{"title": "Automatic Generation of Social Tags for Music Recommendation", "book": "Advances in Neural Information Processing Systems", "page_first": 385, "page_last": 392, "abstract": "Social tags are user-generated keywords associated with some resource on the Web. In the case of music, social tags have become an important component of Web2.0\" recommender systems, allowing users to generate playlists based on use-dependent terms such as \"chill\" or \"jogging\" that have been applied to particular songs. In this paper, we propose a method for predicting these social tags directly from MP3 files. Using a set of boosted classifiers, we map audio features onto social tags collected from the Web. The resulting automatic tags (or \"autotags\") furnish information about music that is otherwise untagged or poorly tagged, allowing for insertion of previously unheard music into a social recommender. This avoids the ''cold-start problem'' common in such systems. Autotags can also be used to smooth the tag space from which similarities and recommendations are made by providing a set of comparable baseline tags for all tracks in a recommender system.\"", "full_text": "Automatic Generation of Social Tags for Music\n\nRecommendation\n\nDouglas Eck\u2217\n\nSun Labs, Sun Microsystems\n\nBurlington, Mass, USA\n\ndouglas.eck@umontreal.ca\n\nPaul Lamere\n\nSun Labs, Sun Microsystems\n\nBurlington, Mass, USA\n\npaul.lamere@sun.com\n\nThierry Bertin-Mahieux\nSun Labs, Sun Microsystems\n\nBurlington, Mass, USA\n\nbertinmt@iro.umontreal.ca\n\nStephen Green\n\nSun Labs, Sun Microsystems\n\nBurlington, Mass, USA\n\nstephen.green@sun.com\n\nAbstract\n\nSocial tags are user-generated keywords associated with some resource on the\nWeb. In the case of music, social tags have become an important component of\n\u201cWeb2.0\u201d recommender systems, allowing users to generate playlists based on\nuse-dependent terms such as chill or jogging that have been applied to particular\nsongs. In this paper, we propose a method for predicting these social tags directly\nfrom MP3 \ufb01les. Using a set of boosted classi\ufb01ers, we map audio features onto\nsocial tags collected from the Web. The resulting automatic tags (or autotags)\nfurnish information about music that is otherwise untagged or poorly tagged, al-\nlowing for insertion of previously unheard music into a social recommender. This\navoids the \u201dcold-start problem\u201d common in such systems. Autotags can also be\nused to smooth the tag space from which similarities and recommendations are\nmade by providing a set of comparable baseline tags for all tracks in a recom-\nmender system.\n\n1 Introduction\n\nSocial tags are a key part of \u201cWeb 2.0\u201d technologies and have become an important source of in-\nformation for recommendation. In the domain of music, Web sites such as Last.fm use social tags\nas a basis for recommending music to listeners. In this paper we propose a method for predicting\nsocial tags using audio feature extraction and supervised learning. These automatically-generated\ntags (or \u201cautotags\u201d) can furnish information about music for which good, descriptive social tags\nare lacking. Using traditional information retrieval techniques a music recommender can use these\nautotags (combined with any available listener-applied tags) to predict artist or song similarity. The\ntags can also serve to smooth the tag space from which similarities and recommendations are made\nby providing a set of comparable baseline tags for all artists or songs in a recommender.\nThis is not the \ufb01rst attempt to predict something about textual data using music audio as input.\nWhitman & Rifkin [10], for example, provide an audio-driven model for predicting words found\nnear artists in web queries . One main contribution of the work in this paper lies in the scale of our\nexperiments. As is described in Section 4 we work with a social tag database of millions of tags\napplied to \u223c 100, 000 artists and an audio database of \u223c 90, 000 songs spanning many of the more\npopular of these artists. This compares favorably with previous attempts which by and large treat\nonly very small datasets (e.g. [10] used 255 songs drawn from 51 artists.)