~/aixsci
weekly updates

learn/what the ai was for

Attribution

Attribution is the work of finding out what accounts for a result: which inputs, conditions or causes produced it, and how much each one contributed, rather than simply what the result was.

What it is

Attribution asks a different question from prediction. A prediction says what will happen; an attribution says what brought something about. There are two common versions. In the first, a model has already produced an answer, and the question is which parts of its input pushed it that way: which measurements mattered, and how much. In the second, something has happened in the world, and the question is which causes account for it. Both end in an apportionment, a division of the outcome between contributors, rather than a single value.

The usual method is comparison. You vary one input, or one supposed cause, while holding the rest as steady as you can, and see how much the outcome moves. Machine learning helps because it can hold many things steady at once. A model is fitted to everything that is known to affect the outcome, such as the weather, the time of day or the state of an instrument. Whatever the model can explain is set aside, and what remains is the share credited to the cause under study.

Why scientists use it

Scientists want attribution because outcomes in the real world are jointly produced. Air pollution falls in winter partly because of policy and partly because the wind blew differently. A signal in a telescope may come from a planet or from a flaw in the instrument. Without some way of separating these, a measured change tells you little about why it occurred. Attribution stands in for the experiment nobody can run: the same winter with the same weather but without the policy, or the same star without the instrument.

It also serves trust. A model that only gives a verdict is hard to check. If the model can show which evidence drove the verdict, a specialist can ask whether that evidence is the kind a human expert would have relied on, and can spot cases where the model is right for the wrong reason. This matters most when a decision will be acted on, or when a finding is added to a catalogue that others will build upon.

In the record

Limits

Attribution can only apportion among the causes it was given. A factor left out of the model does not vanish; its influence is quietly handed to whichever included factor it resembles, and the answer looks just as confident. Causes that move together cannot be told apart at all. Showing which inputs a model leaned on explains the model, not the world: a model may lean on something that merely accompanies the real cause. And a share of an outcome is an estimate, with uncertainty that the single number it is reported as tends to hide.

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

By field

Every article

7articles
0structural-biology
2materials-chemistry
3astronomy
2climate-weather