recsyschallenge opening
TRANSCRIPT
Recommender Systems Challenge
ACM RecSys 2012Dublin
September 13 2012
Organizers
● Nikos Manouselis- Agro-Know & ARIADNE Foundation
● Alan Said- PhD student @ TU Berlin -@alansaid
● Domonkos Tikk- CEO @ Gravity R&D -@domonkostikk
● Benjamin KillePhD tudent @ TU Berlin -@bennykille
● Jannis Hermanns- Moviepilot -@jannis
● Katrien Verbert- KULeuven
● Hendrik Drachsler- Open University - The Netherlands
● Kris Jack- Mendeley
The Challenge - 2 tracks
CAMRa● Previously: 2010 & 2011● Finding users to recom-
mend a movie to● moviepilot.com data● live evaluation● camrachallenge.com
● ~60 participants● 1 submitted paper
ScienceRec● First time● Novel algorithms,
visualisations, services for paper recommendation
● Mendeley data (3 datasets)● Several requested data● 4 submitted papers
What went wrong?
● Initial results indicate that RecSys Challenge was not successful ○ measurable result: 5 submissions, 2 accepted papers + 1 accepted
presentation/talk
● Several issues encountered○ “we downloaded the dataset but could not run extensive simulations
because it was difficult to process”○ “we wanted to combine the dataset with live data from the platform
but we didn’t have enough user info”○ “we used different datasets than the ones suggested because they
were easier to access/use”○ too diverse tracks
○ unawareness / difficulty in spreading information about the challenge
What went right?
Why are we all here?
● finding datasets to experiment with (especially from live, industrial systems) instead of working with the old "favorites"
● learning how existing algorithms can be reused (extended, adapted, evolved) instead of coding from scratch
● finding how our algorithm (unique, novel, amazing, the best) can be contributed to the community conceiving designing/deploying a great recommendation service
● make a business case out of our algorithm/service● (become rich/famous/...)
The real challenge
How to make such contests work, being also useful for...
● ...the data publisher [insight into what can/cannot be done with their data]
● ...the research community [insight into new algorithms, approaches, services + contributions to existing frameworks/libraries]
● ...the deployed platform [insight into new services that could work better / be more useful]
● ...everyone [create publicity/awareness]
Our Workshop
● Follows a simple structure similar to how you would participate in a challenge○ Available Data Sets○ Existing Algorithms/Frameworks○ New Investigated Methods○ Prototyped and/or Deployed Services
Program09:00–09:15 Welcome & intro09:15–10:00 Working with Data
● The MovieLens dataset – Michael Ekstrand● Mendeley’s data and perspective on data
challenges – Kris Jack● Processing Rating Datasets for Recommender
Systems’ Research: Preliminary Experience from two Case Studies - Giannis Stoitsis, George Kyrgiazos, Georgios Chinis, Elina Megalou
10:00–10:30: Algorithms & Experiments● Usage-based vs. Citation-based Methods for
Recommending Scholarly Research Articles - André Vellino
● Cross-Database Recommendation Using a Topical Space - Atsuhiro Takasu, Takeshi Sagara, Akiko Aizawa
11:00–12:30: Real Use● From a toolkit of recommendation algorithms
into a real business: the Gravity R&D experience – Domonkos Tikk
● Selecting algorithms from the plista contest to deliver plista’s ads and editorial content on premium publisher’s websites - Torben Brodt
● Mendeley Suggest: engineering a personalised article recommender system - Kris Jack
12:30–14:30: Lunch break
14:30–15:30: Frameworks, Libs & APIs● Hands-on Recommender System Experiments
with MyMediaLite - Zeno Gantner● Using Apache’s Mahout and Contributing to it-
Sebastian Schelter ● Flexible Recommender Experiments with Lenskit
- Michael Ekstrand15:30–17:30: Hands-on work