In this position paper we discuss the use of information visualization techniques as a mean to find and characterize inconsistencies in music datasets. This idea is supported by empirical findings in a previous work dedicated to the visualization of a large dataset of music metadata called WASABI. During the development process of a visualization technique for exploring the multitude of multimedia attributes in the WASABI dataset (which includes lyrics, chords, audio, graphics describing sound analysis, etc.), we found visual patterns suggesting data inconsistencies (ex. ambiguities, inaccuracies, missing data, conflicts, etc.), which might have occurred during the integration from diverse sources. Traditionally, information visualization techniques are used to understand the data corpus and identify causal relationships, trends, patterns of data concentrations. Nevertheless, our findings suggest that information visualization techniques can be used to inspect data quality and highlight the parts of the datasets that need to be corrected/improved. Furthermore, we suggest that information visualization could be used as an entry point for repairing the dataset. More specifically, our aim is to use information visualization techniques to: communicate data quality issues to users, compare the outcomes of methods (such as crowdsourcing, matrix vectorization, graph reasoning, among other) used to fix the dataset, and observe the evolution of problem solving during the maintenance of the dataset.
This position paper presents our findings and ideas raised whilst creating a knowledge
graph for music datasets. Working on a knowledge graph makes it possible to
interlink resources with rich semantic relationships, which helps increase knowledge. If we
take for example a music dataset, it would offer a large scope of metadata describing
music, including multimedia dimensions that count textual (e.g. title, lyrics), graphical
(e.g. curves), and audio (e.g., the sound produced), as well as metadata qualifying the
works (e.g., date and time of recording, authors, performers...) and uses (e.g. song
covers, classification for a specific use,...). All these resources are given URIs that can be
referenced to documents describing them and can be shared across datasets. In order to
explore music datasets, we have proposed in a previous work [
Musical contents can be described in many multimedia dimensions (ex. lyrics, chords,
sounds, metadata...). Nonetheless, most of music datasets are specialized on a few types
of attributes or on a specific musical genre. In order to have a more complete description
of songs, some new datasets have been created by aggregating multiple data sources:
– The WASABI dataset [
Music datasets are often huge, containing hundreds of thousands, or millions of songs.
In this context, information visualization techniques might be a suitable alternative for
exploring music contents, identify patterns, trends, and correlations. According to a
recent survey done on visualizations for music related data [
Whilst inconsistencies are not frequent in the WASABI dataset, we were able to spot some of them during the exploration process using a visualization tool called MGExplorer. Figure 1 illustrates an example of duplicated entry for artists found during the exploration process. Hereafter we present a short list of the types of data quality defects we have found. We classify the problems in two categories, as follows: 1) Intrinsic defects encompass cases of duplicated data, missing data, disambiguation of artists with similar names, wrong formats for dates, conflicts between values of the same property collected from different data sources, broken links, etc.
2) Scenario problems related are specific to the application domain, such as: i) the same song with more than one producer or performer, according to the source; ii) a big difference between the number of songs produced in one single year, compared to others years (is it normal - related to some real life events? Is it an error during data collection?); iii ) songs classified in multiples genres, which might be the expression of errors in the classification of divergence of opinions among how songs could be classified. 2.4
Some of the methods we found in the literature allows both the detection and the
correction of dataset inconsistencies, as follows:
– Sourcing knowledge [
Whilst several methods for detecting problems with data quality exist, our research questions are focused on how to communicate these problems to the users. For example, it is a duplicated record and not a homonyms artists. In the example given by Figure 1, the disambiguation can be made by allowing users to explore other attributes associated to the artists. 3
As we shall see, detecting and solving data quality problems might require a decisionmaking processing involving users. For example, different algorithms might provide divergent analysis for music genre and the correct classification will depend on the users’ needs. For that, we propose to put users in the loop for detecting data quality problems and for deciding how the problems must be solved. Moreover, rather than propose completion/repairing methods (e.g., link prediction, duplication detection, missing URL, etc.), our approach allows users to apply existing methods and visualize the outcomes in terms of repairs proposed; ultimately users should be able to check if the corrections fulfill their needs (or not). Figure 2 presents a view at glance of our approach for detecting and solving quality problems whilst exploring music dataset using visualization techniques. Our approach encompasses the following premises: – Allow continuous integration of data from several sources, including results of audio and song lyrics analysis; In addition to standard music data, we will be adding emotion analysis, structure detection (verse, chorus, etc.), topic analysis, etc.; – Use information visualization techniques for multivariate data (such as MGExplorer and Parallel Coordinates) to represent all sort of data; – Identify the problems with data quality during the exploration process using visualization techniques; Figure 1 shows an example of a problem detected in the WASABI dataset using MGExplorer, which is a duplicate artist. – Solutions for detecting problems can be: manual, automatic or both; – Visualization techniques can accommodate the results of data correction adding an explanation to the context of use; – Compare the results of methods for detecting problems, allowing the users to decide about the outcomes of the corrections proposed. Validation of data correction/completion can be done quantitatively, by running against some measures (i.e number of missing properties, number of artists that require disambiguation) or qualitatively (i.e present a before/after visualization, and do user testing with real users to evaluate the refined data compared to what we had previously and evaluate the tools themselves).
This paper presented a preliminary work towards the visualization of music datasets. It includes a discussion about problems and potential solutions for integration methods for fixing data quality problems along the visualization process. As we have discussed, these data quality problems arose by continuous data integration (ex. new songs, new analysis of old songs, conflicts detected over time...). So that, we claim that data quality of music datasets is part of a continuous curation process rather than a definite state of the dataset. For that, we propose a methodology aiming at visualizing music datasets, highlighting problems found during the exploration process, as well as allowing to compare solutions to fix those problems. The next steps should be to deal with integration of corrective methods for fixing problems and support the versioning of data (a main requirement for comparison of the methods’ outcomes). For this purpose, MGExplorer will be used as a proof of concept to illustrate the feasibility of the methodology proposed. We are currently working on running prototype that could be used to test our hypothesis with end users. Finally, despite the fact that our work is focused on the music dataset WABASI, we suggest that the solution might be suitable to other types of datasets to be explored in the future.