Sramana Mitra: Can you do one more level of double-clicking? You were distinguishing between clustering algorithms versus one-on-one recommendation and treating each individual as an individual. Can you explain how that works technically?
Jana Eggers: We take all of the product that you have. Let’s take retail. You may have groups of products and offerings. All of those products are connected to each other in a weighted graph to show how similar they are to each other across all of their attributes. Most people think of a product as a molecule. Think of our products as a bunch of atoms.
Those can be connected to each other in different ways and they’re not all the same type of atoms. One may be an iPad that has a certain level of memory. There’s a bunch of information about our product that we’re using and putting all of that information in a very large weighted graph to identify the product. How kid-friendly is a product? That’s part of the weighting in our graph that we figure out using lots of information that we have.
Sramana Mitra: How does the tagging work? Is it manual? How do you figure out what gets tagged as what?
Jana Eggers: That’s part of what we do. That’s what we call the ingestion phase. We take all of the information that an enterprise has. Obviously, it can come from their product catalogs, but we can also take information from customer service for example. We can take structured data as well as unstructured data.
For one of the customers that we’re working with, we did a lot of work to crawl the web for reviews of their products. We would take that information and see what people are really saying about their products to put that into the graph, which may be a different view that they have themselves. At the higher level, each of our neurons in our brain has ten thousand to hundred thousand connections. That’s the kind of level that we’re trying to process at.
Sramana Mitra: My understanding of what I’ve heard from you so far is that the differentiation in your technology is that you’re parametrising at a much extensive level than the level at which, perhaps, your competitors are parametrising at. The second thing is you’re doing it at an individual level as opposed to a cluster level.
Jana Eggers: Yes, you’ve got it exactly right. We got this big graph of products, which like you said, is parametrised and weighted at a different level. When the customer comes in, we have all the information we can get on that customer as well. Typically you can imagine that that is low. Two signals may be like which channels they’re coming in on and which ad they want to put on. Even that to us is two signals.
Even with a couple of signals like that, we’ve shown how much we can differentiate and provide different recommendation sets for people. It’s marrying those two things where we have all the information that the enterprise has, parametrise, and connect it in a knowledge graph in an indexable way that makes it very fast to retrieve for someone coming in for whatever signals we have on them—whether it’s 20 or just two.