In this latest instalment in our series on recommendation engines, we look at ATG - an e-commerce services vendor which, among other things, provides recommendations technology to retailers such as Tommy Hilfiger and BetterWorldBooks. ATG has a similar "blended" approach to recommendations as richrelevance, whom we profiled last week - in other words it uses a mix of personalization and wisdom of the crowds. ATG's current approach to recommendations is heavily influenced by a product it acquired in January 2008, CleverSet. We spoke to ATG this week, to find out more about their recommendations product and what makes it stand out in (what we're discovering) is a crowded market for recommendation technologies.]]>
We spoke to 3 people from ATG: Ryan Hoppe, Marketing Director, ATG e-Commerce Optimization Services; Erik Holm, Product Manager, ATG Recommendations; and joining us later in the call was Bruce D'Ambrosio, Chief Scientist and the founder of CleverSet (also a former Oregon State University computer science professor).
CleverSet, now known as 'ATG Recommendations', is described on ATG's website as "a predictive recommendations service". It's just one part of a suite of e-commerce "optimization services". The company claims that it is differentiated from other recommendations services by its focus on commerce and the fact that it is a stable, profitable company. In the 2008 year the company did $164M revenue "with profitability". That figure includes many services and licensing revenues, of which product recommedations is just one. ATG was founded in 1991 and it did an IPO in 1999, so it does appear to be more experienced than competitors like richrelevance.
Predictive Recommendations
ATG's core product is an e-commerce suite, which it says is used to power hundreds of online store fronts. Its "e-Commerce Optimization Services", which includes recommendations, can be used on other e-commerce sites as well as those powered by ATG. The company told us that their products aim to increase the following 3 things: conversion rates; order value; and customer attention.
ATG calls its approach to recommendations a "blended approach", which aims to predict what the consumer wants to buy next. The recommendations come from a combination of user, site and product data. Elements include purchase history, the time of day, where the user clicked from, what browser they use, product catalog variables, historical shopping information, click-stream data, and more. Out of all this ATG provides what it calls "predictive recommendations".
How is ATG Different From richrelevance?
ATG's approach sounded very similar to that of richrelevance. So we asked ATG: what's different? Bruce D'Ambrosio, founder of CleverSet and ATG's Chief Scientist, responded that richrelevance is similar, but that it is "mostly a subset" of what ATG Recommendations does. He said that ATG is focused on bringing the merchandiser "into the conversation with the visitor", whereas richrelevance perhaps has less focus on merchandiser and more on the user.
D'Ambrosio further said that ATG models the visitor in both current and longer term sessions. The method it uses is called "Statistical Relational Learning" (SRL), whereby ATG integrates information about the actual user with "similar visitors". It also incorporates relational data about the product and "context of use".
Another key piece of data that ATG focuses on is "engagement", by which it means what fraction of users/buyers interact with recommendations. D'Ambrosio said that they see an enormous amount of this engagement activity before users buy products.
The Netflix Prize
As an interesting aside, we asked D'Ambrosio, as a very learned and experienced engineer in this field, what his thoughts were on the Netflix Prize. We mentioned the Napoleon Dynamite problem, whereby products like that are hard to recommend against.
D'Ambrosio replied that to win the Netflix Prize will require a re-definition of the problem, by "dramatically expanding the scope of the information". He said that most of the interesting information required to do Netflix recommendations well is actually off-site - e.g. product data that's not in the catalog on Netflix. He thinks that to win the Netflix Prize, contestants must gather more information about products like Napoleon Dynamite.
Conclusion: Big Player
As well as finding that it's a crowded market, another thing we've discovered in this series is that comparing recommendation engines directly to one another is nearly impossible. Firstly because each customer has different needs and so unless the customer has tried more than one product, they won't be able to accurately compare them. But secondly, each recommendation engine has a different approach and their algorithms are complicated - and well protected. Who knows what is really under the hood of Baynote, richrelevance or ATG.
However, ATG does have a good track record behind it; and we were particularly impressed by the knowledge of Bruce D'Ambrosio - surely one of the company's best assets in recommendations technologies. So if you're an e-commerce shop looking to implement recommendations, ATG seems like a relatively safe bet.
]]>