Baynote: Does Focusing on Real-Time Behavior Trump Amazon's Technology?
Baynote is one of a number of recommendation technology providers which licenses its product to commercial companies. As we'll see in this article, Baynote places particular emphasis on real-time user behaviors - which Baynote claims goes beyond Amazon's "first generation" approach to recommendations. One thing that we've discovered so far in our series on recommendation engines is that every company in this market - including those which create their own platform, like Amazon and Netflix - have differing approaches and ideas on what makes a good recommendation engine. We spoke to Baynote founder and CEO Jack Jia, to find out why he believes their approach trumps Amazon.
Baynote's focus is to analyze the current behavior of a community to come up with recommendations - it purposely puts less emphasis on past user actions such as page views and purchase history.
In positioning his company Baynote against the "king of recommendations" Amazon (as we've termed them), Jia told us that Amazon is a first generation recommendation engine because it still largely relies on collaborative filtering. The gist of Jia's issue with collaborative filtering seemed to be that it focuses too much on analyzing past user data - it doesn't account enough for real-time user behavior, which is what Baynote does.
As Jia described it to us, Baynote observes real-time user behavior on a site and looks for implicit, emergent patterns. It uses collective intelligence and an affinity engine to analyze the data. Common behaviors which it tracks include page refers, queries, mouse movement, time spent on a page, peer behavior (see note about communities below).
The notable exceptions, which Baynote doesn't account for as much as other recommendation engines, are page views and purchases. Jia told us that Baynote considers those two things "not very useful in recommending products or content" because they can lack contextual relevance and aren't necessarily real-time. User engagement, he said, is the true value of measuring what content is useful.
A note on how Baynote defines communities. Basically the system splits users into various communities and sub-communities. Let's take the example of cameras. As Jack Jia explained, cameras can be split into many categories and sub-categories. There may be a community of Nikon cameras, and within that a sub-community of 'high-end' Nikon cameras, and so on.
What Makes Baynote's Personalization Different to Amazon's?
It's obvious that Baynote has some complex technology under its hood. That includes a concept that Baynote calls UseRank - obviously named after Google's famous PageRank, but Baynote's UseRank is focused on user behavior on a website (rather than links). Baynote has patents pending on this technology, which it uses to achieve "real-time contextual personalization". But what makes this form of personalization better than Amazon's?
Jack Jia told ReadWriteWeb that the "personalized profile approach", which the likes of Amazon use, isn't something Baynote puts emphasis on. Baynote's position is that traditional personalization doesn't work, because those approaches don't take into account user intent and context. A classic example is one that we outlined in our post 5 Problems of Recommender Systems: just because you bought something yesterday, doesn't mean you want to buy a similar thing today (e.g. I might buy a birthday present for my sister one day, and a new CD for myself the next). This is why Baynote feels the need to focus on real-time intent and context. As an example Jia showed us how one of their customers uses Baynote technology to recommend items to users.
UPDATE: the below paragraph doesn't refer to Sun & Ski, but another Baynote customer that Baynote showed us at the time of the interview. Baynote has subsequently requested that the original company be removed from the post. We've left the text as-is, however it doesn't refer to Sun & Ski.
In the example, users are shown a "you may also like" section to the right. We asked how this was different to Amazon's "Customers Who Viewed [or Bought] This Item Also Viewed" recommendations? Also we wondered if there might be a bias in Baynote's top recommendation, as most users would naturally click the top item first. To the latter point, Jack Jai reminded us that Baynote disregards page views. He further explained that Baynote measures actions and engagement on pages - so for example if a user clicks on the top link, but then clicks the back button straight away, that would be counted as a negative user measurement in its recommendation engine. But if the user scrolls on the page, highlights things, etc, then it's a positive measurement. Baynote is not just tracking clicks, but engagement on the page. It says this this approach has led to increases in sales for [above customer] and other retail clients.
Conclusion: Don't Count Personalization Out Just Yet
Baynote has about 200 eCommerce, media and enterprise clients - half of those are in the B2B sector. So it's not just the retail sector in which recommendations can be deployed. For example, Baynote told us about the use of its recommendation technology on Motorola's intranet - where the system observes employee activities, so that it can recommend content and improve the search experience.
Baynote is also an example of a Software-as-a-Service. It can be easily implemented, is on-demand and Baynote claims that it works with any site and search engine.
