Single-table tabular machine learning is essentially a solved problem where methods perform about equally well; the real unsolved frontier is multi-table relational data, since flattening a database into one table discards the signal-rich relational structure.
Jure argues that no real organization keeps all its data in a single table, and that the real hard, unsolved machine learning problem lies in the multi-table relational setting, not in the already-well-studied single-table tabular ML space. ✦ AI generated
Jure Leskovec · The TWIML AI Podcast · 2026-05-21 · original ↗
starts at this moment · 22:36
“Would you say the same is true for what you're doing or is it an issue of like there was this missing link and that missing link is the graphical structure and now we have it and we're able to do much more?”
I don't know any organization that has all their data in a single table, right? So the hard part and where the information gets lost is when you go from this rich relational structure into the single table and once you are in a single table you know then we are you know then we are kind of talking almost like second order effects.
verbatim transcript · starts at 22:36
22:36from this rich relational structure into the single table and once you are in a single table you know then we are you know then we are kind of talking almost like second order effects. Did you use this architecture? Did you use that architecture? Did you use this tabular model or this table foundation model or not? Right? like all the information is there in that single table and all the
23:00methods are about equally good at extracting it right I think where the difference happens is if you actually make a step back and say hey single table model or is not the hard is not the hard part is not where is not in a sense uh general or realistic enough where you need to go you need to go to the multitable setting because that's truly now the raw data you have it's not
23:25some summarized featurized data. It's the raw data and and there is much more signal there that that got dropped when the data got flattened into summarized into a single table. So to me single table problems are are you know are solved. I think the differences are kind of second order effects. What is unsolved is the multi-table problem that that's where the vins uh are are are being hidden. And so how do you think
23:55about benchmarking performance for these types of problems? Are there established benchmarks for multitable prediction problems? >> Actually there is quite a lot of single table data out there because of all the history of um of machine learning and I think even when people develop new benchmarks from raw data they just release that single table because everyone learns on the single table right. So uh what we did actually at
24:21Stanford we were like okay so where is a multitable benchmark and there's no multitable benchmark and even if you look at kegle out of thousands of competitions of on kegle you know there are four that are multi-table all the others are features have already been engineered for you there is a single table and you start you know begging and boosting and and and creating tricks until until you win right um so we