Audio-Based Sequential Music Recommendation
Rodrigo Borges
Computer Science Department
University of S
˜
ao Paulo & Tampere University
S
˜
ao Paulo, Brazil
Marcelo Queiroz
Computer Science Department
University of S
˜
ao Paulo
S
˜
ao Paulo, Brazil
Abstract—We propose an audio-based recommendation model
designed to predict the upcoming track within a listening session,
given the audio associated with the current track. Instead of
relying on users’ feedback, as most recommenders, the proposed
model aims to learn intrinsic audio elements that can be leveraged
in the context of sequential recommendation. The proposed model
is evaluated using Mel-spectrogram and raw audio as input data
and, in its best configuration, was able to predict almost 65%
unseen transitions used in the evaluation phase, and 3.5% cold-
start transitions, i.e. transitions from tracks that were never seen
by the model.
Index Terms—Audio-Based music recommendation, Gated Re-
current Unit, Audio Content
I. INTRODUCTION
Sequential music recommendation methods are usually de-
signed to suggest the track that will be listened to in the near
future (next-track), given the information about the tracks that
were listened to in the past [1], [2]. These methods are usually
trained with datasets containing information about track/track
transitions, and the main assumption is that future transitions
can be predicted (suggested) based on historical data [3]–[6].
Music recommender methods may also use the audio signals
associated with each track for calculating recommendations to
users, in which case they are called audio-based methods (as
opposed to rating-based methods). Suggesting tracks based on
audio content is both technically and conceptually challenging,
because it is based on different music listening hypotheses,
i.e. that intrinsic elements, such as rhythm and timbre, are
also important for music recommendation alongside extrinsic
elements, such as listening habits and patterns.
Audio signals associated with tracks may be used by rec-
ommendation methods in two ways: audio may serve as the
primary resource for providing recommendations, or it may be
an auxiliary source of information within rating-based strate-
gies. In the former case, tracks are usually suggested based on
similarities defined within the audio domain, which potentially
produces novel and fairly unbiased results, but imposes severe
limitations on the quality of recommendations [7]. In the latter
case, information extracted from the audio signal is incorpo-
rated in methods that were originally designed to operate with
user/item interaction data, for alleviating situations of item
cold-start, i.e. suggesting items that were never seen by the
model [8], [9].
In this work, we propose Audio-Based GRU4REC,
AGRU4REC for short, a method designed to suggest the next
track within a listening session given an audio representation
(also referred to as audio feature
1
) associated with the current
track. The proposed method is inspired in GRU4REC [10],
which was originally proposed for recommending the next
item to a user, given the information of the previously con-
sumed item. AGRU4REC has no access to any metadata
that identifies the current track, and suggests tracks based
exclusively on the audio contents.
II. RELATED WORK
Music recommenders that rely solely on audio information
have considered the recommendation task as a task of selecting
tracks from an audio-based representation space according to
a set of tracks that a user has listened to before or is currently
listening to [11]. This representation space is built in such
a way that similar tracks are supposed to be located close
to each other, which makes similarity measuring strategies
an essential choice in this context [12]–[15]. There is no
consensus however on what it means for two audios to be
similar to each other, a broad question which is beyond the
scope of this work.
Audio files can also be exploited by rating-based methods
for mitigating the cold-start limitation. In [16], users are
clustered according to listening habits, and track audios are
clustered into music genres. Preferences may be modelled for
each user cluster and musical genre, thus mitigating the lack
of interaction information for new tracks. The idea that the
similarity between tracks can be defined through user access
patterns, and that this similarity can be estimated from the
audio domain, was explored in [17]. A similar idea, based on
learning-to-rank, was proposed in [18]. When given a query
track, the ranking system retrieves other tracks sorted by rel-
evance according to user access patterns, and a corresponding
ranking is simultaneously learned using the query audio as
input. After training, the ranking system is supposed to retrieve
relevant tracks when queried with the audio of a new track,
i.e., as a query-by-example system.