\n\n\u2217Eck and Bertin-Mahieux currently at Dept. of Computer Science, Univ. of Montreal, Montreal , Canada\n\n1\n\n\fThis paper is organized as follows: in Section 2 we describe social tags in more depth, including\na description of how social tags can be used to avoid problems found in traditional collaborative\n\ufb01ltering systems, as well as a description of the tag set we built for these experiments. In Section 3\nwe present an algorithm for autotagging songs based on labeled data collected from the Internet.\nIn Section 4 we present experimental results and also discuss the ability to use model results for\nvisualization. Finally, in Section 5 we describe our conclusions and future work.\n\n2 Using social tags for recommendation\n\nAs the amount of online music grows, automatic music recommendation becomes an increasingly\nimportant tool for music listeners to \ufb01nd music that they will like. Automatic music recommenders\ncommonly use collaborative \ufb01ltering (CF) techniques to recommend music based on the listening\nbehaviors of other music listeners. These CF recommenders (CFRs) harness the \u201cwisdom of the\ncrowds\u201d to recommend music. Even though CFRs generate good recommendations there are still\nsome problems with this approach. A signi\ufb01cant issue for CFRs recommenders is the cold-start\nproblem. A recommender needs a signi\ufb01cant amount of data before it can generate good recom-\nmendations. For new music, music by an unknown artist with few listeners, a CFR cannot generate\ngood recommendations. Another issue is the lack of transparency in recommendations [7]. A CFR\ncannot tell a listener why an artist was recommended beyond the description: \u201cpeople who listen to\nX also listen to Y\u201d. Also, a CFR is relatively insensitive to multimodal uses of the same album or\nsong. For example songs from an album (a single purchase in a standard CFR system) may be used\nin the context of dining, jogging and working. In each context, the reason the song was selected\nchanges.\nAn alternative style of recommendation that addresses many of the shortcomings of a CFR is to\nrecommend music based upon the similarity of \u201csocial tags\u201d that have been applied to the music.\nSocial tags are free text labels that music listeners apply to songs, albums or artists. Typically, users\nare motivated to tag as a way to organize their own personal music collection. The real strength\nof a tagging system is seen when the tags of many users are aggregated. When the tags created by\nthousands of different listeners are combined, a rich and complex view of the song or artist emerges.\nTable 1 show the top 21 tags and frequencies of tags applied to the band \u201cThe Shins\u201d. Users have\napplied tags associated with the genre (Indie, Pop, etc.), with the mood (mellow, chill), opinion\n(favorite, love), style (singer-songwriter) and context (Garden State). From these tags and their\nfrequencies we learn much more about \u201cThe Shins\u201d than we would from a traditional single genre\nassignment of \u201cIndie Rock\u201d.\nIn this paper, we investigate the automatic generation of tags with properties similar to those gener-\nated by social taggers. Speci\ufb01cally, we introduce a machine learning algorithm that takes as input\nacoustic features and predicts social tags mined from the web (in our case, Last.fm). The model\ncan then be used to tag new or otherwise untagged music, thus providing a partial solution to the\ncold-start problem.\nFor this research, we extracted tags and tag frequencies for nearly 100,000 artists from the social\nmusic website Last.fm using the Audioscrobbler web service [1]. The majority of tags describe\naudio content. Genre, mood and instrumentation account for 77% of the tags. See \u201cextra material\u201d\nfor a breakdown of tag types.\nOvercoming the cold-start problem is the primary motivation for this area of research. For new music\nor sparsely tagged music, we predict social tags directly from the audio and apply these automati-\ncally generated tags (called autotags) in lieu of traditionally applied social tags. By automatically\ntagging new music in this fashion, we can reduce or eliminate much of the cold-start problem.