Interview: Cliff Lyon, Stubhub on Mastering Recommendation & Personalization Analytics Part 2
Previous post
We discuss current trends, future vision, interesting correlations, privacy concerns, and advice for Data Science practitioners.
Cliff Lyon joined StubHub.com as the head of technology recommender systems in May of 2013, where he leads a dedicated engineering and research team in the development of applied machine learning software and services. These solutions drive dynamic content delivery in both email and site-side applications. Prior to that, Cliff was Senior Director of Engineering at CBS Interactive, where, in addition to recommendation, he delivered technology for online optimization and multivariate testing.
First part of interview.
Here is second part and last part of my interview with him:
Anmol Rajpurohit: Q4. Over the last few years, what major trends have you observed in the area of recommendation and personalization? What is your vision for near future?
Cliff Lyon: One trend I keep running into recently is deploying recommendation using search technology. Search and recommendation are both strategies for dealing with information overload. The technologies grew up differently, but it turns out that much of what search does well can be used for recommendation, too. This is pretty great, since it gives you some useful tools for free; I am used to DIY technology for recommendation. Ted Dunning, a Mahout contributor now at MapR, has some nice presentations online on this topic.
On the presentation side, given the huge growth of mobile, and the much smaller viewport, I think recommendation and search together can really help users make the most use of the small screen – the more quickly catalogs can cull low-utility suggestions, the better.
AR: Q5. Are there any examples of interesting correlations that you observed (in the pursuit of building recommendations) recently? May be, similar to the famous diapers-and-beers case.
CL: I spend a lot of time looking at recommendations I don’t understand, and trying to decide whether I just don’t have the context, or if it might be a bug. I have seen some surprising similarities for sure, but none that had the sort of explanatory power of the beer-and-diapers case. At StubHub, we sell tickets for concerts, music, and theater. As you might imagine, most often, the correlations stay in category – sports for sports, concerts for concerts, etc. But I did find it interesting that some artists cross over to sports audiences – Jay-Z correlates with the Nets, as he has connections to that organization, for example. And strangely, Jay-Z also shows up for WWE fans. Not exactly beer and diapers, but we do find cross-category correlations.
AR: Q6. What are your thoughts on the privacy concerns of online shoppers? How can the privacy concerns be properly addressed while designing recommendation and personalization systems?
CL: I think the key to this is that the system must be able to adjust to the user’s desired level of privacy. A user must have some control – “clear my personal information” is a baseline – and the system must be robust to a sudden loss of information. The system must have a strategy for handling a graduated amount of information or context, from the total stranger to the customer who has more than enough data points.
AR: Q7. What motivated you to work in Data Science? What is the best advice you have got in your career in Data Science?
CL: Like any data geek, I suppose, when I first encountered the power of machine learning as a way of extracting information from data, I was completely hooked. I’d worked with data for years to that point, doing
summaries and aggregations, and machine learning – it wasn’t "data science" then – just really amazed me. I happened to be working in the Data Warehouse group at CNET at the time, and I convinced my manager to let me start my own group, so I could become a customer of the data, and apply it back to make user experience better. Never looked back, still think it is important, and there are plenty of challenges left in doing it well.
Best advice – I feel lucky in that I had a lot of good teachers and counselors, too many to count.
One thing I learned from a teacher early on -- and this sounds simple but can be difficult for many reasons -- is to follow the data wherever it leads, which may or may not be to certainty. Often, we explore data only to find an outcome or result that challenges our basic assumptions. We need to question those assumptions, and find another model.
AR: Q8. What key qualities do you look for when interviewing for Data Science related positions on your team?
CL: I look for depth in one or maybe two areas depending on the level of the position, and passion. It is such a broad discipline, chances are you are going to be out of your comfort zone, learning new things, and you need passion to get you through that.
AR: Q9. What was the last book that you read and liked?
CL: I really liked Ted Dunning’s Practical Machine Learning; a quick read, and a really great take on the growing trend of deploying recommendation via search.
Related:
