Music listening patterns can be influenced by contextual factors such as the activity a listener is involved in, the place one is located or physiological constants. As a consequence, musical listening choices might show some recurrent temporal patterns. Here we address the hypothesis that for some listeners, the selection of artists and genres could show a preference for certain moments of the day or for certain days of the week. With the help of circular statistics we analyze playcounts from Last.fm and detect the existence of that kind of patterns. Once temporal preference is modeled for each listener, we test the robustness of that using the listener's playcount from a posterior temporal period. We show that for certain users, artists and genres, temporal patterns of listening can be used to predict music listening selections with above-chance accuracy. This finding could be exploited in music recommendation and playlist generation in order to provide user-specific music suggestions at the “right” moment.
Among the requirements of good music recommenders we can point to, not only delivering the right music but, delivering it at the right moment. This amounts to consider the context of listening as a relevant variable in any user model for music recommendation. As existing technologies also make it possible to track the listening activity every time and everywhere it is happening, it seems pertinent to ask ourselves how this tracking can be converted into usable knowledge for our recommendation WOMRAD 2010 Workshop on Music Recommendation and Discovery, colocated with ACM RecSys 2010 (Barcelona, SPAIN).
Copyright ©. This is an open-access article distributed under the terms of
the Creative Commons Attribution License 3.0 Unported, which permits
unrestricted use, distribution, and reproduction in any medium, provided
the original author and source are credited.
systems. Music listening decisions might seem expressions of free
will but they are in fact influenced by interlinked social,
environmental, cognitive and biological factors [
While recommendations based on content analysis or on
collaborative filtering may achieve a certain degree of
personalization, they do miss the fact that the users interact with
the systems in a particular context [
Although research on behavioral rhythms has a long and solid
tradition, we are not aware of many studies about their influence
on music listening activities. The exception is a recent paper [
Getting access to yearly logs of the musical choices made by a large amount of listeners is not an easy task. Many music playing programs store individual users’ records of that, but they are not publicly accessible. As a workable solution, we have taken advantage of Last.fm API, which makes possible to get the playcounts and related metadata of their users. As raw data we have started with the full listening history of 992 unique users, expressed as 19,150,868 text lines and spanning variable length listening histories from 2005 to 2009. The data contained a user identifier, a timestamp, Musicbrainz identifiers for the artist and track, and a text name for the listened track.
The artist genre information was gathered from Last.fm using the Last.fm API method track.getTopTags(), which returns a list of tags and their corresponding weight1.This list of tags, however, may relate to different aspects of music (e.g. genre, mood, instrumentation, decades...). Since in our case we need a single genre per track, we first clean tags in order to remove special characters or any other undesirable characters, such as spaces, hyphens, underscores, etc. Then irrelevant tags (i.e., those having 1 Last.fm relevance weight of tag t to artist a, ranging from 0 to 100. a low weight) are removed and the remaining ones are matched against a predefined list of 272 unique musical genres/styles gathered from Wikipedia and Wordnet. From the genre tags we obtained for each song, we select the one with the highest weight. If there are several tags with the highest weight, we select the one with the least popularity (popularity is computed as the number of occurrences of a specific genre in our data-set).
Data coming from Lastfm.com contain playcounts that cannot be attributable to specific listening decisions on the side of users. If they select radio-stations based on other users, on tags or on similar artists there are chances that songs, artists and genres will not recur in a specific user’s profile. In general, even in the case of having data coming from personal players obeying solely to the user’s will, we should discard (i) users that do not provide enough data to be processed, and (ii) artists and genres that only appear occasionally. We prefer to sacrifice a big amount of raw data provided those we keep help to identify a few of clearly recurring patterns, even if it is only for a few users, artists or genres. In order to achieve the above-mentioned cleaning goals we first compute, for each user, the average frequency of each artist/genre in his/her playlist. Then, for each user’s dataset, we filter out all those artists/genres for which the playlist length is below the user’s overall average playlist length. Finally, in order to get rid of low-frequency playing users, we compute the median value of the number of artists/genres left after the last filtering step, which we will name as “valid” artists/genres. Those users whose number of “valid” artists/genres is below the median percentage value are discarded.
Once we get rid of all the suspected noise, we split our dataset in two groups. One will be used to generate the temporal predictions while the other one will be used to test them. The test set contains all the data in the last year of listening for a given subject. The prediction-generation set contains the data coming from two years of listening previous to the year used in the test set.
Circular statistics are aimed to analyze data on circles where
angles have a meaning, which is the case when dealing with daily
or weekly cycles. In fact, circular statistics is an alternative to
common methods or procedures for identifying cyclic variations
or patterns, which include spectral analysis of time-series data or
time-domain based strategies [
After this transformation, vectors ri are vector-averaged by ⎛cosα i ⎞ ri = ⎜⎜⎝ sinα i ⎠⎟
⎟ r = 1 ∑ ri
N i
The first step in circular analysis is converting raw data to a
common angular scale. We chose the angular scale in radians, and
thus we apply the following conversion to our dataset:
α =
2πx
k
where x represents raw data in the original scale, α is its angular
direction (in radians) and k is the total number of steps on the
scale where x is measured. In fact, we denote α as a vector of N
directional observations αi (i ranging from 1 to N). For the daily
hour case, x would have values between 0 and 24, and k = 24.