Overall, we were impressed with Baynote's recommendations system. However, we're not necessarily convinced that it trumps the personalization approaches used by the likes of Amazon and richrelevance (founded by an ex-Amazon executive). We're certainly not ready to write off Amazon's recommendations as mere "first generation" technology quite so quickly as Baynote. Further, we think that up-and-coming companies such as richrelevance, which we will profile shortly in this series on recommendation engines, have interesting innovations in the area of personalized recommendations.
So, there's room for many approaches - although we hold Baynote's technology in high regard, based on what we learned about it. Tell us your thoughts on Baynote's approach in the comments below.
Share
Facebook
Comments
Subscribe to comments for this post OR Subscribe to comments for all ReadWriteWeb posts
We're experimenting with recommendation engines too, but I am confused about Baynote's strong "real-time" claims.
Is Jia saying real-time when he simply means that their algorithms have a decaying recency bias? i.e. Baynote gives newer information a higher score than older information?
When I read Baynote patent apps, it seems they make a big deal out of preferring newer information, which is pretty standard. Frankly I'd be shocked if Amazon doesn't employ the same techniques.
And I think it's a safe bet that Amazon's recommendation engine been improved and upgraded more than a few times since its first generation. But then the RR guys would know for sure.
So if we say that Amazon also decays old information, what then is Baynote's trump? That they track when a user scrolls the page but not actual page views or purchases?
Is anyone else confused?
From all 3 baynote patent apps:
I believe Baynote's trump card is the ability to focus on real-time intent and context. As mentioned in the article, a user can have many contexts in which they act during a particular day, week, or month. In any given day I am a husband, dad, programmer, Sales Engineer, technology enthusiast, sports fanatic, etc. I might go to a Walmart.com one day to try and find my daughter the newest Elmo doll. However, the next time I go to the site (that same day or a month from now) I may be looking at a new tech gadget that Walmart is selling. Baynote is able to understand that I am not there looking for kids toys and show me relevant recommendations to my current browsing behavior. Other recommendation engines tend to focus on my past behaviors (the Elmo doll) versus Baynote which focuses on my current intent of why I am on the site today. I feel this is very powerful in helping to drive a better experience for customers and ultimately lead to more sales in a B2B scenario or a more satisfied employee in an Intranet environment.
I should disclose I work for a vendor that has an OEM relationship with Baynote. So, a large portion of my experience with recommendation engines is with Baynote. However, as we talk to customers, the real-time intent and context based recommendations is quickly seen as important and beneficial to customers trying to drive a more engaging experience on their site.
Israel, your points are well taken. Amazon's approach is certainly not dated, but it is different. The time of observations make is just part of the picture, as noted in the article above. Baynote takes a different approach on how to observe users as well as when in time to observe those users. By focusing less on clicks and purchases, Baynote focuses more on the not so obvious actions of users that express their intent in more subtle ways. This data in aggregate paints a fuller picture of user interactions with a website's content, and it is those interactions between users and content that expose the many different contexts on a website. These different contexts are Baynote's main targets.
With that said, Baynote does not completely ignore personal preferences of users exhibited explicitly within a user profile or implicitly through pass purchase decisions. This information can be used as a filter to combine context with the personal bias of a user.
@Mark, @Warren - I agree that sometimes being more aggressive at decaying the weight of older data can certainly be useful. Sorry if that wasn't clear.
Still, I don't quite follow how subtle signals like "highlighting text" are more reliable inputs than clicks or purchases but I guess that's why customers like to run A/B tests between recommendation engines: Are KFC's 13 secret herbs and spices really better than the special sauce on a Big Mac? Test both on your customers!
Good luck and thanks for the comments.
they should just give their software away for a while to build up their data set.
Congratulations on your emergent technology. In the race to be King of the Hill, I think focusing on merit instead of "the newest, brightest, shiniest" is just as important i.e. collaborative filtering still has its place in the sun. For example, Cogenuity uses collaborative filtering. You can check out its beta at http://dynamicalsoftware.com/cogenuity
Great in helping to drive a better experience for client and ultimately lead to more sales or a more satisfied employee in an Intranet environment. Baynote is also an example of a Software-as-a-Service. One aspect of this adaptation involves giving precedence to usage patterns that occur more recently over those that occur in the past. It can be easily implemented, is on-demand and Baynote claims that it works with any site and search engine.
Nice to see a tutorial website that realizes good design. Thank you so much for everything you do to improve our creative.