In the next Recommender Systems Amsterdam - the first organised on meetup.com - we will have a combination of academic and industry talks. Nava Tintarev from TU Delft will describe interaction paradigms and explanation methods for recommender systems, Daan Odijk will discuss personalisation at Blendle, and Barend Linders will discuss online recommendation at the Dutch public broadcast, the NPO.
Doors open: 18.00. XITE will provide snacks and drinks.
Talks start: 18:30.
1. Nava Tintarev, 'Explain yourself! Arguing with Recommender Systems'
Nava is is an Assistant Professor and Delft Technology Fellow in the Web Information Systems group, Faculty of Electrical Engineering, Mathematics and Computer Science at TU Delft.
The complexities of many advice-giving systems often lead to people struggling to establish why a system chose what it did, to identify which alternatives were considered, and to determine why these alternatives were not selected or suggested. In other words, such systems are opaque, and a human (and particularly a non-expert) often struggles to understand their reasoning. During her talk Nava will introduce interaction paradigms, and methods for generating explanations (text and graphics) for recommender systems. She will also address how explanations can be designed to not only improve trust and transparency, but also improve the discovery of novel content and help users identify their own blindspots.
2. Daan Odijk, Blendle, Real-time Recommendations for News
Every morning, at Blendle, we have a huge cold-start problem when over 8.000 new articles from the latest newspapers arrive in our system. These articles are read by virtually no-one yet when we generate personalized newsletters for over a million users. We can thus not rely on collaborative filtering, nor can we use the popularity of the articles as clues for what our user might want to read. We overcome our cold-start problem by a mix of curation by our editorial team and automated content analysis using enrichments such as named entities, semantic links, authors, the language and plenty of stylometrics. Our editorial team get up at around 5am and is done reading and recommending their selection of articles around 7am, which is also the time we would ideally send out the newsletter. Starting a batch process only then would mean a prohibitively long delay. In this talk, I will outline our solution for real-time recommendations to address both challenges, based on a streaming infrastructure with Kafka at the core.
3. Barend Linders and Robbert van Waardhuizen, NPO, A sneak peek into the recommender system of the NPO
At the Dutch Public Broadcasting Organisation (NPO) the Marketing Intelligence team is responsible for producing the recommendations served on the online portal (npo.nl (http://npo.nl/) and corresponding apps). This is currently done via collaborative filtering on a series level (a stream or episode is part of a series). By using the instances of what content is watched together by users we calculate what content is expected to match what is currently being watched. I will discuss some of the limitations of this approach and future improvements we would like to make, as it is a continuous work in progress. Finally we will discuss current explorations on how to make our recommendations more diverse. Since it is part of our mission to ‘connect and enrich the Dutch audience…’ we want to showcase the immense diversity present in our video content.
Barend is a Hydrogeologist turned Data Scientist currently working within the Marketing Intelligence team at the NPO. Robbert van Waardhuizen is currently doing his Data Science Msc. thesis research at the NPO and will talk about his improvements on the algorithm to suggest more diverse recommendations.