Alternatively, for the weekday analysis, x would have a scale from
0 (Monday) to 6 (Sunday) and thus, k = 6. As noted, the effect of
this conversion can be easily transformed back to the original
scale. Once we have converted our data to angular scale, we
compute the mean direction (a central tendency measure) by
transforming raw data into unit vectors in the two-dimensional
plane by
The quantity r is the mean resultant vector associated to the
mean direction, and its length R describes the spread of the data
around the circle. For events occurring uniformly in time R
values approach 0 (uniform circular distribution) whereas events
concentrated around the mean direction yield values close to 1
(see figure 1 for an example). A null hypothesis (e.g., uniformity)
about the distribution of data can be assessed using Rayleigh’s
[
As a consequence of the cleaning process, our working dataset now contains data from 466 valid users. The cleaning process has kept 62% of their total playcounts, which corresponds to 4,5% of the initial amount of artists. This dramatic reduction of the artists should not be surprising as many listening records show a “longtail” distribution, with just a few of frequently played artists, and many of them seldom played. On the other hand, when focusing on musical genre listening, the working dataset includes 515 users, from which 78% of their playcounts has been kept. These playcounts comprise 8.6% of the total number of genres. Again, a long-tail distribution of the amount of listened genres is observed.
Once we have cleaned our dataset, we compute the mean circular direction and the mean resultant vector length for each artist and user. Therefore, these values can be considered as a description of the listening tendencies for each artist by each user. Both parameters were calculated for the daily and for the weekly data.
In order to assess the relevance of these listening trends, we tested that the distribution of playcounts was different from uniform, and that it was modally distributed (i.e, showing a tendency around an hour or around a day of the week) and discarded those that were not fulfilling these requirements (a null hypothesis rejection probability p<0.05 was set for the tests).
In the hour prediction problem, for each listener’s clean dataset almost 93% (σ=13) of the artists passed on average the uniformity test (i.e., listening to them is meant to be concentrated around a specific hour). However, considering the raw dataset, only a peruser average of 7% (σ=3.2) of the artists show a listening hour tendency. For the weekly approach, the per-user average in the clean dataset is 99.8% (σ=0.8), indicating that there are some artists showing a clear tendency towards a preferred listening day. Considering the original raw dataset, they correspond to a 7.5% (σ= 3.2) of all the played artists.
Data from 466 users, including 7820 different songs and a grand total of 23669 playcounts were used in the validation of the temporal listening patterns of artists. For each user and artist we computed a “hit” if the absolute difference between the playing day in the prediction and test conditions, expressed as a circular mean value in radians, was less than 0.45 (the equivalent to a halfa-day error). For the time of the day a half-an-hour error was accepted, corresponding to a difference between the predicted and the observed time of less than 0.13 radians.
When predicting the day of listening, an overall 32.4% of hits was found for the songs in the test collection, which exceeds by far the chance expectations (1/7=14.28%). As the final goal of the model is providing user-specific contextual recommendation, an additional per-user analysis yielded 34.5% of hits (σ=17.8). Identical data treatment was done with the time of the day yielding an overall 17.1% of hits (chance expectation baseline: 1/24=4.1%) and a per-user hit rate of 20.5% (σ=16.4).
Data from 456 users, including more than 5100 songs and 117 genres, were used for the validation of the genre-related patterns. In order to consider a “hit” in the prediction of listening time and day for a given genre, we set the same thresholds than for evaluating the artist prediction. For the time of the day an overall 22.6% (and per-user 23.2%) of accurate predictions was found. It is interesting to note that relaxing the required accuracy of the prediction to plus/minus one hour error we reached 39.9% of average hits and per-user average 41% (σ=28.4). For the day of the week, the overall hit percent was 40.9%, while the per-genre average and the per-user average were, respectively, 40.7% (σ=24.1) and 41.7% (σ=26.3). It is interesting to note that among the best predictable genres we find many of infrequent ones but also many of the most frequent ones.
The present study is, as far as we know, the first one inquiring the possibility that our music listening behavior may follow some detectable circadian and ultradian patterns, at least under certain circumstances. We have discovered that a non-negligible amount of listeners tend to prefer to listen to certain artists and genres at specific moments of the day and/or at certain days of the week. We have also observed that, respectively for artists and for genres, 20% and 40% time-contextualized music recommendations can be successful. In our future work agenda, more sophisticated prediction models will be tested, and also ways to implement them into existing music recommenders.
Our thanks to Òscar Celma who kindly shared the Last.fm data file, accessible from this URL: http://www.dtic.upf.edu/~ocelma/MusicRecommend ationDataset/lastfm-1K.html