A novel approach is proposed for a dynamic content-based
music recommender in [19], [20]. Ratings given by users are
modelled as a combination of two factors, an affinity for
1
This method presupposes a choice of a specific audio feature representation
for the tracks, which is simply referred to in the sequel as “audio feature”; in
the experimental part of this paper, we use Mel spectrograms and raw audio
waveforms.
421ISBN: 978-9-4645-9360-0 EUSIPCO 2023
the audio content, and a factor responsible for diversity. The
affinity for audio features is modelled as an inner product of
a user preference variable and the audio features of listened
tracks. The diversity is implemented with an exponential
curve that prevents the recommender to repeat a song that
was recently suggested. The system, however, iterates through
every track for maximizing the quantile value of the estimated
distribution, inspired by Bayesian-UCB [21], and this can be
time-consuming.
To the best of our knowledge, one single method was
already proposed in the specific context of audio-based se-
quential music recommendation [22]. The method, named
Adaptive Linear Mapping Model (ALMM), adapts the content-
boost methodology [8] to the next-track recommendation task.
ALMM decomposes a set of personalized transition matri-
ces as a product of three latent matrices: user embedding,
previous-track embedding, and next-track embedding, in a
similar fashion to FPMC [23]. The two last matrices, the
ones associated with the previous and next tracks, are also
factorized as linear products of an audio feature matrix and
auxiliary matrices that are learned during the optimization
process. The final recommendation score for a specific track
can be calculated directly from its audio features with the help
of the auxiliary matrices.
III. METHOD
Audio-Based GRU4REC is an audio-based recommenda-
tion model composed of one Convolutional Neural Network
(CNN), one Gated Recurrent Unit (GRU), and one Multi-Layer
Perceptron (MLP). The model is trained to predict the next
track within a listening session, given approximately 3 seconds
of the audio associated with the current track. More details
about the model are presented in the sequel.
A. Problem Definition
A listening session of size T is denoted as
{s
(1)
, s
(2)
, . . . , s
(T )
}, where s
(t)
∈ S is the track observed at
instant t, with 0 < t ≤ T . Typically, a temporal dependency
among consecutive tracks is assumed according to the
conditional probabilities p(s
(t)
|s
(t−1)
, . . . , s
(t−m)
), taking
the previous m tracks into consideration. Here, we assume a
dependency between the current track and the previous audio
features, expressed as p(s
(t)
|A
(t−1)
, . . . , A
(t−m)
), where A
(t)
is the audio feature (a Mel spectrogram or any other selected
representation) associated with track s
(t)
observed at instant
t.
Our aim is to train a model that is able to predict the
upcoming track s
(t+1)
given the audio feature associated with
the current track A
(t)
. In other words, a model that estimates
p(s
(t+1)
|A
(t)
).
B. Audio-Based GRU4REC
Audio-Based GRU4REC (AGRU4REC) was inspired in the
method GRU4REC [10], originally proposed as a track/track
transition model. AGRU4REC suggests the next track within
a listening session given an audio feature.
The model consists of three stages described as follows.
First, a function f(·) maps an audio feature A
(t)
to an audio
embedding D
(t)
, in such a way that f (A
(t)
) = D
(t)
. Second,
another embedding is calculated by a function g(·) with
memory, i.e. a function that is able to store its parameters
so they can be used in the next round of recommendation.
Let g(·) be the function that maps the audio embedding to the
new embedding, named sequence-aware embedding E
(t)
, and
let H
(t)
be the current state of the function g(·). At instant t, a
sequence-aware embedding is calculated considering the state
stored at instant t − 1, in such a way that g(D
(t)
, H
(t−1)
) =
E
(t)
. When a listening session ends, the state H is reset,
assuming that listening sessions are independent of each other.
Finally, a function q(·) maps the session-aware embedding to
the scores corresponding to the next track in session Y
(t+1)
, in
such a way that q(E
(t)
) = Y
(t+1)
. The output Y
(t+1)
has size
|S|, and contains the scores attributed to each track s ∈ S. The
highest the score attributed to a track, the higher the probability
that this track is the next one in a current listening session.
