=Paper=
{{Paper
|id=Vol-1533/paper7
|storemode=property
|title=Stating and Discussing Challenges of Radio Recommender Systems in Contrast to Music Recommendation
|pdfUrl=https://ceur-ws.org/Vol-1533/paper7.pdf
|volume=Vol-1533
|authors=Stefan Hirschmeier,Daniel A. Döppner,Detlef Schoder
|dblpUrl=https://dblp.org/rec/conf/dmrs/HirschmeierDS15
}}
==Stating and Discussing Challenges of Radio Recommender Systems in Contrast to Music Recommendation==
https://ceur-ws.org/Vol-1533/paper7.pdf
Stating and discussing challenges of radio recommender
systems in contrast to music recommendation
Stefan Hirschmeier, Daniel A. Döppner, and Detlef Schoder
University of Cologne
{lastname}@wim.uni-koeln.de
Keywords: Radio Recommendation, Audio-Text Recommendation, Recom-
mender Systems, Public Radio, Interactive Player design
Abstract. Recommendation for public radio is a challenging area. In contrast to
music recommendation, which has been intensively discussed in research as well
as in practice, recommendation for radio has some special additional challenges
as it mixes diverse formats, such as news, articles, audio dramas, biographies,
etc. Focusing on the interaction design of a radio player, we bring up and discuss
these special challenges and find that in a radio recommender context, optimal
interaction design has conflicting issues with optimal algorithmic design.
1 Motivation and problem
While the entertainment industry has created popular apps, such as Spotify and Netflix,
with powerful recommendation systems behind them, public radio has received less
attention in recommender systems research. As in a lot of countries, people pay for
broadcast receiving licenses in order to finance public radio, and there is always a latent
need to come up with appealing solutions that provide value as perceived by the user.
Apart from the public value provided by public radio programs, individuals’ apprecia-
tion of public radio depends on not only the content but also, and perhaps more im-
portantly, the user’s interaction with the radio application.
What makes recommendation approaches for public radio so much different from
the well-known approaches for music, movies, or TV recommendation, and what im-
plications do these differences have for the design of recommender systems?
Challenge 1. First, from a content perspective, radio is a mixture of diverse for-
mats, such as news, articles, audio dramas, biographies, long features, talks, interviews,
and—last but not least—music. Additionally, live formats, like dial-in games, are an
option as well. A direct implication of these heterogeneous formats is that each format
requires its own recommendation algorithms. For instance, as research on recommen-
dation has brought up the discussion of where and whether to draw the border of a filter
bubble [1,2], in the context of news, this has to be considered carefully for a radio
broadcaster with a public mission. In contrast, recommendation for music is deliber-
ately intended to build a closed world and filter out music that the user dislikes. Even
interaction behavior might differ for each format. This mixture of formats makes build-
ing a recommender system for radio significantly different from building recommender
systems for homogeneous content types like movies, series, or music.
Challenge 2. From an interaction perspective, portable radio accompanies the user
everywhere while he/she does other things, such as driving or washing dishes. In the
core use case of radio listening, there is no choice process as would be the case for a
“lean-back experience” (e.g., browsing), a term coined by movie recommender sys-
tems. That is, because users do not spend two hours with one choice (as in the case of
a movie) but rather get a stream of audio pieces that are a couple of minutes in length
each, users are less willing to spend time with their choice. Also, there are no interactive
recommendation games like in Netflix. Rather, the interaction with traditional linear
radio is reduced to the max: tune in and tune out; listen or ignore. In order to stay with
this paradigm, interactive radio must stay as simple as possible. Few interaction possi-
bilities, however, result in few (distinguishable) data.
Challenge 3. One-time consumption. Whereas users do not mind hearing songs
they like over and over again, they do not usually feel this way about news. Users might
want to repeat news in some cases, but generally not more than one time. Whereas
eliminating duplication and not playing the same content again sounds simple, it is not.
It is challenging to decide whether a news audio piece provides new details compared
to other news that has already been played or whether it is better to skip that particular
piece.
Challenge 4. There is a steady content stream of new media coming in every day
that a radio recommender system has to cope with. Whereas a lot of recommender sys-
tems researchers tackle the “cold-start problem” [5,6]—when little is known about the
interaction of users of new media—radio recommender systems also face another prob-
lem: after four to six weeks, the content disappears from the media library again, re-
ducing the value of past interaction data for, as an example, collaborative filtering.
Challenge 5. Context-Dependency. Radio is highly contextual, depending on the
day of the week, the time, user mood, and the things users do while listening. The hy-
pothesis of a context-aware recommendation is that it can significantly improve recom-
mendation accuracy because playlists can be tailored according to the actual condition
of the context [9]. Whereas considering contextual information in recommendations is
one challenge, getting the contextual information is another. In a mobile radio context,
mobile device sensors can determine if the user is moving, relaxing, or playing sports
and then incorporate this information into the recommendation playlist [10].
Challenge 6. Program management. Program managers have gained experience
over the years regarding which content to play at which time and in which order as well
as how to assemble and mix the content so that the program will be entertaining for the
user. When assembling radio playlists automatically, this program managing experi-
ence has to be reflected in recommender algorithms (sequential recommendation [8]).