\n\n3 An autotagging algorithm\n\nWe now describe a machine learning model which uses the meta-learning algorithm AdaBoost [5]\nto predict tags from acoustic features. This model is an extension of a previous model [3] which won\nthe Genre Prediction Contest and was the 2nd place performer in the Artist Identi\ufb01cation Contest at\nMIREX 2005 (ISMIR conference, London, 2005). The model has two principal advantages. First\nit selects features based on a feature\u2019s ability to minimize empirical error. We can therefore use the\n\n2\n\n\fTag Freq\nThe Shins\n2375\nIndie\nFavorites\n1138\nIndie rock\nEmo\n841\nIndie pop\nMellow\nAlternative\n653\n512\nRock\nFolk\n298 Alternative rock\nSeen Live\nPop\n231\nAcoustic\n\nTag Freq\n190\n138\n113\n85\n85\n83\n54\n\nTag Freq\n49\nPunk\n45\nChill\n41\nSinger-songwriter\nGarden State\n39\n37\nFavorite\n36\nElectronic\nLove\n35\n\nTable 1: Top 21 tags applied to The Shins\n\nFigure 1: Overview of our model\n\nmodel to eliminate useless feature sets by looking at the order in which those features are selected.\nWe used this property of the model to discard many candidate features such as chromagrams (which\nmap spectral energy onto the 12 notes of the Western musical scale) because the weak learners\nassociated with those features were selected very late by AdaBoost. Second, though AdaBoost may\nneed relatively more weak learners to achieve the same performance on a large dataset than a small\none, the computation time for a single weak learner scales linearly with the number of training\nexamples. Thus AdaBoost has the potential to scale well to very large datasets. Both of these\nproperties are general to AdaBoost and are not explored further in this short paper. See [5, 9] for\nmore.\n\n3.1 Acoustic feature extraction\n\nThe features we use include 20 Mel-Frequency Cepstral Coef\ufb01cients, 176 autocorrelation coef\ufb01-\ncients computed for lags spanning from 250msec to 2000msec at 10ms intervals, and 85 spectro-\ngram coef\ufb01cients sampled by constant-Q (or log-scaled) frequency (see [6] for descriptions of these\nstandard acoustic features.)\nThe audio features described above are calculated over short windows of audio ( 100ms with 25ms\noverlap). This yields too many features per song for our purposes. To address this, we create \u201caggre-\ngate\u201d features by computing individual means and standard deviations (i.e., independent Gaussians)\nof these features over 5s windows of feature data. When \ufb01xing hyperparameters for these experi-\nments, we also tried a combination of 5s and 10s features, but saw no real improvement in results.\nFor reasons of computational ef\ufb01ciency we used random sampling to retain a maximum of 12 ag-\ngregate features per song, corresponding to 1 minute of audio data.\n\n3.2 Labels as a classi\ufb01cation problem\n\nIntuitively, automatic labeling would be a regression task where a learner would try to predict tag\nfrequencies for artists or songs. However, because tags are sparse (many artist are not tagged at all;\nothers like Radiohead are heavily tagged) this proves to be too dif\ufb01cult using our current Last.fm\n\n3\n\nArtist A80s rockcoolSong 180s rockcoolSONG TAGGINGLEARNING \u201980s\u2019 TAGSong 1audio features target: \u201980s\u2019none/some/a lot\u201980s\u2019 boostertrainingPREDICTIONSET OF BOOSTERSnew songpredicted tags\fdataset. Instead, we chose to treat the task as a classi\ufb01cation one. Speci\ufb01cally, for each tag we try to\npredict if a particular artist has \u201cnone\u201d, \u201csome\u201d or \u201ca lot\u201d of a particular tag relative to other tags.\nWe normalize the tag frequencies for each artist so that artists having many tags can be compared to\nartists having few tags. Then for each tag, an individual artist is placed into a single class \u201cnone\u201d,\n\u201csome\u201d or \u201ca lot\u201d depending on the proportion of times the tag was assigned to that artist relative\nto other tags assigned to that artist. Thus if an artist received only 50 rock tags and nothing else, it\nwould be treated as having \u201ca lot\u201d of rock. Conversely, if an artist received 5000 rock tags but 10,000\njazz tags it would be treated as having \u201csome\u201d rock and \u201ca lot\u201d of jazz. The speci\ufb01c boundaries\nbetween \u201cnone\u201d, \u201csome\u201d and \u201ca lot\u201d were decided by summing the normalized tag counts or all\nartists, generating a 100-bin histogram for each tag and moving the category boundaries such that\nan equal number of artists fall into each of the categories. In Figure 2 the histogram for \u201crock\u201d is\nshown (with only 30 bins to make the plot easier to read). Note that most artists fall into the lowest\nbin (no or very few instances of the \u201crock\u201d tag) and that otherwise most of the mass is in high bins.