Manually choosing aggregation statistics (count, median, cheapest, most expensive, etc.) to summarize a customer's transactions into a single feature for churn prediction is an arbitrary, unprincipled process that a neural network attending over raw transactions can replace.
Using a churn-model example, Leskovec shows how feature engineering over relational data devolves into endless, arbitrary debates about which aggregation statistic to use, which relational deep learning avoids entirely. ✦ AI generated
Jure Leskovec · The TWIML AI Podcast · 2026-05-21 · original ↗
starts at this moment · 26:11
“It also makes me wonder if there's a way to reuse existing benchmarks by like denormalizing, you know, wide single tables or something like that. Is that something you've looked into?”
I have a customer and here are historic transactions of the customer. I need to aggregate them. So first I say I'll count how many purchases you made last month and then I'll maybe take the median price of those purchases and then you know some other data scientist says no no let's take the cheapest price of everything you bought right and then somebody says no no you should take the most expensive one
verbatim transcript · starts at 26:11
26:11>> Exactly. Exactly. And I I know I can dwell on this point a bit, right? Imagine you do you are doing a churn model, right? So I a customer churn model could be I have a customer and here are historic transactions of the customer. I need to aggregate them. So first I say I'll count how many purchases you made last month and then I'll maybe take the median price of
26:35those purchases and then you know some other data scientist says no no let's take the cheapest price of everything you bought right and then somebody says no no you should take the most expensive one and then somebody says no it's the average person says oh but distributions are skewed you should take the media then another person wakes up and says hey it's about shopping in the morning.
26:58That's what's predictive of churn. Let's add another feature, right? And then somebody says, "Oh, but we have to account for holiday. >> Give me the data." >> You know, like that's what I mean, right? And then you're like, "Oh, holidays. People sleep longer on holidays. Let's now create a new feature that accounts for holidays. Oh, but then there is summer daylight change. Let's account for that." You see how kind of
27:18ridiculous this gets? just attend over the transactions and let the attention figure out what predict >> when I introduced you I mentioned that you were co-founder at Kumo in addition to the research talk about the relationship between the research and what you're doing at Kumo >> what we built at Kumo is a commercial enterprisegrade platform that allows us to to to do large scale um relational deep learning uh models um and we are
27:46using this platform to two effects one is to allow um uh partners, customers to train uh tune single task models over the multitabular relational data. Um and I can talk about uh that part. But the recent breakthrough uh that we had and we just released um the second version uh is our what we call a relational foundation model. Um, and that's a pre-trained foundation model uh that can