Classification Of Artist Genre Through Supervised Learning

Classification of Artist Genre through SupervisedLearningRichard Ridley and Mitchell DumovicAbstract – The goal of this paper is to classify the genre ofan artist given a set of quantitative measures for each oftheir associated songs. We utilized supervised learning inthis task, and relied upon a dataset composed ofquantitative song data that Spotify provides for themajority of the songs that they stream. We collected datafor 14 different features for over 35 million songs, whichspanned over 100 thousand artists and over 12 hundredgenres. Before applying classification methods to thisdata, we collapsed our set of genres by a factor of ten,and collapsed our song data by applied differentstatistical measures to reduce to a set of features forevery artist. Then, we applied Stochastic GradientDescent, Gaussian SVM, and Nearest-Neighbors asclassifiers, and results were somewhat successful.INTRODUCTIONGenre classification is immensely important to thefield of music. Accurate genre classification canaid in the effectiveness of music recommendationengines, help unearth similarities betweendifferent music genres, and reduce the need for thehand labeling of genres in streaming services.This project focuses on attempting to accuratelypredict the genres that an artist belongs to giveninformation about the songs that they haveproduced. The input to our algorithm is a datasetthat we pulled from Spotify containinginformation about artists (with their associated listof classified genres) and tracks (with theirassociated song features). We then use variousclassification algorithms (KNN, SGD, SVM) tooutput predicted genre classifications for artists.RELATED WORKMost related works we found dealt with songclassification instead of artist classification, butwe still were able to get valuable ideas andinformation from various sources.For feature extraction, [2] and [4] used “SpectralCentroid, Spectral Roll-Off, Spectral Flux, TimeDomain Zero Crossings, Pitch Distribution and 13different Mel-Frequency Cepstral Coefficients ofindividual tracks” whereas [5] and [1] deal withextracted features related to timbre, melody,rhythm, and pitch. In general, most papers wefound pulled features from the raw songwaveform, so we knew that we would have toextract features derived from these waveforms aswell. However, we found that some of theseapproaches like [4], [5], and [6] suffered fromrelatively small datasets or limited numbers ofgenre labelings.In terms of models, the model that performed bestin [4] was k-nearest neighbors applied to themulti-label data. [5] suggests using support vectormachines as well as attempting one versus one andone versus all classification algorithms for themulti-label data. [6] suggests the use of anensemble technique for multi-label classification,another classification algorithm which involvestraining multiple classifiers for the data andcombining the results of the multiple classifiersinto one single classification. In general, it becamequickly apparent from these related works that wewould need to use one of the multi-label methodsfor classifying our data, and would really benefitfrom attempting algorithms involving the trainingof multiple classifiers, as these seemed to performbest over the similar papers that we looked at.DATASETTo collect our data, we relied upon a publicallyavailable Spotify API that allowed us to captureinformation about specific songs, artists, and theiralbums. To collect data about the songs of a widerange of different artists, we first needed toassemble a set of artists for which we couldcollected song data for. To do so, we first handpicked a “root” set of 50 different artists that webelieved to be representatives of a wide variety of