Function f (·) is implemented with a CNN, function g(·)
is implemented with a GRU network, and function q(·) is
implemented with an MLP. The hidden state of the GRU
network H is initialized containing zeros, and the training
process is summarized in the sequel.
The audio embedding D
(t)
is first obtained from its corre-
sponding audio feature A
(t)
(Figure 1, left) and it propagates
to the GRU network. The reset (R
(t)
) and update (Z
(t)
) gates
of the GRU network are the first parameters to be adjusted,
respectively, with equations:
R
(t)
= σ(W
rs
D
(t)
+ W
rh
H
(t−1)
+ B
r
) (1)
Z
(t)
= σ(W
zs
D
(t)
+ W
zh
H
(t−1)
+ B
z
) (2)
where W
xy
are weight matrices for mapping x to y, to be
adjusted during training, and B
r
and B
z
are biases. Sigmoid
is applied to transform the input values to the range (0,1).
When presenting the audio embedding corresponding to the
first track of each listening session, H
(t−1)
is set equal to zero
for ensuring independency between sessions, and the second
terms of both equations are not considered in the calculation
of R
(t)
and Z
(t)
.
A candidate hidden state N
(t)
is calculated, incorporating
the reset gate:
N
(t)
= tanh(W
ns
D
(t)
+ W
nh
(R
(t)
⊙ H
(t−1)
) + B
n
)) (3)
where ⊙ is the Hadamard (elementwise) product and tanh
is applied to ensure that the values remain in the interval (-
1,1). For now, when entries in the reset gate are set to 1, then
the candidate’s new state reminds the hidden state calculated
for standard RNN. When the reset gate is set equal to 0 the
architecture resembles a standard MLP having D
(t)
in the
input.
The final hidden state incorporates the update gate, and is
calculated with:
H
(t)
= (1 − Z
(t)
) ⊙ N
(t)
+ Z
(t)
⊙ H
(t−1)
, (4)
422
GRU
Y
(t+1)
A
(t)
CNN
D
(t)
E
(t)
H
(t-1)
MLP
Fig. 1. Audio-Based GRU4REC (AGRU4REC), a spatio-temporal recommendation model inspired in GRU4REC [10].
where H
(t−1)
is the hidden state at time t − 1. The update
gate Z
(t)
determines to which extent the new hidden state H
(t)
is inherited from the previous hidden state H
(t−1)
, and how
much of the new candidate state is considered.
In this model, the session-aware embedding E
(t)
is a copy
of H
(t)
, and Y
(t+1)
is obtained from E
(t)
, considering that
Y
(t+1)
= tanh(q(E
(t)
)) (Figure 1, right).
The model is trained in mini-batches, and the goal is to
minimize the TOP1 loss function, calculated as [10]:
Loss =
1
|S|
|S|
X
j=1
σ(ˆy
j
− ˆy
i
) + σ(ˆy
2
j
), (5)
where ˆy
i
is the score given to the right track s
(t+1)
, and ˆy
j
is the score given to any other track observed within a mini-
batch (negative samples). An extra regularization term forces
negative samples to have scores close to zero.
IV. EXPERIMENTS
A. Dataset
LFM-1b is among the biggest datasets publicly available
containing music consumption information [24]. It contains
data extracted from the LastFM
2
streaming platform from
2005 to 2014 in the format (user, artist, album, track, times-
tamp), where each row is associated with a listening event.
We separated the user-track interactions from the year 2013,
taking into account that this was the most recent year available
with a relevant number of interactions.
In order to separate dataset entries in listening sessions,
we separated the tracks listened by the same user, ordered
these events by timestamp, and sessions are assumed as non-
interrupted sequences of listening events. More specifically, a
session is assumed as starting with the first track of the list, and
whenever an interval between adjacent tracks is longer than
30 minutes, the current session is finished and the following
track is assumed to belong to a new session.
Audio files corresponding to tracks that were listened to by
at least 10 users were downloaded from the Spotify website
with the help of their API
3
and of the Spotipy
4
Python
library. The URL of a 30s mp3 preview for each song is
included and was used to download the corresponding files.