Interdependencies between audio pieces have to be taken into account in general. There
are some well-known examples of misplaced advertisements (e.g., the headline “11-
year-old charged with driving drunk” followed by a beer advertisement). The ability to
mix content can therefore range from a faux pas to enjoyable radio.
Whereas all these challenges relate to algorithmic design, especially challenges 1, 2,
and 3, also have a major impact on interaction design. In the following section, we take
a closer look at interaction design challenges for radio recommender systems.
2 Interaction design in radio players
When designed to be interactive, radio necessarily takes the shape of an audio player,
typically providing play/pause, previous, and next interaction buttons. Whereas from a
user’s perspective, it is generally sufficient to have these three buttons, for a recom-
mender system, these are rather inadequate as it is hard to distinguish the intention be-
hind hitting a button (e.g., the next button). In a music context, to understand the pref-
erences of the user, the following questions arise: Why did the user click on the next
button? Does he/she generally dislike this kind of music, or does he/she just not want
to listen to it right now? This discrimination does not matter to the user in the moment
he/she hits the button, but it matters very much to the recommender system. A way to
find out the user’s preference could be to provide additional like and dislike buttons, so
when a user hits a button, it does not need to be interpreted (though there is a trend
toward implicit feedback [7]). Following this approach, in some popular music player
apps, there are like and dislike buttons in addition to previous and next buttons.
Going back to the radio context again, the situation gets much more complex. In a
radio context, when news, radio dramas, and other formats are mixed, the reasons be-
hind users’ hitting the next button are much more diverse and are important to under-
stand for a good recommendation: the user may not find the topic to be interesting,
he/she may have heard the content before, he/she may find the topic interesting but is
uncomfortable with the speaker’s voice, the user may find the topic interesting but may
not want to hear it right then, etc. If buttons were added for every possible cause (e.g.,
“I don’t like the speaker”), the user interface would inevitably become complex and
messy. As a result, we find that in a radio player context, the tasks of designing inter-
actions and designing recommendation algorithms have conflicting objectives.
The concept of traditional radio interactions (i.e., almost no interaction at all)
matches the design paradigms of mobile app interaction very well. Mobile apps have
made a simple interaction design imperative: “Not only are the apps easy, but so are the
processes,” (translated from German) said Jens Wehrmann, CEO of Mobile Software
AG [3]. Thus, a very reduced interaction design of a player fits well with the radio
domain.
In contrast, recommender systems follow a different approach: the more data (in
terms of amount as well as diversity), the better. Recommender systems cannot reach
their full potential with little data or with data that while available, still needs to be
interpreted. Especially in a radio player context, we find a mismatch between optimal
interaction design and optimal algorithm design. Hence, a tradeoff has to be made be-
tween optimal interaction design and optimal algorithmic design in this context.
3 Further research
The findings herein give rise to research on how users behave when having multiple
buttons to express their interests and preferences in a radio player context. Spotify has
made interesting contributions with their combination of a like and next button: namely,
the ratio of like and dislike does not always correlate with user engagement. In other
words, implicit and explicit feedback does not lead to the same results. According to
Spotify, “what users say they like is not always what they watch/listen” [4]. This puts
the usefulness of additional buttons in question and limits their value. Further research
has to show which buttons make sense to include in player design and which buttons
eventually produce data that enable recommender systems to give appropriate recom-
mendations.
4 References
1. Pariser, Eli (2001). The Filter Bubble: What The Internet Is Hiding From You, Penguin
Books.
2. Nguyen, T. T., Hui, P.-M., Harper, F. M., Terveen, L., and Konstan, J. A. (2014). Exploring
the filter bubble: the effect of using recommender systems on content diversity. In: Proceed-
ings of the 23rd international conference on World wide web, pp. 677-686.
3. Wehrmann, Jens (2013). Nicht Apps sind einfach, sondern die Prozesse. In: Mobile Tech-
nology Magazin 4/2013, pp. 10-11, Frankfurt.
4. Steck, H., van Zwol, R. and Johnson, C. (2015). Tutorial Interactive Recommender Systems.
ACM Recommender Systems, Vienna.
5. Tkalčič, M., Kunaver, M., Košir, A., Tasič, J.F. (2011). Addressing the New User Problem
with a Personality Based User Similarity Measure. In: Proceedings of the 2nd Workshop on
User Models for Motivational Systems.
6. Schein, A. I., Popescul. A., Lyle H. Ungar, L. H., and Pennock, D. M. (2002). Methods and
metrics for cold-start recommendations. In: Proceedings of the 25th ACM SIGIR con-
ference, pp. 253-260.
7. Moling, O, Baltrunas, L., and Ricci, F. (2013). Optimal Radio Channel Recommendations
with Explicit and Implicit Feedback, Epilog Conference, Vienna, Technical University of
Vienna, June 13, 2013, p. 8.
8. G. Shani, Heckerman, D., and Brafman, R. I. (2005). An mdp-based recommender system.
Journal of Machine Learning Research, 6:1265-1295.
9. Hidasi, B. (2015): Context-aware Preference Modeling with Factorization, ACM RecSys
2015, pp. 371-374.
10. Unger, M. (2015). Latent Context-Aware Recommender Systems, ACM RecSys 2015, pp.
383-386.