\nThis was the trend for most tags and one of our motivations for using only 3 bins. As described in\nthe paper we do not directly use the predictions of the \u201csome\u201d bin. Rather it serves as a class for\nholding those artists for which we cannot con\ufb01dently say \u201cnone\u201d or \u201ca lot\u201d. See Figure 2 for an\nexample.\n\nFigure 2: A 30-bin histogram of the proportion of \u201crock\u201d tags to other tags for all songs in the dataset.\n\n3.3 Tag prediction with AdaBoost\n\nAdaBoost [5] is a meta-learning method that constructs a strong classi\ufb01er from a set of simpler\nclassi\ufb01ers, called weak learners in an iterative way. Originally intended for binary classi\ufb01cation,\nthere exist several ways to extend it to multiclass classi\ufb01cation. We use AdaBoost.MH [9] which\ntreats multiclass classi\ufb01cation as a set of one-versus-all binary classi\ufb01cation problems.\nIn each\niteration t, the algorithm selects the best classi\ufb01er, called h(t) from a pool of weak learners, based\non its performance on the training set, and assigns it a coef\ufb01cient \u03b1(t). The input to the weak\nlearner is a d-dimensional observation vector x \u2208 <d containing audio features for one segment of\naggregated data (5 seconds in our experiments). The output of h(t) is a binary vector y \u2208 {\u22121, 1}k\nl = 1 means a vote for class l by a weak learner while h(t), \u22121 is a vote\nover the k classes. h(t)\nagainst. After T iterations, the algorithm output is a vector-valued discriminant function:\n\nTX\n\ng(x) =\n\n\u03b1(t)h(y)(x)\n\n(1)\n\nAs weak learners we used single stumps, e.g. a binary threshold on one of the features. In previous\nwork we also tried decision trees without any signi\ufb01cant improvement. Usually we obtain a single\nlabel by taking the class with the most votes i.e f(x) = arg maxl gl(x), but in our model, we use\nthe output value for each class rather than the argmax.\n\nt=1\n\n3.4 Generating autotags\n\nFor each aggregate segment, a booster yields a prediction over the classes \u201cnone\u201d, \u201csome\u201d, and \u201ca\nlot\u201d. A booster\u2019s raw output for a single segment might be (none:\u22123.56) (some:0.14) (a lot:2.6).\n\n4\n\n\fThese segment predictions can then be combined to yield artist-level predictions. This can be\nachieved in two ways: a winning class can be chosen for each segment (in this example the class \u201ca\nlot\u201d would win with 2.6) and the mean over winners can be tallied for all segments belonging to an\nartist. Alternately we can skip choosing a winner and simply take the mean of the raw outputs for an\nartist\u2019s segments. Because we wanted to estimate tag frequencies using booster magnitude we used\nthe latter strategy.\nThe next step is to transform these class for our individual social tag boosters into a bag of words to\nbe associated with an artist. The most naive way to obtain a single value for rock is to look solely\nat the prediction for the \u201ca lot\u201d class. However this discards valuable information such as when a\nbooster votes strongly \u201cnone\u201d. A better way to obtain a measure for rock-ness is to take the center\nof mass of the three values. However, because the values are not scaled well with respect to one\nanother, we ended up with poorly scaled results. Another intuitive idea is simply to subtract the\nvalue of the \u201cnone\u201d bin from the value of the \u201ca lot\u201d bin, the reasoning being that \u201cnone\u201d is truly\nthe opposite of \u201ca lot\u201d. In our example, this would yield a rock strength of 7.16. In experiments\nfor setting hyperparameters, this was shown to work better than other methods. Thus to generate\nour \ufb01nal measure of rock-ness, we ignore the middle bin (\u201csome\u201d). However this should not be\ntaken to mean that the middle \u201csome\u201d bin is useless: the booster needed to learn to predict \u201csome\u201d\nduring training thus forcing it to be more selective in predicting \u201cnone\u201d and \u201ca lot\u201d. As a large-\nmargin classi\ufb01er, AdaBoost tries to separate the classes as much as possible, so the magnitude of the\nvalues for each bin are not easily comparable. To remedy this, we normalize by taking the minimum\nand maximum prediction for each booster, which seems to work for \ufb01nding similar artists. This\nnormalization would not be necessary if we had good tagging data for all artists and could perform\nregression on the frequency of tag occurrence across artists.