2
https://www.last.fm/
3
https://developer.spotify.com/documentation/web-api/
4
https://github.com/plamere/spotipy
In total, mp3 previews for 237,705 tracks were downloaded.
All the information downloaded from the Spotify website was
exclusively applied for research purposes.
B. Data Partition and Feature Extraction
Around 19,000,000 non-interrupted listening sessions were
derived from user-track interaction data, considering intervals
shorter than 30 minutes between listening events as a criterion
for including tracks in the same session. Among these sessions,
889,968 included only tracks associated with downloaded au-
dio previews and were considered in the experiments. Listen-
ing sessions containing less than 5 and more than 100 events
were removed, as well as sessions with less than 2 unique
tracks (i.e. sessions containing a single song multiple times).
Finally, we split the whole set into training/validation/test
subsets according to proportions of 80/10/10% and ordered
by timestamp (training on the oldest 80% sessions, validating
on the next 10% and testing on the newest 10%). The idea is to
simulate a situation when the system is exposed to interactions
that happened in the past and evaluate its performance with
listening sessions that happen in the future.
The raw audios were extracted from the mp3 previews with
the Librosa Python library
5
, with a sampling rate of 22,050 Hz.
Mel-spectrograms were also computed from these excerpts
with the same library, using 128 Mel filters and FFT window
and hop sizes of 2048 and 512 samples, respectively, and the
Hann window function, resulting in Mel-spectrograms of di-
mension 128 × 1292. The magnitudes of the Mel-spectrograms
were compressed by a nonlinear curve log(1 + C|A|) where
|A| is the magnitude and C is set to 10, as suggested in [25].
C. CNN Architectures and Training
In order to compare the performances of AGRU4REC using
raw waveforms and Mel-spectrograms as input, as suggested
in [26], we implemented the method using one-dimensional
(1D) and two-dimensional (2D) CNNs. The 1D CNN was
inherited from [25], and the 2D CNN was implemented in
such a way that its architecture (number of layers, activation
functions, dropout rate, normalization layers) was kept as
similar as possible to the 1D architecture.
The models were trained with audio features corresponding
to approximately 3 seconds of the audio previews: 59,049
samples in the case of 1D CNN, and 115 FFT frames in the
5
https://librosa.org
423
TABLE I
RESULTS MEASURED FOR THE NEXT-TRACK PREDICTION TASK. RESULTS ARE REPORTED SEPARATELY FOR ALL TRANSITIONS IN THE TEST SUBSET
(OVERALL), AND FOR TRANSITIONS FROM TRACKS THAT APPEAR IN THE TEST SLICE FOR THE FIRST TIME (COLD-START). WHENEVER THE RESULTS
MEASURED FOR ONE METHOD ARE BETTER THAN THE OTHERS, THE VALUES ARE HIGHLIGHTED.
REC@1 REC@20 REC@100 MRR@1 MRR@20 MRR@100
Overall
ALMM 0.018 0.222 0.421 0.018 0.053 0.058
AGRU4REC (MEL) 0.222 0.497 0.640 0.222 0.292 0.296
AGRU4REC (RAW) 0.245 0.512 0.651 0.245 0.314 0.317
Warm-Start
ALMM 0.019 0.234 0.445 0.019 0.056 0.061
AGRU4REC (MEL) 0.237 0.529 0.681 0.237 0.311 0.315
AGRU4REC (RAW) 0.262 0.547 0.694 0.262 0.335 0.339
Cold-Start
ALMM 0.001 0.005 0.021 0.001 0.002 0.002
AGRU4REC (MEL) 0.001 0.008 0.027 0.001 0.002 0.002
AGRU4REC (RAW) 0.002 0.013 0.036 0.002 0.003 0.004
case of 2D CNN. The architecture selected for mapping Mel-
spectrograms to audio embeddings has 5 convolutional layers,
followed by a linear layer, applied for reducing the embedding
size. The architecture selected for mapping raw waveforms to
audio embeddings has 11 convolutional layers, followed by a
linear layer, also applied for reducing the embedding size. The
source code for reproducing the experiments is publicly avail-
able
6
. In order to improve the model’s generalization ability,
30-second audio features were separated into 10 equally-sized
slices, and at each training round a random slice is chosen to
train the model.
