The Importance of Song Context in Music Playlists
Enabling Recommendations in the Long Tail
Andreu Vall, Markus Schedl, Gerhard Widmer Massimo Quadrana, Paolo Cremonesi
Department of Computational Perception Politecnico di Milano, Italy
Johannes Kepler University Linz, Austria massimo.quadrana@polimi.it
andreu.vall@jku.at
ABSTRACT 1.1 Song Popularity
Music recommender systems often operate in sequential mode by This is a unigram model that computes the popularity of a song
suggesting a collection of songs that constitute a listening session. according to its relative frequency in the training playlists. At
This task is usually called automated music playlist generation and it test time, the next-song candidates are ranked by their popularity,
has been previously studied in the literature with different success- regardless of the current and previous songs in the playlist.
ful approaches based on, e.g., variations of collaborative filtering or
content-based similarity. Some of the proposed playlist models take 1.2 Song-based Collaborative Filtering
into consideration the current song and a number of previous songs, This is an item-based Collaborative Filtering (CF) model. A song s is
i.e., the song context, in order to predict the next song. However, it represented by the binary vector ps indicating the playlists to which
is not yet clear to what extent knowing this song context improves it belongs. The similarity of each pair of songs si , s j in the training
next-song predictions. To shed light on this question, we conduct set is computed as the cosine between psi and ps j . At test time, the
a numerical experiment on two datasets of hand-curated music next-song candidates are ranked according to their similarity to the
playlists, where we compare playlist models that account for dif- current song, but previous songs in the playlist are ignored.
ferent song context lengths. Our results indicate that knowing the
song context seems, at first, uninformative. However, we explain 1.3 Recurrent Neural Networks
this effect by a strong bias in the data towards very popular songs
Recurrent Neural Networks (RNNs) are a class of neural network
and observe that, in fact, songs in the long tail are more accurately
models particularly suited to processing sequential data. They have
predicted when the song context is considered.
a hidden state that accounts for the input at each time step while
recurrently incorporating information from previous hidden states.
CCS CONCEPTS We adopt the approach proposed in [2], where an RNN model
•Information systems →Data mining; Recommender systems; with one layer of gated recurrent units is combined with a loss
function designed to optimize the ranking of next-item recommen-
KEYWORDS dations. At test time, given the current and all the previous songs
music playlist generation, music recommender systems, recurrent in the playlist, the RNN outputs a vector of song scores that is used
neural networks to rank the next-song candidates.
ACM Reference format: 2 DATASETS
Andreu Vall, Markus Schedl, Gerhard Widmer and Massimo Quadrana,
The “AotM-2011” dataset [3] is a playlists collection derived from
Paolo Cremonesi. 2017. The Importance of Song Context in Music Playlists.
the Art of the Mix1 database. The “8tracks” dataset is a private
In RecSys 2017 Poster Proceedings, Como, Italy, August 27–31.
playlists collection compiled from 8tracks,2 an on-line platform
where users can share playlists and listen to other users’ playlists.
We keep only the playlists with at least 3 unique artists and
1 MODELING MUSIC PLAYLISTS with a maximum of 2 songs per artist. This is to discard artist- or
We describe the playlist models considered in our experiment. We album-themed playlists, which may correspond to a less careful
generally assume that: two sets of playlists are available, one for compilation process. We also keep only the playlists with at least
training and one for test; given the current song and a number 5 songs. Songs occurring in less than 10 playlists are removed to
of previous songs from a test playlist, a playlist model has to be ensure that the models have sufficient observations for each song.
able to rank all next-song candidates according to how likely they We randomly assign 80% of the playlists to the training set and
are to be the next song in that playlist. Note that we refer to the the remaining 20% to the test set. As in any recommendation task
current and the previous songs in the playlist as the song context blind to item content, the songs that occur only in test playlists
as it is commonly done in language models, but this should not be need to be removed because they can not be modeled at training
confused with the incorporation of general contextual information time. This affects the final playlist length and song frequency.
into recommendation systems. The filtered AotM-2011 dataset has 17,178 playlists with 7,032
songs by 2,208 artists. The filtered 8tracks dataset has 76,759 playlists
RecSys 2017 Poster Proceedings, Como, Italy 1 www.artofthemix.org
© 2017 Copyright held by the author(s). 2 https://8tracks.com
RecSys 2017 Poster Proceedings, August 27–31, Como, Italy A. Vall et al.