different genre clusters. After doing so, we used afunctionality that would allow us to find all of theartists related to a specific artist. Through the useof recursive formulation, we assembled a list ofover 100,000 artists that would form the basis forour training data. Each of these artists had labeledgenres, which indicated to which genres they werea part of. Then, for each of these artists, wecollected a set of all of their associated songs. Foreach of these songs, we were then able to collect avariety of different quantitative measures overdifferent qualitative aspects of our songs. Thefeatures that we collected are below.This succeeded in reducing the dimensionality ofour data set, and also reduced our training set toabout100,000examples,consistingofagglomerated song data for every artist that wehad collected. However, this did not come withouta cost, as the use of these summary statisticsreduced the precision of the data that we hadcollected by a significant degree. Nonetheless,artists of different genres generally differedsignificantly in their values for these statistics, asthe below example illustrates.AVAILABLE SONG criptionA confidence measure of how acousticthe track is.A metric assessing the degree to whichthe track is danceable.Represents a perceptual measure of thesongs intensity and activity.A confidence measure dictatingwhether the track contains vocals.An integer indicating the key the trackis in.A measure that details the likelihoodthat the track was performed live.The overall loudness of the track asmeasure in decibels.The confidence in the appearance ofspoken words in a track.The overall estimated tempo of a trackin beats per minute.FEATURESBecause songs do not have a genre labeling, weneeded to build a feature set for every artist thataccurately represented the set of songs associatedwith them. And since artists could have a highlyvariable number of songs, we needed to develop astrategy that would allow us to have a staticnumber of features defined over a variable numberof input songs. To do so, we calculated differentstatistical measures over each of our individualsong features for each of our artists. We decided touse mean, median, variance, and skew.GENRE COLLAPSINGIn our original dataset, Spotify classified songswith 1241 unique genres. As some of ouralgorithms involve training classifiers for eachgenre labeling, it was necessary to try to reducethe number of genre labelings as much as possible.To do this, we relied heavily on association rulealgorithms in order to find labelings that appearedfrequently together. These algorithms workedroughly as follows:First, define support for some set of genres G 𝑠𝑢𝑝𝑝(𝐺) - as the number of times that an artist isclassified with all the genres in the set G. Ouralgorithm started by removing genre labelings forsingle genre set below a certain support thresholdover the entire dataset. Next, define confidenceand lift for some pair of genres G1 and G2 asfollows:𝑠𝑢𝑝𝑝 𝐺! 𝐺!𝑐𝑜𝑛𝑓 𝐺! 𝐺! 𝑠𝑢𝑝𝑝 𝐺!𝑠𝑢𝑝𝑝 𝐺! 𝐺!𝑙𝑖𝑓𝑡 𝐺! 𝐺! 𝑠𝑢𝑝𝑝 𝐺! 𝑠𝑢𝑝𝑝(𝐺! )A confidence value of 1 indicates thatevery time that G1 appears it appears with G2,

meaning that it is a good candidate for genrecollapsing. For example, in our analysis we foundthat the genre “college a capella” appearedtogether with the genre “a capella” the exact samenumber of times the genre “college a capella”appeared by itself. This indicates a confidencevalue of 1, and resulted in us collapsing “college acapella” together with “a capella.”In general, our algorithm worked by iterativelycollapsing the two genres with the highestconfidence and lift values in our dataset,terminating when the confidence and lift valuesfell below a certain threshold. Finally, we scannedover the entire dataset once more and removed anymore collapsed genres that had a support valuelower than a second, higher threshold. Thisresulted in our final collapsed list of 95 uniquegenres. This also had the added effect of reducingthe number of labelings an artist was classified ason average, further simplifying our dataset. Beforethis genre collapsing, each artist was classifiedwith an average of approximately 5.27 differentgenre labelings. After the collapse, this numberwas reduced to an average of approximately 1.88different labelings.machines in this multi-label classificationproblem, we decided to try using both “one versusone” and “one versus all” approaches.A “one versus all” approach in the context ofmulti-label classification problem involves thetraining of a binary classifier for each of thelabels. This classifier discriminates between itsassociated label and all of the other labels in thetraining set. In our case, when using this approach,we trained 95 different classifiers which would tellus whether an artist either belonged to or did notbelong to a specific genre, and labeled that artistwith that genre if the classifier outputted that theartist does belong to that genre.We also employed a “one versus one”classification scheme when using SVM. “Oneversus one” in the context of multi-labelclassification involves the training of a classifierfor every pair of labels, and learns to distinguishwhich of the two labels an example is more likelyto belong to. After training these classifiers, anexample is assigned a label according to the classthat got the highest number of positive predictions.Thus, in our case, we trained a classifier for everypair of musical genres, and when evaluating whatgenre that an artist belonged to, we returned thegenre that was predicted most often in the 95 pairsin which it appeared. So as to ensure that artistscould attain multiple genre labelings, we assignedadditional genre labelings if the number of timesthat these individual genres were predicted wasclose to the number of times the mostly commonlypredicted genre was predicted.Both “one versus one” and “one versus all” can beconsidered valid, but it is generally the case that“one versus one” multi-label classification takessignificantly more processing time and requiresmore data than does “one versus all” approaches.CLASSIFICATION MODELSTo classify the genres for different artists, we usedthe supervised learning approaches of StochasticGradient Descent, K-nearest neighbors, andSupport Vector Machines. In configuringStochastic Gradient Descent and Support VectorMulti-label K-Nearest Neighbors ModelWhen attempting to classify a testing example x,let A be a sorted list containing the Euclideandistances of the training examples to the test point.Let Ai, y denote the label of the training example