The Benefits of Attention

---

Participants

• Session Leader ** - Paul Martino, Aggregate Knowledge

Paul Martino

Anu Nigam

Mike Prince

Erik Adibard

Pete Kazanjy

Dan Beltramo

Steve Portigal

Patrick Launay

Natalie Quizon

Asha Vellaikal

Introduction

Paul started with giving a background on Aggregate Knowledge (AK):

• Insights from his time at Tribe Networks: Attention data and finding stuff that I care about is an interesting problem.
• Again, discovery of relevant content and products on major websites is a huge problem. Solving this discovery problem leads to more sales, more page views and a better user experience.
• Search is only part of the puzzle. Attention data (behavior) is key
• Attention data is being ignored for most parts today. Largely, the data is used by marketing teams - reports for page views/session, unique visitors etc. Sometimes the product people use it for A/B testing. The customer gets no immediate benefit.
• Recommendations are the killer app of attention data
  • Attention (behavior) based recommendations
  • making your site sticky
  • promote discovery
• Web2.0 principles make this easy to do. Legacy recommedation systems involve enterprise software integration. With AK products, this can be done in real-time. Integration as easy as 2 API calls.
• Multiple examples: In Tribe Networks, such recommendations became the primary navigation metaphor - no one went through the top-down directory.
• Click-stream behavior - let users be your editors. Some people call this user-bsed editorial. People who read this also read that.
• AK can produce highly contextual recommendations based on mining attention data. Can be done for events (Mercury News), products (ABC Stores retail)
• Relative to clickstreams, ratings are not that useful in the sense of using them for computation-oriented recommendations.
• Privacy is intrinsic to the design of the system. Only keeps anonymized session data with just behavior patterns. Need to pay attention to third-party cookies and other related issues.

Q & A Session

Q: Do you have clickfraud?

A: Yes - but it needs massive volume to be successful. Would need cluster of servers all around the world with geo-targeting.

Q: Why can't you have ratings in parallel with attention-based recommendations?

A: Its subjective. For example, currently talking to Power Reviews (http://powerreviews.com), which is a web service to add comments to products to do a joint package with them. Recomputing math based on a rating on 1-5 starts is hard with more noise than signal.

Q: For MySpace, there is one distinct gesture -add me. This signifier is too broad that it means nothing.

A: Its hard to expect people to classify their social network as acquaintance, real-friend etc. And in addition, even if they do it, its hard to do computations on them. The key is using groups as a better filter as opposed to friends. The problem is to find the right filtering group.

Its hard also to add weights to the algorithm. Should Purchase data have a different rate of integrity as compared to view data. Weighting by buy data could be more useful than view data for cross-selling purposes. However, just for discovery, view data is important.

Q: Do you use the "twins" approach?

A: No, AK uses an item-based collaborative filtering solution. IBM uses another model, based on mentor-based training. Item-based filtering keeps AK clean on privacy issues which is a traditional battle. AK optimized for privacy-compliant, anonymous-item, etc. In general, performance of item-based filtering is far superior.

Q: What about suggesting things that you already viewed?

A: There is definitely a risk in item-based filtering to show products thta the user has already purchsaed. One of the biggest lessons is to tell the user why you are doing it - For example, Amazon says: "people who bought this also bought.." which is better than "we recommend you ..". Works better with user psychology as well.

Q: What are the metrics by which AK is judged?

A: Page views/session, Engagement - how often they come back in any month, How often they post back in the collective.

Discussion: Digg vs. Netscape

There is a danger providing financial incentives to only 1% of the contributors. You need all the participants in a natural ecosystem - core contributors, secondary contributors, audience etc.

Q: Incentive for fraud.

A: As AK's footprint gets bigger, there will be more incentive to game the system. However, it remains to be seen whether any clear economic incentive exists to game AK. Bottomline - any network effect business will have gaming issues. Level of effort needed to spoof AK woul dbe hard.

Asides

• According to Paul, Google hardly does PageRank anymore. On 2-3 billion pages, hard to do graph-based computation(?). Current Google implementation is a loose approximation of PageRank.

• Key pitch: Be as good as Amazon recommendations iwth just two lines of Javascript.

• 35-40% of all Amazon revenue comes from the usage of the extra systems that Amazon provides - reviews, wiki, bought this bought that etc.

• Tivo recommendations are horrible in spite of having a goldmine of clean data. Need a change of algorithm/approach.

• Multiple viewers or users in same household on the same device is one of the hardest issues.

• Guys with the most data wins the game and not necessarily the one with the best algorithms.