On Sept 25th, 2020, the PodRecs: Workshop on Podcast Recommendations at ACM RecSys 2020 introduced researchers in other domains of RecSys to the special characteristics and challenges of podcast recommendations, and showcased the leading research being done in this area. We hope this workshop will help grow a community of researchers and foster an active research and innovation in the field of podcast recommender systems.
2. The average podcast
listener spends 6 hrs
and 39 minutes
listening to podcasts
Up from 90M in 2019
Edison Research Infinite Dial 2020
Why Podcasts?
Up from 900k in 2019
listennotes.com
eMarketer.com Edison Research Infinite Dial 2020
There are more than
1.6M podcasts in the
world
Around 104M people in the
US listened to a podcast in
the last month
Podcast ad spending in
the US forecast to
surpass $1B in 2021
3. Unlike linear radio, podcast
listeners have full
on-demand control, and can
choose exactly when and
how to listen and binge on
their favorite shows.
Though podcasts cover a
broad variety of topics and
genres like TV and film,
they lack visuals, and are
generally more host-driven
than video entertainment.
Podcasts have much longer
running times than music,
are mostly speech, and are
released more often than
music.
Not quite music Not quite radio Not quite video
Familiar, yet unique
4. Research Challenges
Podcasts are longer in duration, and many are released more
often than music or video. People also listen to podcasts in
different contexts than other media, often while doing something
else How do we need to adapt traditional RecSys techniques to
these new interaction patterns?
New interaction
patterns
Since podcasts are an audio, and mainly spoken format, many
techniques from the audio and NLP domain such as speech
recognition, summarization, and sentiment analysis would seem
to apply. But how are podcasts different from news, websites,
and other audio and text media?
Content
understanding
Many podcasts are used as sources of information or education.
However, like video, podcasts can sometimes contain
controversial and, in some cases, harmful content. How can
recommender systems ensure listeners are getting accurate and
unbiased information, and avoid rabbit-holing and radicalization?
Fairness and safety
6. Program Committee
Benjamin Carterette, Spotify
Christophe Charbuillet, Spotify
Maarten de Rijke, ICAI & University of Amsterdam
Maria Eskevich, CLARIN ERIC
Ben Fields, British Broadcasting Corporation,
Amit Goyal, Applied Machine Learning Lead, Amazon Music
David Graus, Randstad Groep Nederland
Gareth Jones, Dublin City University
Matthew McCallum, Pandora
Zahra Nazari, Spotify
Sole Pera, Boise State University
Özlem Özgöbek, Norwegian University of Science and Technology
Massimo Quadrana, Pandora
Scott Waterman, Pandora
Hao Wu, Apple
Hamed Zamani, Microsoft Research
7. Accepted Papers
The Spotify Podcast Dataset
by Ann Clifton, Aasish Pappu, Sravana Reddy, Yongze Yu, Jussi Karlgren, Ben Carterette, and Rosie Jones (Spotify)
Trajectory Based Podcast Recommendation
by Greg Benton (NYU), Ghazal Fazelnia, Alice Wang, and Ben Carterette (Spotify)
A Baseline Analysis for Podcast Abstractive Summarization
by Rachel Chujie Zheng, Harry Jiannan Wang (U. of Delaware), Kunpeng Zhang (U. of Maryland), and Ling Fan (Tongji U., China)
A Review of Metadata Fields Associated with Podcast RSS Feeds
by Matthew Sharpe (Spotify)
PodSumm: Podcast Audio Summarization
by Aneesh Vartakavi and Amanmeet Garg (Gracenote)
8. Keynote
What do we really know about podcast listeners?
In this talk, I'll discuss lessons learned from my time as a research intern with Mozilla's podcast
and voice initiatives. Combining quantitative surveys of podcast listeners and qualitative
interviews with podcast enthusiasts, I'll consider how podcast listeners aren't a homogenous
population. Some of the topics I'll touch on include: how podcast listeners discover new content,
how the the listening practices of podcast newcomers and seasoned podcast listeners differ,
and how people "listen" to video content on platforms that weren't created specifically for
listening.
Jordan Wirfs-Brock
University of Colorado, Boulder
@jordanwb
9. Schedule
14:00 - 14:10 UTC Opening Remarks
14:10 - 14:30 Paper 1: A Review of Metadata Fields Associated with Podcast RSS Feeds, Matthew Sharpe (Spotify)
14:30 - 14:50 Paper 2: The Spotify Podcast Dataset, Rosie Jones (Spotify)
14:50 - 15:10 Paper 3: A Baseline Analysis for Podcast Abstractive Summarization, Rachel Chujie Zheng (U. Delaware)
15:10 - 15:40 Break
15:40 - 16:40 Workshop Activity: Podcast Recommendation Game, Rosie Jones (Spotify)
16:40 - 17:10 Keynote: What We Know About People Who Listen to Podcasts, Jordan Wirfs-Brock (U. Colorado, Boulder)
17:10 - 17:30 Paper 4: PodSumm: Podcast Audio Summarization, Aneesh Vartakavi (Gracenote)
17:30 - 17:50 Paper 5: Trajectory Based Podcast Recommendation, Greg Benton (New York University)
17:50 - 18:00 Closing Remarks
10. Activity
What are 3 podcast shows that you really like?
Just think about it.
11. Activity
At 15:40 UTC we’ll have an interactive human podcast
recommendation game!
Check out the Whova chat for Zoom link.
13. Schedule
14:00 - 14:10 UTC Opening Remarks
14:10 - 14:30 Paper 1: A Review of Metadata Fields Associated with Podcast RSS Feeds, Matthew Sharpe (Spotify)
14:30 - 14:50 Paper 2: The Spotify Podcast Dataset, Rosie Jones (Spotify)
14:50 - 15:10 Paper 3: A Baseline Analysis for Podcast Abstractive Summarization, Rachel Chujie Zheng (U. Delaware)
15:10 - 15:40 Break
15:40 - 16:40 Workshop Activity: Podcast Recommendation Game, Rosie Jones (Spotify)
16:40 - 17:10 Keynote: What We Know About People Who Listen to Podcasts, Jordan Wirfs-Brock (U. Colorado, Boulder)
17:10 - 17:30 Paper 4: PodSumm: Podcast Audio Summarization, Aneesh Vartakavi (Gracenote)
17:30 - 17:50 Paper 5: Trajectory Based Podcast Recommendation, Greg Benton (New York University)
17:50 - 18:00 Closing Remarks
14. Activity
What are 3 podcast shows that you really like?
Think about it, and write them down somewhere.
16. Research Challenges
Podcasts are longer in duration, and many are released more
often than music or video. People also listen to podcasts in
different contexts than other media, often while doing something
else How do we need to adapt traditional RecSys techniques to
these new interaction patterns?
New interaction
patterns
Since podcasts are an audio, and mainly spoken format, many
techniques from the audio and NLP domain such as speech
recognition, summarization, and sentiment analysis would seem
to apply. But how are podcasts different from news, websites,
and other audio and text media?
Content
understanding
Many podcasts are used as sources of information or education.
However, like video, podcasts can sometimes contain
controversial and, in some cases, harmful content. How can
recommender systems ensure listeners are getting accurate and
unbiased information, and avoid rabbit-holing and radicalization?
Fairness and safety