Table 1: Descriptive statistics for the AotM-2011 and the 8tracks All songs Popular songs Long tail songs
datasets. “Song popularity” corresponds to the song frequency in
the dataset, i.e., the number of playlists in which each song occurs. 4500
predicted rank
3000
dataset statistic min 1q med 3q max
AotM-2011 Songs per playlist 5 6 7 8 34 1500
Artists per playlist 3 5 7 8 34
Song popularity 1 8 12 20 249 1
N
F
N
F
N
F
.
.
.
p.
p.
p.
8tracks Songs per playlist 5 5 6 7 46
nd
nd
nd
C
C
C
N
N
N
Po
Po
Po
Ra
Ra
Ra
R
R
R
Artists per playlist 3 5 6 7 41 (a) AotM-2011 dataset
Song popularity 1 9 15 30 2,320
All songs Popular songs Long tail songs
11000
with 15,649 songs by 4,290 artists. Table 1 reports the distribution
predicted rank
of unique songs per playlist, unique artists per playlist and song 8000
popularity in the datasets.
5000
3 EVALUATION
1500
The evaluation is based on the prediction of next songs, considering 1
all the songs in the dataset as next-song candidates. Given a trained
N
F
N
F
N
F
.
.
d.
p.
p.
.
nd
nd
p
C
C
C
N
N
N
model, the following procedure is repeated over all the test playlists.
n
Po
Po
Po
Ra
Ra
Ra
R
R
R
We show the model the first song in a playlist. The model then has (b) 8tracks dataset
to rank all the next-song candidates according to their likelihood Figure 1: Distribution of predicted ranks for the actual next songs
to be the second song in that playlist. We only keep track of the (lower is better). The boxplots indicate the first quartile, median
rank assigned to the actual second song. We then show the model and third quartile ranks. Left: All songs are considered. Center:
the first and the second actual songs. The model has to rank all the Only the 10% most popular songs in the dataset are considered.
next-song candidates for the third position, having now a longer Right: Only the 90% least popular (long tail) songs in the dataset are
song context. In this way, we progress until the end of the playlist, considered. “Pop.” “CF” and “Rand.” correspond to the popularity-
always keeping track of the rank assigned to the actual next song. based, the song-based CF and the random models, respectively. The
Finally, the performance of a playlist model is characterized by the scale of the y-axis relates to the number of songs in each dataset.
distribution of its predicted ranks for the actual next songs.
Hyperparameter tuning is performed on a validation split. Along In our view, this is a remarkable finding given that the bias
with the described models, we also consider a random model that as- towards popular songs is a characteristic feature of playlist datasets
signs scores to songs uniformly at random, yielding random ranks. in the music recommendation domain [1]. Playlist models that
account for a longer song context are therefore better suited to
4 RESULTS next-song prediction for songs in the long tail.
Considering the ranks achieved for all next-song predictions (left
panels in Figure 1), the RNN and the popularity-based model per- 5 CONCLUSION
form comparably well, and significantly better than the song-based In this work we investigated the importance of the song context for
CF model. This is a surprising result given the simplicity and null next-song recommendations in music playlists. Our results indicate
song context of the popularity-based model, but might be explained that the bias towards very popular songs masks the importance
by the impact of a small number of very popular songs in both of the song context. However, we observe that a playlist model
datasets (Table 1). To investigate this effect, we consider the ranks based on an RNN, which considers the full song context of a playlist,
achieved when the actual next songs belong to the 10% most pop- clearly outperforms simpler models when recommending songs
ular songs in the dataset (central panels in Figure 1), and when belonging to the long tail of non popular songs.
the actual next songs belong to the long tail of 90% least popular
songs in the dataset (right panels in Figure 1). As expected, the 6 ACKNOWLEDGMENTS
popularity-based model performs well on the most popular songs,
We thank Matthias Dorfer, Bruce Ferwerda, Rainer Kelz, Filip Ko-
but is close to the random model for songs in the long tail. Thus,
rzeniowski, Rocı́o del Rı́o and David Sears for helpful discussions.
the overall performance of the popularity-based model is similar to
that of the RNN model only because the datasets are biased towards REFERENCES
popular songs. The song-based CF model is not competitive in this [1] Òscar Celma. 2010. Music recommendation and discovery. Springer.
experiment and, especially in the 8tracks dataset, is also affected [2] Balázs Hidasi et al. 2016. Session-based recommendations with recurrent neural
by the bias towards popular songs. In contrast, the performance networks. In Proc. ICLR.
[3] Brian McFee and Gert Lanckriet. 2012. Hypergraph models of playlist dialects.
of the RNN model, which keeps track of the full song context, is In Proc. ISMIR. 343–348.
robust to the popularity of the actual next song.