\n\n4 Experiments\n\nTo test our model we selected the 60 most popular tags from the Last.fm crawl data described in\nSection 2. These tags included genres such as \u201cRock\u201d, \u201cElectronica\u201d, and \u201cPost Punk\u201d, mood-\nrelated terms such as \u201cChillout\u201d. The full list of tags and frequencies are available in the \u201cextra\nmaterials\u201d. We collected MP3s for a subset of the artists obtained in our Audioscrobbler crawl.\nFrom those MP3s we extracted several popular acoustic features. In total our training and testing\ndata included 89924 songs for 1277 artists and yielded more than 1 million 5s aggregate features.\n\n4.1 Booster Errors\n\nAs described above, a classi\ufb01er was trained to map audio features onto aggregate feature segments\nfor each of the 60 tags. A third of the data was withheld for testing. Because each of the 60\nboosters needed roughly 1 day to process, we did not perform cross-validation. However each\nbooster was trained on a large amount of data relative to the number of decision stumps learned,\nmaking over\ufb01tting a remote possibility. Classi\ufb01cation errors are shown in Table 2. These errors are\nbroken down by tag in the annex for this paper. Using 3 bins and balanced classes, the random error\nis about 67%.\n\nSegment\nSong\n\nMean Median Min Max\n49.6\n40.93\n37.61\n46.6\n\n43.1\n39.69\n\n21.3\n17.8\n\nTable 2: Summary of test error (%) on predicting bins for songs and segments.\n\n4.2 Evaluation measures\n\nWe use three measures to evaluate the performance of the model. The \ufb01rst TopN compares two\nranked lists, a target \u201cground truth\u201d list A and our predicted list B. This measure is introduced in\n[2], and is intended to place emphasis on how well our list predicts the top few items of the target\nlist. Let kj be the position in list B of the jth element from list A. \u03b1r = 0.51/3, and \u03b1c = 0.52/3,\n\n5\n\n\fas in [2]. The result is a value between 0 (dissimilar) and 1 (identical top N),\n\nPN\nPN\nr\u03b1kj\nj=1 \u03b1j\nl=1(\u03b1r \u2217 \u03b1c)l\n\nc\n\nsi =\n\n(2)\n\nFor the results produced below, we look at the top N = 10 elements in the lists.\nOur second measure is Kendall\u2019s T au, a classic measure in collaborative \ufb01ltering which measures\nthe number of discordant pairs in 2 lists. Let RA(i) be the rank of the element i in list A, if i is not\nexplicitly present, RA(i) = length(A) + 1. Let C be the number of concordant pairs of elements\n(i, j), e.g. RA(i) > RA(j) and RB(i) < RB(j). In a similar way, D is the number of discordant\npairs. We use \u03c4\u2019s approximation in [8]. We also de\ufb01ne TA and TB the number of ties in list A and\nB. In our case, it\u2019s the number of pairs of artists that are in A but not in B, because they end up\nhaving the same position RB = length(B) + 1, and reciprocally. Kendall\u2019s tau value is de\ufb01ned as:\n\n\u03c4 =\n\nsqrt((C + D + TA)(C + D + TB))\n\nC \u2212 D\n\n(3)\n\nUnless otherwise noted, we analyzed the top 50 predicted values for the target and predicted lists.\nFinally, we compute what we call the TopBucket, which is simply the percentage of common ele-\nments in the top N of 2 ranked lists. Here as in Kendall we compare the top 50 predicted values\nunless otherwise noted.\n\n4.3 Constructing ground truth\n\nAs has long been acknowledged [4] one of the biggest challenges in addressing this task is to \ufb01nd a\nreasonable \u201cground truth\u201d against which to compare our results. We seek a similarity matrix among\nartists which is not overly biased by current popularity, and which is not built directly from the\nsocial tags we are using for learning targets. Furthermore we want to derive our measure using\ndata that is freely available data on the web, thus ruling out commercial services such as AllMusic\n(www.allmusic.com). Our solution is to construct our ground truth similarity matrix using correla-\ntions from the listening habits of Last.fm users. If a signi\ufb01cant number of users listen to artists A\nand B (regardless of the tags they may assign to that artist) we consider those two artists similar.