer to the truth. CatBoost is one of several libraries that do this. Its particular concern is categorical data, meaning values that are names rather than quantities, such as a material, a species or a sample label. Most methods need those names turned into numbers first. CatBoost converts them internally, in a way designed to stop the model from quietly memorising the answer.

Why scientists use it

Much scientific data arrives as a table: rows of measurements with a column to be predicted. On tables of this kind, boosted trees often match or beat more elaborate approaches while needing less data and less tuning, and they run on an ordinary computer in minutes. Trees also handle messy input gracefully, as they do not assume the relationship between a measurement and the outcome is a straight line, and they are untroubled by columns measured on wildly different scales. Scientists reach for CatBoost as a dependable stand-in for a physical or statistical model they do not have: something that learns the pattern from examples when the underlying mechanism is too tangled to write down.

In the record

materials-chemistryRegression models predict how graphene loading changes aluminium's electron emissionPredicted field emission behaviour of aluminium composites reinforced with graphene from a table of electric field, current density, emission current and time.structural-biologyMachine learning sorts antibody heavy chains by the germ they targetClassified which antigen an antibody targets, using descriptors computed from its heavy chain amino acid sequence.

Limits

A boosted tree model describes the data it was trained on and does not explain why. It cannot extrapolate: asked about conditions outside the range it saw, it returns the nearest edge of what it learnt, confidently and wrongly. Because each tree chases the previous one's errors, it will happily learn noise and quirks of the particular samples, so apparent accuracy can collapse on fresh data. It also gives no causal account. A column that merely accompanies the outcome will be used just as readily as one that produces it.

Written by AIxSci: the background from general knowledge, the examples from checked records.