All posts filed under “recsys

RecSys 2016 – Highlights

This year’s RecSys was jam packed and fantastic. The dual track 3-day conference took place in beautiful MIT in Boston and showcased lots of great talks around advances in recommendation system algorithms and approaches.  Below are some of my highlights from the conference.  Full proceedings are available here

#1 Awesome Industry Track

I co-chaired the industry track along with Paul Lamere, Director of Developer Platform at Echonest / Spotify and Hrishi Aradhye, Engineering Director at Google. This involved selecting and curating a set of industry track talks / speakers from a diverse range of companies who actively work in the recommender systems space.  The industry track resulted in a set of 15 talks across 3 sessions featuring speakers from Mendeley, Meetup, Bloomberg, Foursquare, Spotify, Netflix, Pandora, Stitch Fix, Expedia, Nara Logics, GraphSQL, Retail Rocket, Quora, Google and Pinterest. Here’s just some of the awesome industry track talks that were presented.

#2 Record Attendance at Women’s Lunch!

I co-organized a women’s lunch with Tao Ye, Principal Scientist at Pandora. We had a record number of women attend (almost 50). The lunch resulted in a range of action items for next year’s conference including (hopefully) the organization of day care and a listing of female speakers so that future organizers can choose from a set of talented female researchers and practitioners who want to speak at conferences.

RecSys2016_WomensLunch

#3 Combining Machine Learning with Human Curation

A prominent theme at this year’s RecSys is the combination of machine learning techniques with large-scale human curation to improve recommendations. Stitch Fix, a clothing delivery service where customers get their own personalized selected of five clothing items called a “Fix”, gave a great talk on this topic. Stitch fix employ over 3000 stylists across the globe! They have a large algorithms team who build recommendation algorithms to help the stylists choose what to send to customers. Customers keep only what they like and send back the rest so it’s important that the recommendations are good! Katherine Levins, a data scientist from Stitch Fix talked about how they approach understanding, measuring and optimizing the role of human selection in a recommendation system. It turns out that they employ a combination of cognitive research, eyetracking, and machine learning models to tune the behavior of stylists. As data related initiatives like merlin take shape in Intercom, it feels like this combination of algorithm and human curation is something we could learn from and try going forward. Katherine’s talk is available here. Earlier this year Katherine published a great related blog post.

#4 Deep Learning

There were lots of talks around deep learning at this year’s RecSys, with several papers accepted accepted to the main conference track and an entire workshop dedicated to deep learning for recommender systems. Paul Convington presented a really interesting talk on deep neural networks for YouTube recommendations. The problem he’s working on is how to predict what movie a user will want to watch next. He discussed how they’ve implemented an age feature designed to remove the bias towards recommending movies / videos from the past. What was most intriguing is that despite the promise of deep learning to advance the field of recommendation, he said that they still have to do lots of feature engineering in YouTube. And overall they have found that user’s interactions with similar items are the best features for improving recommendation.

#5 The Explainability Spectrum & Signal Decay

Shashi Thakur, a Distinguished Engineer and head of the Google Now team gave an interesting keynote on personalization, recommendation and exploration in Google Now. Shashi talked about the importance of setting the right expectations for users and explaining why recommendations are being made at a given point in time. He talked about the explainability spectrum where on one end a recommendation is so clear that it doesn’t require any explanation, and on the other end the recommendation is higher risk / less clear and so needs a clear, concrete explanation. Depending on the circumstances surrounding the recommendation and the person for whom the recommendation is being made, the spectrum of explainability will shift in one direction or the other. I thought this was interesting to consider in the context of merlin and how we surface merlin’s recommendations.

Another interesting point raised across a few talks (including Google and Netflix) related to signal decay. The general advice is to revisit recommendation features often because signals that were once useful may not longer be as useful / as impactful as when the recommender system was first built.

2016 Conference Org – Mobile HCI & RecSys

2016 looks set to be a busy year for conference organization with 2 super exciting announcements to make.

(1) I’m serving as paper co-chair for ACM Mobile HCI, a conference very near and dear to my heart. My fellow papers chairs are Jonna Hakkila from University of Lapland (Finland), Antonio Kruger from DFKI (Germany) and Marcos Serrano from University of Toulouse (France). Please check out the Mobile HCI 2016 website and get thinking about your paper submissions! Deadlines for paper submissions are 12th February 2016.

(2) I’m also serving as industry track co-chair for ACM RecSys 2016 alongside. My fellow chairs are Paul Lamere from Spotify and Hrishi Aradhye from Google. Our aim is to devise a super interesting industry line up so if (a) you work in industry and (b) your work involves “recommending things to people” why not consider submitting a proposal to the track?

Frappé: Paper on arXiv & Context-Aware App Usage Dataset Release

During my last few months in Telefonica Research in 2013 I worked with wonderful colleagues and RecSys gurus Linas Baltrunas and Alexandros Karatzoglou along with scientific director Nuria Oliver on a context-aware mobile app recommendation service called Frappé. Frappé was specifically designed to support novel app discovery experiences. In order to assess it’s effectiveness we deployed Frappé in-the-wild on Google Play and ran a smaller-scale user study with 33 users designed to evaluate user perceptions of using and engaging with an app recommendation service.

Yes, it’s been a while since working on this specific project, however, I have 2 very exciting announcements to share about Frappé.

  • Firstly, a paper describing the Frappé application, our large-scale Google Play deployment and insights from our smaller scale user study has been published on arXiv. In particular we describe actionable lessons learned related to designing, deploying and evaluating mobile context-aware recommender systems in-the-wild with real users. Details and PDF are available here.
  • Secondly, we have released the anonymized Frappé data set!! It can be downloaded from Linas’s website HERE. The dataset contains 96,202 records by 957 users for 4,082 apps. We’re very excited to see what the RecSys and Mobile HCI communities end up doing with this rich dataset, in particular in terms of pushing the envelop in the context-aware recommender systems domain.

If you end up using the data, we ask that you please cite the following paper:

@Article{frappe15,
title={Frappe: Understanding the Usage and Perception of Mobile App Recommendations In-The-Wild},
author = {Linas Baltrunas, Karen Church, Alexandros Karatzoglou, Nuria Oliver},
date={2015},
urldate={2015-05-12},
eprinttype={arxiv},
eprint={arXiv:1505.03014}
}

Happy researching folks!!!