\nOne challenge, of course, is that some users listen to more music than others and that some artists\nare more popular than others. Text search engines must deal with a similar problem: they want\nto ensure that frequently used words (e.g., system) do not outweigh infrequently used words (e.g.,\nprestidigitation) and that long documents do not always outweigh short documents. Search engines\nassign a weight to each word in a document. The weight is meant to represent how important that\nword is for that document. Although many such weighting schemes have been described (see [11]\nfor a comprehensive review), the most popular is the term frequency-inverse document frequency\n(or TF\u00d7IDF) weighting scheme. TF\u00d7IDF assigns high weights to words that occur frequently in a\ngiven document and infrequently in the rest of the collection. The fundamental idea is that words\nthat are assigned high weights for a given document are good discriminators for that document from\nthe rest of the collection. Typically, the weights associated with a document are treated as a vector\nthat has its length normalized to one.\nIn the case of LastFM, we can consider an artist to be a \u201cdocument\u201d, where the \u201cwords\u201d of the\ndocument are the users that have listened to that artist. The TF\u00d7IDF weight for a given user for a\ngiven artist takes into account the global popularity of a given artist and ensures that users who have\nlistened to more artists do not automatically dominate users who have listened to fewer artists. The\nresulting similarity measure seems to us to do a reasonable enough job of capturing artist similarity.\nFurthermore it does not seem to be overly biased towards popular bands. See \u201cextra material\u201d for\nsome examples.\n\n4.4 Similarity Results\n\nOne intuitive way to compare autotags and social tags is to look at how well the autotags reproduce\nthe rank order of the social tags. We used the measures in Section 4.2 to measure this on 100 artists\nnot used for training (Table 3). The results were well above random. For example, the top 5 autotags\nwere in agreement with the top 5 social tags 61% of the time.\n\n6\n\n\fTopN 10 Kendall (N=5) TopBucket (N=5)\n\nautotags\nrandom\n\n0.636\n0.111\n\n-0.099\n-0.645\n\n61.0%\n8.1%\n\nTable 3: Results for all three measures on tag order for 100 out-of-sample artists.\n\nA more realistic way to compare autotags and social tags is via their artist similarity predictions.\nWe construct similarity matrices from our autotag results and from the Last.fm social tags used for\ntraining and testing. The similarity measure we used wascosine similarity scos(A1, A2) = A1 \u2217\nA2/(||A1|| ||A2||) where A1 and A2 are tag magnitudes for an artist. In keeping with our interest in\ndeveloping a commercial system, we used all available data for generating the similarity matrices,\nincluding data used for training. (The chance of over\ufb01tting aside, it would be unwise to remove The\nBeatles from your recommender simply because you trained on some of their songs). The similarity\nmatrix is then used to generate a ranked list of similar artists for each artist in the matrix. These lists\nare used to compute the measures describe in Section 4.2. Results are found at the top in Table 4.\nOne potential \ufb02aw in this experiment is that the ground truth comes from the same data source as\nthe training data. Though the ground truth is based on user listening counts and our learning data\ncomes from aggregate tagging counts, there is still a clear chance of contamination. To investigate\nthis, we selected the autotags and social tags for 95 of the artists from the USPOP database [2]. We\nconstructed a ground truth matrix based on the 2002 MusicSeer web survey eliciting similarity rank-\nings between artists from appro 1000 listeners [2]. These results show much closer correspondence\nbetween our autotag results and the social tags from Last.fm than the previous test. See bottom,\nTable 4.\n\nGroundtruth Model\nLast.FM\n\nTopN 10 Kendall 50 TopBucket 20\n\nMusicSeer\n\nsocial tags\nautotags\nrandom\nsocial tags\nautotags\nrandom\n\n0.26\n0.118\n0.005\n0.237\n0.184\n0.051\n\n-0.23\n-0.406\n-0.635\n-0.182\n-0.161\n-0.224\n\n34.6%\n22.5%\n3.9%\n29.7%\n28.2%\n21.5%\n\nTable 4: Performance against Last.Fm (top) and MusicSeer (bottom) ground truth.\n\nIt is clear from these previous two experiments that our autotag results do not outperform the social\ntags on which they were trained. Thus we asked whether combining the predictions of the autotags\nwith the social tags would yield better performance than either of them alone. To test this we blended\nthe autotag similarity matrix Sa with the social tag matrix Ss using \u03b1Sa + (1 \u2212 \u03b1)Ss. The results\nshown in Figure 3 show a consistent performance increase when blending the two similarity sources.