D. Previous Approaches
The original ALMM method suggests the use of personal-
ized transition matrices, but preliminary results showed that
using a single transition matrix produced better results, and
so results reported here use the latter strategy. The ALMM
method was trained and evaluated on the LFM-1b dataset using
the same audio codeword histograms defined in [9], in order
to preserve the structure of the original method.
E. Evaluation Metrics
The AGRU4REC and ALMM methods were trained for 35
epochs, and were evaluated according to their performances in
the test subset. The Recall (REC@K) was used for measuring
the recommendation accuracy, and Mean Reciprocal Rank
(MRR@K) was used for measuring the quality of the ranking
in the results. Both metrics were implemented according
to [27].
In the case of AGRU4REC, the input audio is sliced in
10 equally-sized slices, as mentioned before, and K tracks
are recommended for each slice. The results are measured
considering the recommendations calculated for all slices.
V. RESULTS
All methods were evaluated in transitions between tracks
that both appeared in the training and test set, a context
referred to as warm-start (this is the most common recom-
mendation scenario). In this context, which may be considered
the easiest scenario, AGRU4REC produced better results than
6
https://www.github.com/rcaborges/AGRU4REC
ALMM and the best results for each considered metric.
In Table I it can be seen that improvements obtained by
AGRU4REC relative to ALMM range from factors of 1.6×
(REC@100) up to 11× (REC@1 and MRR@1), where best
improvements do occur for the more demanding metrics which
only consider the first position of the corresponding ranked
lists.
The training/validation/test splits refer to sessions (and not
tracks), but it is important to differentiate between tracks
that did appear in training sessions and those that did not.
Among all 237,705 tracks used in the experiments, 5,459 were
observed in the test subset for the first time. Transitions from
these tracks to tracks appearing in the training subset were
considered cold-start transitions and were evaluated separately
(Table I). When assessing these transitions, AGRU4REC pre-
sented better results compared to ALMM, both in terms of
Recall and MRR, for all values of K, with improvements
ranging from 50% (REC@100) to 160% (REC@20). The low
values are indicative of the difficulty of sequential cold-start
prediction, but nevertheless, allow a comparison between these
methods.
The 1D CNN trained with raw waveforms turned out
to be more versatile than the 2D CNN trained with Mel-
spectrograms, achieving better results than the latter in every
metric and every scenario considered. This difference might
be attributed both to the audio feature used, respectively raw
waveforms or Mel-spectrogram, and to the CNN architecture,
which is 1D or 2D, respectively.
One example of an audio-based recommendation instance is
presented in Table II. AGRU4REC was able to recommended
non-obvious track/track transitions, from different artists, and
with consistent results.
VI. CONCLUSIONS
The proposed audio-based recommendation models
achieved satisfactory accuracy and ranking quality. These
models can be also used as auxiliary recommendation
models, to be consulted whenever the current track that a
user might be listening to is not known by a feedback-
based recommendation model. According to the results
presented here, AGRU4REC can improve the accuracy of any
current-track cold-start recommendation model up to 3.6%.
424
TABLE II
ONE EXAMPLE OF AUDIO-BASED RECOMMENDATION. THE TABLE SHOWS
ONE TRACK/TRACK TRANSITION FROM THE TEST SUBSET, AND THE
TRACKS RECOMMENDED BY AGRU4REC SORTED BY RELEVANCE.
