This paper summarizes our contribution (team DBIS) to the AcousticBrainz Genre Task: Content-based music genre recognition from multiple sources as part of MediaEval 2017. We utilize a hierarchical set of multilabel classifiers to predict genres and subgenres and rely on a voting scheme to predict labels across datasets.
In the MediaEval AcousticBrainz Genre Task, the goal is to classify
tracks into main and subgenres, using content-based features
computed with Essentia [
There are multiple factors that make the posed task dificult to solve, particularly the large amount of data to handle and the multilabel nature of the classification problem make the tasks highly challenging. Subtask 2 is further complicated by the fact that genre and subgenre labels are hardly consistent across the four provided training sets, hence providing a heterogeneous set of labels.
In the following, we firstly sketch our approach to mitigate these dificulties. Next, we detail the classification approaches we used throughout subtasks 1 and 2 and lastly, present the obtained results. We make our implementation available for reproducibility and for promoting research in this direction1.
Reducing the amount of data. To reduce the amount of data and make the task computationally feasible within the limited time frame, we at first skipped detailed features describing low level energy bands of the energy spectrum and verified on a preliminary basis that the respective central moments are suficient in terms of classification accuracy. This allowed us to reduce the number of features used for training the genre classifiers to 395 (from over 3,000 features originally provided). The full list of features can be found in our GitHub repository1. 1https://github.com/dbis-uibk/MusicGenreClassification
Multilabel classification. The fact that any track may feature multiple genres and subgenres complicates the classification problem, since not all classification algorithm inherently support multilabel classification. We solved this problem by applying the one-vs.-therest strategy, efectively training a separate binary classifier for every label.
Diferent genre labels across data sets. As subtask 2 allows to
combine all datasets for training, the (vastly) difering genre labels
used in the four available training sets posed a challenge. We
tackled this problem by computing a direct mapping between the main
class labels of all training sets aiming to find equivalent genre labels
across all datasets. Therefore, we applied the Levenshtein string
distance measure [
Classification Algorithms. We implemented our solution with
two diferent classification methods 2: (1) a linear C-support vector
machine [
Predict main genres for each track
Predict subgenres for each main genre predicted for each track Fallback: Assign most popular main genre to tracks with no predicted label
The workflow underlying our approach for subtask 1 is outlined
in Figure 1. First, we train one classifier for main genres and a
separate classifier for each main genre’s subgenres. After that, we
2We relied on the python library scikit-learn [
To exploit the possibility to submit five submission runs, this basic approach was implemented with the following configurations of classification algorithms for main and subgenre, which are also listed in Table 1: • Run #1 uses a SVM with C = 1.0 and no class weight balancing for the main genre classifier; an extra-trees classifier with 50 trees, p| f eatures | features considered when searching for the best split and balanced class weights for the subgenre classifiers. • Run #2 uses a SVM with C = 1.0 and balanced class weights for the main genre classifier; an extra-trees classifier with 50 trees, p| f eatures | features considered when searching for the best split and balanced class weights for the subgenre classifiers. • Run #3 includes a SVM with C = 10.0 and balanced class weights for the main- and subgenre classifiers.
The C value for the SVMs was selected after a grid search on a smaller test set of 10,000 randomly sampled tracks. The chosen amount of features and trees for the extra trees classifier was a trade of between classification runtime and accuracy, as more features would possibly have provided more accuracy. For runs #4 and #5, the results of run #3 were used. 2.2
For subtask 2, the set of all provided datasets could be utilized to classify each of the four test sets. We chose to implement this using a voting mechanism. First, SVMs as main genre classifiers were trained as in subtask 1, independently for every training set. These classifiers were then used to predict the main genres of a given track as follows: (1) Predict the main genres of the track with all four classifiers. (2) Utilize the genre mapping as described above to map the predicted genres to the genre labels of the current test set (otherwise, the predicted labels would not be compatible and hence, create false positives). Thereby, classification results where no class label was contained in (or could be mapped to) the test set were discarded. (3) For every genre predicted by any of the four classifiers, count the number of classifiers that predicted this genre (using two diferent weighing schemes) and weigh this by the number of classifiers that produced a usable result.
To arrive at the final set of main genres for every track, we applied two diferent variants, which can be seen in Table 1 for runs #4 and #5: (1) weigh every prediction equally and retain genres predicted by at least 50% of the usable classifiers—for example, if three of the four classifiers predict the label rock/pop, that label was predicted by 75% of the classifiers and is retained (run #4); (2) double the weight of the prediction of the classiefir trained specifically on the training set corresponding to the current test set and retain genres predicted by at least 60% of the usable classifiers (e.g., if we did predictions for the Lastfm test set and the Lastfm and Discogs classifiers predicted rock/pop, then that label was assigned three votes out of five (i.e., 60%) and retained (run #5)). This puts more emphasis on the predictions of the training set and hence, classifier that is trained on the naturally best training data (stemming from the same data source as the current test set).
Prediction of subgenres and handling of tracks with no predicted labels was handled the same way as in subtask 1. For this subtask, support vector machines were used as classifiers, with C = 10.0 and balanced class weights as determined in preliminary experiments. 3
The results of the evaluation of our approach for subtasks 1 and 2 can be found in Tables 2 and 3, respectively. Table 2 shows the results of run #3, which provided the best overall performance in both subtasks. Table 3 contains the results of run #5, which performed better in some measures (in bold font) compared to run #3 in subtask 2. Due to space limitations, the results of the other runs are omitted.
Possible improvements of the presented approaches include different classifying methods such as deep neural networks and a more detailed feature selection process. These steps were rendered impossible due to time constraints and technical limitations of the available hardware.