Tag Archives: Aggregate Knowledge

[repost]Introducing Aggregate Knowledge, Powerful Recommendation Engine

This week on HyveUp, I am trying out a new publishing format. The intention is to offer video content that is more adapted to the discussional Web. I basically break down the video interview into 3-4 parts, that each constitute a single blog post. Thanks for letting me know what you think.

Last week, Facebook was announcing the Live Search integration into the popular social network search utility. The reactions on popular tech blogs were almost unanimous: Why?

Chris Law is the co-Founder and CEO of Aggregate Knowledge, a Web startup that specializes in serving recommendations on medium and high-traffic sites. Previously, Chris launched Tribe.net, an early social networking site. Social graphs are poor predictors for advertisers is the major lesson he learned from this startup.

Aggregate Knowledge doesn’t tap into social graphs to serve recommendations to its visitors. Instead, it uses a complex algorithm that analyzes two major dimensions of the visitor:
1. Behavioral patterns
2. Contextual patterns
The behavioral analysis is anonymous, so it doesn’t raise the same issues Beacon did. The contextual analysis is based on semantics. By mixing those two ingredients, Aggregate Knowledge serves up quality recommendations to its clients, who just have to insert a snippet of code in their site’s sidebar to get the service up and running.

Aggregate Knowledge is a good example of a startup which development plans drift away from the social hype of the Web 2.0. In this post, Alex Iskold describes the challenges of building a recommendation engine. Since he is in the recommendation business, he has great analytical skills on that subject, and the post makes us understand how complex designing a behavioral/contextual recommendation engine can be.

In the same spirit of heavy data storing and crunching, Aggregate Knowledge’s approach is complex and powerful: Both Google and Microsoft invest heavily in behavioral targeting technologies, and semantics has been publicized as the new big trend many times (some competing recommendation engines focus exclusively on semantics).

However, Aggregate Knowledge isn’t a search tool, but a discovery tool. Discovery happens in a different context than search, a topic I will further expand on in the next post.

Read related items:
Comparing Discovery Tools (Whonu, Evri, Aggregate Knowledge)
3 Different Approaches to Automated Recommendation (Pandora, Strands, Aggregate Knowledge)
Recommendation Engines: Future for Retailers and Content Providers?


[repost]Comparing Discovery Tools (Whonu, Evri, Aggregate Knowledge)

Aggregate Knowledge, the recommendation engine, is a tool for discovery, not search. When Web users search something, it means they know what they are looking for. Recommendations lead to serendipity. There are quite a few providers of that service online, so let’s look at a few.

Let’s start by a recommendation engine which technology is very close to search. Whonu looks like a search engine. The first action on the site is to type a keywork in the search box. As Whonu’s creator Derek Franklin explains: “I felt that if I could bring together the power of all these great search tools into a single interface, I just might have the research, discovery, idea-generating tool I’ve been looking for.

Whonu mashes up different search engines (Yahoo!, Live, Google…) into one search box, which makes the experience a discovery because resulting SERPs are unique.

Evri is another discovery site. Like Whonu, Evri users start by searching a term in a search box. However, the search term gets processed in a more constructive way on Evri, and the machine delivers a whole range of useful results to discover more about the search term.

As you can see in this search for Zinedine Zidane, the graph offers contextual links to the soccer player (France, Madrid, Beckham), a box to sharpen your search, article headlines, an about section, and pics & vids results that couldn’t fit in this screenshot. In Evri, discovery happens through rich contextual results.

Both examples above require a pro-active search behavior to launch the discovery process.

In a lot of situations, you just want to step in a store and let sales persons smooth-talk you into your next purchase. You don’t know what you’re looking for, and browsing around will help you figure it out. This is the kind of discovery Aggregate Knowledge provides. By applying complex contextual and behavioral algorithms, the engine is able to serve visitors with tailored suggestions of items they could be interested in.

No initial search implied here.

Why should a Website integrate a recommendation engine into their sidebar? Visits that imply clicking on the discovery window usually generate 6 to 10 more page views than visits that do not imply the discovery box. Case closed!
Read related items:
Introducing Aggregate Knowledge, Powerful Recommendation Engine
3 Different Approaches to Automated Recommendation (Pandora, Strands, Aggregate Knowledge)
Recommendation Engines: Future for Retailers and Content Providers?


[repost]3 Different Approaches to Automated Recommendation (Pandora, Strands, Aggregate Knowledge)

In the previous post about Aggregate Knowledge, I quickly compared the recommendation engine to a few other discovery tools online (Whonu and Evri). In this post, I will compare Aggregate Knowledge to other recommendation engines to better understand how their technology sticks out from the crowd.

In the video below, Chris Law mentions Pandora, and explains how the music site is laser-focused on the DNA of the music files, a technology that enables them to recommend songs to their users based on structural data.

Together our team of fifty musician-analysts has been listening to music, one song at a time, studying and collecting literally hundreds of musical details on every song. It takes 20-30 minutes per song to capture all of the little details that give each recording its magical sound – melody, harmony, instrumentation, rhythm, vocals, lyrics … and more – close to 400 attributes!” (Pandora, about page)

This project is called the Genome Project (consult the Human Genome Project, and apply the principles for music). Basically, Pandora’s technology is based on a deep structural analysis of music files. It detects subtle musical patterns, and organizes those patterns into groups, hence the ability to recommend other songs based on the structure of one initial song.

Strands is another recommendation engine, which technology is based on social behaviors. Strands’ goal is to personalize your online experience, by understanding who likes what, and generating suggestions based on the social feedbacks of users (a technological approach very similar to FFWD’s video technology).

Strands develops technologies to better understand people’s taste and help them discover things they like and didn’t know about. Strands has created a social recommender engine that is able to provide real-time recommendations of products and services through computers, mobile phones and other Internet-connected devices.“(Strands, about page)

Strands deploys its technology through several different products: Finance, Social Media (mostly music and video items) and Business (helps people discover the content on your site). By mixing people’s likes and dislikes in their tech blender, Strands offers a powerful recommendation algorithm for their users, making them one of the leaders in recommendation technologies today.

Aggregate Knowledge is yet another way to approach the question. As you remember in the first post of Aggregate Knowledge episodes, Chris Law explained that his previous experiences taught him that the tastes of your friends usually poorly reflect yours. This is why Aggregate Knowledge is more focused on analyzing the context of the visit (traffic source, landing page, semantics, visitors’ demographics?…) and the behavior of the visitors (page views, clicks, time spent…).

From there, they run multiple algorithms on their servers’ blender and extract the best recommendation possible, to serve it up nice and fresh on their clients’ Website. As it says on the company profile on Crunchbase:

The word in Silicon Valley is that they are doing one hell of a job for their partners, which include the Washington Post and Overstock.com

to conclude, we have 3 different approaches:
– Pandora: deep structural analysis of an item
– Strands: intensive social behavior analysis around an item
– Aggregate Knowledge: structural analysis of an item, paired with behavioral analysis around the item

Read related items:
Introducing Aggregate Knowledge, Powerful Recommendation Engine
Comparing Discovery Tools (Whonu, Evri, Aggregate Knowledge)
Recommendation Engines: Future for Retailers and Content Providers?