Recently, I have been attracted by the paper "Stable learning establishes some common ground between causal inference and machine learning" published in Nature Machine Intelligence journal. After perusing it, I met with a problem regarding the connection between model explainability and spurious correlation.
I notice that in the paper, after introducing the three key factors(stability, explainability and fairness) ML researchers need to address, the authors make a further judgement that spurious correlation is a key source of risk. So far I have figured out why spurious correlation can cause ML models to lack stability and fairness, however, it is still unknown to me why spurious correlation can obstruct the research on explainable AI. As far as I know, there have been two lines of research on XAI, explaining black-box models and building inherently interpretable models. But I'm wondering if there are some concrete explanations about why spurious correlations are such a troublemaker when trying to design good XAL methods?
Hi, I'm afraid to disappoint, but the black box model is a formal model. And the idea of false correlation is a meaningful idea (an assumption about causes and consequences). No formal mathematical methods can solve substantive issues. A meaningful model of the learning process is needed, within which the question of correlation should be discussed: false or not.
In my opinion, spurious correlation have effect from another causality(independent variable beyond) which are key risk for machine learning.
Briefly, the real phenomenon are constructed by some serial causality rather than order invariant behavior.
That is quite simple in theory and quite hard in practice. A spurious correlation appears if you investigate two events/processes etc. which are the consequences/results/manifestatin of the same reason. However, such manifestation might be absolutely conspired: at least, one mey not expect to meet it thus facing the spurious correlation. Classic example is the "brithish scientists" observation between the positive correlation between Downism and angina: they indeed exhibit a positive correlation with (obvious) no reason/consequence relation. The trick is that downism patients were included into a tested set, and such people ALWAYS have decreased immune protection, so being almost permanently sick of angina and related disorders.
- Badges
- Science topic
- Similar topics
- Biological Science
- Bioecology
- Systems Ecology
- Landscape Ecology
- Connectivity