Notes from Neel Sundaresan’s keynote speech at RecSys 2011
He started by stating that he won’t have any greek symbols in the talk.
Arch West was the inventor of Doritos and David Pace we the inventor of Pace sauce. What they did was that they noticed the can sell more if they advertise the two products together. There is a lesson in cross-selling and recommender systems that we can learn from this story.
eBay started when Pierre Omidyar wanted to sell his broken laser pointer. He listed the laser pointer online for 99 cents and finally sold for 14 dollar. Was wondering if the person who bought it knew that it is broken. The guy responded “Yes I am a collector of broken laser pointers”
Why people buy something? it is hard to say. Some people buy stories! remember the toast that sold for 27K that was the cheapest marketing campaign that a casino had. One man’s trash is another’s treasure.
The long tail in eBay’s context mean most people sell very few items and most of eBay’s revenue comes from these people. (i.e the mean is way larger than the median
The users constantly are running experiments to maximize their revenue. They are constantly testing to see if free shipping can sell more, different selling strategies are being tested by users at any time on eBay.
This causes an interesting behavior. If you promote a user’s product on the homepage they may increase their price! This is an interesting dynamics between the user and the seller (eBay)
One of the problems that locations like eBay have is the problem of big data. Complex algorithms are often impossible to work with in that scale. If you are looking for a job at eBay you need to know how to work with data in that scale. A goal at eBay lab is that when a new scientist joins the lab on Monday he got access to all the data by Friday.
This amount of data has changed how economics is doing experiments. They can now run experiments on 400 million data points.
What are you optimizing for at eBay? is it profit maximization? do you want to increase the shopping cart size? are you looking for maximum customer satisfaction?
The other thing is how do you measure success?
Everything we do at eBay is a recommendation.
I KEEP six honest serving-men
(They taught me all I knew);
Their names are What and Why and When
And How and Where and Who.
— from The Elephant’s Child
When we look at the tag cloud of eBay you see keywords like “used”, “vintage” and “antique” a lot more than “new”
The search is an interesting problem some people are looking for “ipod nano 4gb black new” and some are looking for the skin for their ipod. Our search engine should be able to differentiate between “ipod nano 4gb black new” “ipod nano 4gb black new skin”. This proposes a hard and challenging research questions.
Click trails can help us tremendously with building recommender systems that can capture these behavior and improve recommender systems. At eBay a data cleanup is an important part of recommender. Specially when they use click trails.
eBay has a language like pig that allows them to do pattern recognition at scale. Sometimes a search is followed by some page views and another search. This pattern is useful to do recommendation to other users who have similar initial search queries. See two recent papers from Sundaresan for the results and model.
Fashion item buyer on eBay are very brand aware. Sometimes ebay does not have enough inventories and needs to recommend proper products from outside websites.
one of the challenges at eBay is that we do not have a catalogue of items (remember the laser pointer story?) Amazon does not have such a problem, you cannot sell anything on Amazon unless it is on the catalog.
eBay uses its own matrix factorization see their ICML paper. The sparsity in eBay’s data is fascinating it is 100 times the sparsity in the Netflix data.
eBay clusters items into pseudo products using LDA. He shows an example of a recommendation for a broken blackberry cellphone.
The most important thing is “why” are you recommending this to the user and “why” they should buy it. HCI is a useful tool here. reveal to the user why you are recommending. Something like “52% of the people who bought this item also bought …” are very effective. Be very explicit on why certain recommendations are made.
Let’s look at “When”. things like reminders, post purchases, urgency, upgrades, seasonal sales fall into this. Reminders can be like “you have viewed this item” that reminds people that they can still go and buy. There is a temporal element to this problem too. They may not need the same item until after 30 days but need to buy it again after 30 days is passed.
See this wired article on persuasion based profiling and recommendation systems. (thanks to twitter).
There is a lot of seasonality on eBay. Mother and Father’s days, Christmas. There are other events that are we don’t know (so my question is how can we find them algorithmically)
We get more data from mobile devices than we get from online. It is a huge research opportunity.