The consumption history of online media content such as music and video o ers a rich source of data from which to mine information. Trends in this data are of particular interest because they re ect user preferences as well as associated temporal contexts that can be exploited in systems such as recommendation or search. This paper classies songs associated with a holiday temporal context using a large, realworld dataset of user listening data. Results show strong performance of classi cation of Christmas music with Gaussian Mixture Models.
Consumption of media content such as music and video often exhibits patterns
when associated with a temporal context. Identifying and understanding these
contexts can improve the quality of recommendations as shown by [
Previous research has studied the dynamics and classi cation of time series
signals. In the web search domain, [
Approach
Listen counts of a track will exhibit a di ering and detectable pattern around a period of time if it has an association with that period, such as a Christmas track around December 25th. This pattern can be exploited by training a classi er using features of this signal. The features in this study are listening rates of a track i for day j in a window of time localized around the target context Rij =
PW l=1 PU k=1 cijk PU k=1 cilk (1) where c is an element of C, and C 2 RT W U where T is the number of tracks, W is the number of time periods, and U is the number of users. To control for the signi cant di erences in the overall popularity of tracks in a large data set, we normalize the listen counts of each track across the selected periods.
For classi cation, we chose the Gaussian Mixture Model (GMM) with full covariance matrix because it is fast to train and the listening rates resemble a normal distribution. A GMM is trained using tracks from the target holiday in a training portion of the data set, and classi cation is performed on the test set using the likelihood of the data given the model. 2.2
Dataset Number of Records Number of Users Number of Tracks Date Range 4,819,992,847 1,648,796 13,227,376
January 2012 - February 2013
This study uses an internal Gracenote dataset of online radio listening records in North America with some basic statistics of the dataset shown in Table 2.2. Each record of the dataset represents one listen of a track by one user and provides User ID, Date, Time, and Track ID. From the Track ID some associated metadata such as track name and album name is used for keyword search and post-experiment analysis. It is necessary to use a large dataset to get good classi cation results as shown in section 2.3. Other public datasets similar to Table 2.2 such as \Last.fm Dataset - 1K users" dataset available at http://www.dtic. upf.edu/~ocelma/MusicRecommendationDataset/lastfm-1K.html are too small. 2.3
Experiment - Christmas We chose Christmas as the target for seasonal music identi cation because of its popularity and large volume of associated music. We hypothesize that a classi er trained with features in section 2.1 can identify Christmas tracks. We generated an initial set of Christmas tracks by searching for \Christmas" keyword in the Large Scale Discovery of Seasonal Music track name and album name - totaling 87,554 Christmas tracks or 0.7% of the entire track population - and maintained a second list of tracks without the keyword. This is not a comprehensive list of Christmas tracks, but is generally free of non-Christmas tracks. Expert curation of a comprehensive set is infeasible with such a large dataset, and using tags from external sources or a more complex text search is error prone.
We chose a consecutive 15 day span centered on December 25th, Christmas, as the listening rate inputs to the classi er. Training and classi cation (60% train, 40% test) using Gaussian Mixture Models were performed on subsets of the dataset given by tracks with more than some minimum total listens in the whole dataset. To validate performance of the Christmas model, ROC and AUC score were calculated on the test set and are in Figure 1.
1.0 0.8 e taR0.6 e v ii t s o P eu0.4 r T 0.2
Christmas Model ROC 0.00.0 0.2 0.4 0.6
False Positive Rate The model performed quite well even though the experiment used an incomplete list of Christmas tracks. At the highest threshold, an inspection of tracks with >1500 Listens (AUC = 0.986) >500 Listens (AUC = 0.973) >200 Listens (AUC = 0.957) >100 Listens (AUC = 0.938) >10 Listens (AUC = 0.819) >1 Listens (AUC = 0.753) 0.8 high probability according to the Christmas model without the \Christmas" keyword shows that many are other Christmas songs well-known in North America such as \The First Noel" and \Santa Claus Is Coming To Town." This suggest that the model is not just identifying tracks with the \Christmas" keyword, but would likely accurately classify a more complete list of Christmas tracks.
One notable observation is the change in AUC as the threshold for total minimum listens of track is lowered. Classi cation su ers when including unpopular tracks. This is likely due to the natural variance in the listen counts of tracks with fewer listens. Normalizing smaller listen counts has a disproportionate e ect on computation of listen rates.
The dataset contains only a single year of data, which is a limitation for analyzing seasonal temporal contexts. Multiple years of data could provide better information for classi cation and show changing listening preferences over time. This is a topic of future work. 4
This study demonstrated on a large, real-world dataset that user listening data could be utilized to detect seasonal music content for Christmas. Classi cation with a Gaussian Mixture Model showed that the listen rates are sensitive to variance in unpopular tracks and quality results require detection to be performed on a large database of listening records.