This week The ACM Recommender Systems Conference (RecSys 17) took place. RecSys is the premier international forum for the presentation of new research and techniques in the field of recommender systems. At the conference companies such as Netflix, Spotify and of course Strossle discussed the future of recommendations. We took the chance to interview our Strossle representative Max Leander, who is a part of our machine learning team.
Hi, Max! In short – what is a recommendation system?
– A system that concludes from data to figure out what will be the most useful or interesting thing right now for a person, and then gives recommendations based on that.
Where do we see recommendation systems in our everyday life?
– In almost all content providing services. Obvious ones are of course when reading the news, where you get recommended by us at Strossle, or browsing for products online. But we are also exposed to recommendations in less obvious ways. For instance, when playing a new video game, the difficulty level might be set for you based on your playing history. It is also becoming more prominent in navigation apps. Since most people are equipped with a smartphone 24/7 nowadays, all of our movement is tracked and it turns out that much of it is very predictable.
What were the hottest topics during the conference?
– Without a doubt, it was deep learning. Deep learning has solved some problems which many thought would remain unsolved for a long time, so the hopes are up that this technique will be able to address any problem.
Whats the most fun part of working with recommendation systems?
– The most satisfying thing is when you get to see the results of a working system, not only regarding increased business value, but also when it is evident that the recommendations are becoming more relevant to the users, or in my case news readers. I love that you need to grasp a lot of interesting theory on algorithms, mathematics, and statistics. But it is equally important to understand the users you are targeting. In Strossle’s case, this translates to analyzing how people consume news and how to present recommended news articles.
What is the biggest challenge for people working with recommendation systems today?
– One big challenge is to figure out how to measure the added value of the recommendations. There is a phenomenon called cannibalization. It means that the recommendations seem to be providing value but have a negative effect on other parts of the site, for instance by “stealing” traffic. Cannibalization is challenging to measure because when you track the performance of a whole site you get a lot of data with a lot of noise.
You’re a part of Strossle’s machine learning team – how do you use machine learning to make better article recommendations?
– Machine learning is the fuel of our system. Our algorithms figure out what article has the highest probability of gaining interest.
How is that done?
– More specifically, we use it for text analysis and user behavior analysis. Our algorithms can analyze the content of articles and the past actions of users to learn things like which articles are similar, what categories a user prefers or what subjects are trending right now. This information can, in turn, be used to recommend content which is more relevant to a particular user or in a specific context.
One last thing, can you think of something in our everyday life where we could have more use of recommendations?
– I would like more recommendations in my home. For instance, when I’m cooking, the recommender could keep track of what I’ve put in the pan and recommend suitable spices. Or my refrigerator could keep track of my eating habits and remind me of when to buy new milk or even better wine on Fridays.