\nIt seems clear from these results that the autotags are of value. Though they do not outperform the\nsocial tags on which they were trained, they do yield improved performance when combined with\nsocial tags. At the same time they are driven entirely by audio and so can be applied to new, untagged\nmusic. With only 60 tags the model makes some reasonable predictions. When more boosters are\ntrained, it is safe to assume that the model will perform better.\n\n5 Conclusion and future work\n\nThe work presented here is preliminary, but we believe that a supervised learning approach to au-\ntotagging has substantial merit. Our next step is to compare the performance of our boosted model\nto other approaches such as SVMs and neural networks. The dataset used for these experiments\nis already larger than those used for published results for genre and artist classi\ufb01cation. However,\na dataset another order of magnitude larger is necessary to approximate even a small commercial\ndatabase of music. A further next step is comparing the performance of our audio features with other\nsets of audio features.\n\n7\n\n\fFigure 3: Similarity performance results when autotag similarities are blended with social tag simi-\nlarities. The horizontal line is the performance of the social tags against ground truth.\n\nWe plan to extend our system to predict many more tags than the current set of 60 tags. We expect\nthe accuracy of our system to improve as we extend our tag set, especially as we add tags such as\nClassical and Folk that are associated with whole genres of music. We will also continue exploring\nways in which the autotag results can drive music visualization. See \u201cextra examples\u201d for some\npreliminary work.\nOur current method of evaluating our system is biased to favor popular artists. In the future, we\nplan to extend our evaluation to include comparisons with music similarity derived from human\nanalysis of music. This type of evaluation should be free of popularity bias. Most importantly, the\nmachine-generated autotags need to be tested in a social recommender. It is only in such a context\nthat we can explore whether autotags, when blended with real social tags, will in fact yield improved\nrecommendations.\n\nReferences\n[1] Audioscrobbler. Web Services described at http://www.audioscrobbler.net/data/webservices/.\n[2] A. Berenzweig, B. Logan, D. Ellis, and B. Whitman. A large-scale evaluation of acoustic and subjective\nIn Proceedings of the 4th International Conference on Music Information\n\nmusic similarity measures.\nRetrieval (ISMIR 2003), 2003.\n\n[3] J. Bergstra, N. Casagrande, D. Erhan, D. Eck, and B. K\u00b4egl. Aggregate features and AdaBoost for music\n\nclassi\ufb01cation. Machine Learning, 65(2-3):473\u2013484, 2006.\n\n[4] D. Ellis, B. Whitman, A. Berenzweig, and S. Lawrence. The quest for ground truth in musical artist\nsimilarity. In Proceedings of the 3th International Conference on Music Information Retrieval (ISMIR\n2002), 2002.\n\n[5] Y. Freund and R.E. Shapire. Experiments with a new boosting algorithm. In Machine Learning: Pro-\n\nceedings of the Thirteenth International Conference, pages 148\u2013156, 1996.\n\n[6] B. Gold and N. Morgan. Speech and Audio Signal Processing: Processing and Perception of Speech and\n\nMusic. Wiley, Berkeley, California., 2000.\n\n[7] Jonathan L. Herlocker, Joseph A. Konstan, and John Riedl. Explaining collaborative \ufb01ltering recommen-\n\ndations. In Computer Supported Cooperative Work, pages 241\u2013250, 2000.\n\n[8] Jonathan L. Herlocker, Joseph A. Konstan, Loren G. Terveen, and John T. Riedl. Evaluating collaborative\n\n\ufb01ltering recommender systems. ACM Trans. Inf. Syst., 22(1):5\u201353, 2004.\n\n[9] R. E. Schapire and Y. Singer. Improved boosting algorithms using con\ufb01dence-rated predictions. Machine\n\nLearning, 37(3):297\u2013336, 1999.\n\n[10] Brian Whitman and Ryan M. Rifkin. Musical query-by-description as a multiclass learning problem. In\nIEEE Workshop on Multimedia Signal Processing, pages 153\u2013156. IEEE Signal Processing Society, 2002.\n\n[11] Justin Zobel and Alistair Moffat. Exploring the similarity space. SIGIR Forum, 32(1):18\u201334, 1998.\n\n8\n\n\f", "award": [], "sourceid": 892, "authors": [{"given_name": "Douglas", "family_name": "Eck", "institution": null}, {"given_name": "Paul", "family_name": "Lamere", "institution": null}, {"given_name": "Thierry", "family_name": "Bertin-mahieux", "institution": null}, {"given_name": "Stephen", "family_name": "Green", "institution": null}]}