KXEN Interview: Eric Marcade on Recommendations, User communities, Kaggle MSD, Predictive Offers

KDnuggets talks with KXEN CTO Eric Marcade on new KXEN Recommendations, integration with business rules, successful applications, KXEN results in Kaggle MSD (6.6x more accuracy than the baseline in only 4 hours) , Predictive Offers, and more.

Gregory Piatetsky, April 13, 2013. Here is the first part of the Eric Marcade (KXEN) interview, which also includes a longer bio. This is the second part of the interview.

Erik MarcadeErik Marcade, founder and CTO of KXEN, is responsible for software development and information technologies. He has over 25 years of experience in the artificial Intelligence and machine learning industry.

Gregory Piatetsky: How do you find user communities and how you integrate community information with recommendations?

Erik Marcade: Good question! Using social network analysis, KXENInfiniteInsight® Social analyzes transactional data to "link" individuals together, such as if they bought (or clicked on) the same products. InfiniteInsight® then analyzes this graph and automatically detects "communities", using a modularity optimization technique. A community is a group of people who are linked, or connected, because they have similar behaviors.

We then match communities to a company's existing metadata (like product categories) to profile the communities. Similarly to the propensity scores, this information can be added to the rulesets that were previously built, and the probability of each rule will be modified.

GP: How do you integrate recommendations with business rules? Can it be automated?

EM: This is a crucial point. You very often need to fine-tune your recommendations by including additional characteristics and constraints like product brand and category, margin, price or inventory. As I said, because InfiniteInsight® Recommendation was designed for flexibility, it's easy to blend rulesets with your existing applications and processes and into major e-Commerce platforms and content management systems.

As I mentioned, when we founded KXEN our philosophy was to make analytics usable and accessible by the business. In the case of recommendations, this means being able to easily integrate resulting rulesets into your production environments in a timely fashion so our customers can get quick time to value. For example, Allocine, a movie database in Europe with over 35 million unique visitors, built movie recommendations with KXEN and integrated them into their content database in only two months with no prior analytical experience. By boosting total pages views, Allocine realized a 9% increase in advertising revenue. That's what I mean by time to value! It's core to KXEN, it's in our blood!

GP: Tell us about Kaggle MSD competition and KXEN results there. How automatic was KXEN submission and how much manual tuning did it require?

EM: Kaggle's Million Song Dataset (MSD)Kaggle's Million Song Dataset (MSD) Challenge was a competition to build a music recommendation system to predict the 500 best songs to recommend to each user.

The data that was provided was mostly "which user listens to which song", consisting of 48 million records (user, song, count) gathered from the listening histories of 1.2 million distinct users. The aggregates themselves make up a 280 GB file.

With a beta release of KXEN's InfiniteInsight Recommendation, we placed #7 worldwide in this competition in August 2012 among 153 distinct team entries. KXEN produced recommendations 6.6x more accurate than the baseline in only 4 hours. Today, six months later, using the GA release of InfiniteInsight® Recommendation, we're producing results that would have taken the #3 spot in only a few clicks! It's awesome.

GP: Which companies use KXEN recommendations now? Can you give us an example from Allegro?

EM: We've had some killer results with our customers. For example:

  • AllegroAllegro - the biggest non-eBay marketplace worldwide delivers over 80 million personalized product recommendations to its 75 websites daily. As a result, the company has increased page views per visit by 30% and click-through rate on recommended products by 500%.
  • Allocine - the second largest online movie information service in the world, serves 600 million monthly movie recommendations to its 35 million unique visitors, increasing pages per visit and as a result advertising revenue by 9%.
  • Skyrock.com - makes friend recommendations to the social networking site's 18 million unique visitors, driving up connections among community members and resulting in increased stickiness of its site.

The red box on the figure below highlights KXEN-generated recommendations.

Allegro recommendations

GP: What is KXEN's Predictive Offers™ and how does it differ from recommendations?

EM: Predictive Offers™ is a Cloud-based packaged application that we launched a few weeks ago. It's available today on Salesforce.com's AppExchange, and aims at personalizing customer conversations with a real-time "next best activity" approach. Predictive Offers™ takes all the rich CRM and social profile data residing in Salesforce to deliver personalized offer recommendations, with minimal set-up. In contrast to rule-based offer systems which are complicated to setup and maintain and often are based on intuition or outdated analysis, a predictive model-based next best activity solution learns by itself, analyzing all available information to create a mathematically optimal score to determine the next best offer. Predictive Offers is best suited for recommendations in a CRM environment where you may have hundreds of offers and want to serve up the best one to be presented in front of a customer, say on the phone by a call center agent.

InfiniteInsight® Recommendation is an on-premise solution. It typically connects to the data warehouse or the Big Data infrastructure, analyzes many data sources, no matter what the size, and builds rules, as I have explained before. InfiniteInsight® Recommendation produces personalized rulesets that will then define the best recommendations for, say, a visitor on your website and is well suited for business scenarios with thousands or millions of distinct items or SKU's - like products, movies or music.

GP: What recent book have you read and liked?

EM: The last technical book I read was "Programming Hive" by E. Capriolo, D. Wampler and J. Rutherglen. I often go back to basics and read books like this when I'm traveling and have downtime on the plane (being French, I do a lot of travel to the US and internationally) as well as zillions of articles I print from the Web (such as a very interesting one called "Scalable K-Means++" from Bahmani and al)...

I'm in the middle of that long string of books... I like to read things that can be subject of conversation with my two teens at home... and I like to read several books at the same time so I can choose the atmosphere of the night... (Besides Platon's books which lead to very different lively conversations). For instance, I very much like "Game of Thrones" and another book I'm currently reading is called "Le Roi des Ombres" in French describing the living of people building the Versailles castle for Louis XIV. It gives you a feeling of the pain (and what would be nowadays unacceptable treatments) to go through to achieve beautiful things.