2 ms·
AI Predicts Obesity Prevalence from Space
- tlb 8y agoPaper at https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2698635 https://jamanetwork.com/journals/jamanetworkopen/fullarticle... This paper abuses cross-validation to claim an unreasonably high accuracy. They trained their model on census tracts from 6 US cities, and tested it with 5-fold cross-validation. That means you split the data randomly into 5 groups, and train with 4/5 and validate with the other 1/5. Suppose, for the sake of argument, that census tracts come in groups that are basically identical, because they're built by the same subdivision builder and moved into at the same time, and are served by the same grocery stores and restaurants and schools, and they have similar incidences of obesity. (In US suburbia, a census tract is typically 1 block with ~30 inhabitants.) A random leave-1/5-out cross validation can report high accuracy, even with a highly overfitted network that just memorized the tracts it saw. Because the 5-fold validation still preserves at least one member from every group of identical tracts. It'd be more impressive if it could predict obesity for a city it hadn't seen.
- jmpman 8y agoCan it also predict wealth? Aren’t obesity and wealth correlated? Suspect green lawns and widely separated houses predict wealth compared to mobile homes in a barren lot packed together.