Track Artist
Previous Track Blunderbuss Jack White
Next Track Speak to Me/Breathe Pink Floyd
Rec. Tracks
Atom Heart Mother Pink Floyd
Cirrus Minor Pink Floyd
Astronomy Domine Pink Floyd
Comfortably Numb Pink Floyd
If Pink Floyd
Speak to Me/Breathe Pink Floyd
Shine on You Crazy Diamond Pink Floyd
One of These Days Pink Floyd
Let There Be More Light Pink Floyd
Wish You Were Here Pink Floyd
One potential application for audio-based models is the
generation of playlists given a local collection of tracks stored
on a user’s device. In this specific case, AGRU4REC could be
applied for generating recommendations based on the stored
audio files, even without having access to metadata associated
with these tracks. This can be particularly interesting in a sit-
uation where users are trying to expand their music collection
with tracks that are related to the ones they already have.
VII. ACKNOWLEDGEMENTS
During the development of this project, the first
author received financial support from CAPES Grant
88881.189985/2018-01 and the second author received finan-
cial support from CNPq Grant 310141/2022-2.
REFERENCES
[1] Brian McFee and Gert R. G. Lanckriet, “The natural language of
playlists,” in Proceedings of the 12th International Society for Music
Information Retrieval Conference, ISMIR 2011, Miami, Florida, USA,
October 24-28, 2011, 2011, pp. 537–542.
[2] Bruno L. Pereira, Alberto Ueda, Gustavo Penha, Rodrygo L. T. Santos,
and Nivio Ziviani, “Online learning to rank for sequential music
recommendation,” in Proceedings of the 13th ACM Conference on
Recommender Systems, 2019, RecSys ’19, p. 237–245.
[3] Mehdi Hosseinzadeh Aghdam, Negar Hariri, Bamshad Mobasher, and
Robin Burke, “Adapting recommendations to contextual changes using
hierarchical hidden markov models,” in Proceedings of the 9th ACM
Conference on Recommender Systems, 2015, RecSys ’15, p. 241–244.
[4] Malte Ludewig and Dietmar Jannach, “Evaluation of session-based
recommendation algorithms,” User Modeling and User-Adapted Inter-
action, vol. 28, no. 4-5, pp. 331–390, 2018.
[5] Massimo Quadrana, Paolo Cremonesi, and Dietmar Jannach, “Sequence-
aware recommender systems,” ACM Comput. Surv., vol. 51, no. 4, 2018.
[6] Chengfeng Xu, Pengpeng Zhao, Yanchi Liu, Jiajie Xu, Victor S.Sheng
S.Sheng, Zhiming Cui, Xiaofang Zhou, and Hui Xiong, “Recurrent
convolutional neural network for sequential recommendation,” in The
World Wide Web Conference, 2019, WWW ’19, p. 3398–3404.
[7] Arthur Flexer, Martin Gasser, and Dominik Schnitzer, “Limitations of
interactive music recommendation based on audio content,” in AM ’10,
The 5th Audio Mostly Conference, Pite
˚
a, Sweden, September 15-17,
2010. 2010, p. 13, ACM.
[8] Peter Forbes and Mu Zhu, “Content-boosted matrix factorization for
recommender systems: Experiments with recipe recommendation,” in
Proceedings of the Fifth ACM Conference on Recommender Systems,
2011, RecSys ’11, p. 261–264.
[9] A
¨
aron van den Oord, Sander Dieleman, and Benjamin Schrauwen,
“Deep content-based music recommendation,” in Proceedings of the
26th International Conference on Neural Information Processing Sys-
tems - Volume 2, 2013, NIPS’13, p. 2643–2651.
[10] Bal
´
azs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos
Tikk, “Session-based recommendations with recurrent neural networks,”
in 4th International Conference on Learning Representations, ICLR,
2016.
[11] Pedro Cano, Markus Koppenberger, and Nicolas Wack, “Content-based
music audio recommendation,” in Proceedings of the 13th Annual ACM
International Conference on Multimedia, 2005, MULTIMEDIA ’05, p.
211–212.
[12] Malcolm Slaney, Kilian Q. Weinberger, and William White, “Learning
a metric for music similarity,” in ISMIR 2008, 9th International Confer-
ence on Music Information Retrieval, Drexel University, Philadelphia,
PA, USA, September 14-18, 2008, 2008, pp. 313–318.
[13] B. Logan and A. Salomon, “A music similarity function based on signal
analysis,” in IEEE International Conference on Multimedia and Expo,
2001. ICME 2001., 2001, pp. 745–748.
[14] Dmitry Bogdanov, Joan Serr
`
a, Nicolas Wack, Perfecto Herrera, and
Xavier Serra, “Unifying low-level and high-level music similarity
measures,” IEEE Transactions on Multimedia, vol. 13, no. 4, pp. 687–
701, 2011.
[15] Matthew D. Hoffman, David M. Blei, and Perry R. Cook, “Content-
based musical similarity computation using the hierarchical dirichlet
process,” in ISMIR 2008, 9th International Conference on Music Infor-
mation Retrieval, Drexel University, Philadelphia, PA, USA, September
14-18, 2008, 2008, pp. 349–354.
[16] Qing Li, Byeong Man Kim, Dong Hai Guan, and Duk whan Oh, “A
music recommender based on audio features,” in Proceedings of the
27th Annual International ACM SIGIR Conference on Research and
Development in Information Retrieval, 2004, SIGIR ’04, p. 532–533.
[17] B. Shao, D. Wang, T. Li, and M. Ogihara, “Music recommendation
based on acoustic features and user access patterns,” IEEE Transactions
on Audio, Speech, and Language Processing, vol. 17, no. 8, pp. 1602–
1611, 2009.
[18] B. McFee, L. Barrington, and G. Lanckriet, “Learning content similarity
for music recommendation,” IEEE Transactions on Audio, Speech, and
Language Processing, vol. 20, no. 8, pp. 2207–2218, 2012.
[19] Zhe Xing, Xinxi Wang, and Ye Wang, “Enhancing collaborative filtering
music recommendation by balancing exploration and exploitation,” in
Proceedings of the 15th International Society for Music Information
Retrieval Conference, ISMIR 2014, Taipei, Taiwan, October 27-31, 2014,
2014, pp. 445–450.
[20] Xinxi Wang, Yi Wang, David Hsu, and Ye Wang, “Exploration
in interactive personalized music recommendation: A reinforcement
learning approach,” ACM Trans. Multimedia Comput. Commun. Appl.,
vol. 11, no. 1, 2014.
[21] Emilie Kaufmann, Olivier Cappe, and Aurelien Garivier, “On bayesian
upper confidence bounds for bandit problems,” in Proceedings of
the Fifteenth International Conference on Artificial Intelligence and
Statistics, 2012, vol. 22 of Proceedings of Machine Learning Research,
pp. 592–600.
[22] Szu-Yu Chou, Yi-Hsuan Yang, Jyh-Shing Roger Jang, and Yu-Ching Lin,
“Addressing cold start for next-song recommendation,” in Proceedings
of the 10th ACM Conference on Recommender Systems, 2016, RecSys
’16, p. 115–118.
[23] Steffen Rendle, Christoph Freudenthaler, and Lars Schmidt-Thieme,
“Factorizing personalized markov chains for next-basket recommenda-
tion,” in Proceedings of the 19th International Conference on World
Wide Web, 2010, WWW ’10, pp. 811–820.
[24] Markus Schedl, “The lfm-1b dataset for music retrieval and recommen-
dation,” in Proceedings of the 2016 ACM on International Conference
on Multimedia Retrieval, 2016, ICMR ’16, pp. 103–110.
[25] Jongpil Lee, Jiyoung Park, Keunhyoung Luke Kim, and Juhan Nam,
“Sample-level deep convolutional neural networks for music auto-
tagging using raw waveforms,” CoRR, vol. abs/1703.01789, 2017.
[26] Sander Dieleman and Benjamin Schrauwen, “End-to-end learning for
music audio,” in 2014 IEEE International Conference on Acoustics,
Speech and Signal Processing (ICASSP), 2014, pp. 6964–6968.
[27] Arthur Tofani, Rodrigo Borges, and Marcelo Queiroz, “Dynamic
session-based music recommendation using information retrieval tech-
niques,” User Model. User Adapt. Interact., vol. 32, no. 4, pp. 575–609,
2022.
425
