Music4All A+A: Artist & Album Dataset
- Music4All A+A is a multimodal dataset for MIR that organizes data at artist and album levels, offering a higher semantic granularity than track-level datasets.
- It integrates diverse modalities such as metadata, genre labels, images, and text to support cross-level modeling and robustness studies.
- Empirical benchmarks demonstrate that visual cues from artist and album images are more effective for genre classification than textual descriptions.
Searching arXiv for papers on Music4All Artist and Album and closely related metadata-based MIR work. Music4All Artist and Album (Music4All A+A) is a multimodal dataset for Music Information Retrieval (MIR) organized at the levels of artists and albums rather than individual tracks. It was introduced to address a structural limitation of prior multimodal MIR resources, which predominantly operate at track granularity even though many music concepts, including genre characterization, recommendation, and browsing, are often instantiated at artist or album level (Geiger et al., 18 Sep 2025). Music4All A+A is built on top of Music4All-Onion and provides metadata, genre labels, image representations, and textual descriptors for 6,741 artists and 19,511 albums, while retaining linkage to track-level multimodal data and user–item interaction data inherited from the underlying ecosystem (Geiger et al., 18 Sep 2025). This positioning makes the dataset relevant not only for multimodal genre classification, but also for hierarchical MIR, metadata-based representation learning, recommendation, and studies of missing-modality robustness. Earlier work on supervised music representation learning from factual metadata provides a direct conceptual basis for the use of artist- and album-level supervision, showing that artist and album labels encode distinct but complementary similarity concepts and that joint supervision improves downstream transfer performance (Lee et al., 2019).
1. Dataset definition and design rationale
Music4All A+A was created as a higher-level counterpart to track-centric multimodal datasets. Its central premise is that music can be described at several granularities, and that artists and albums constitute semantically meaningful units for MIR tasks such as genre classification or recommendation (Geiger et al., 18 Sep 2025). In contrast to datasets where visual and textual signals are attached to tracks through music videos or lyrics, Music4All A+A curates direct artist- and album-level information, aiming to better align modalities with the semantic identity of those entities (Geiger et al., 18 Sep 2025).
The dataset is built on top of Music4All-Onion rather than being an isolated corpus. This is consequential because it preserves access to track-level data, including audio, video clips, and user–item interaction data, while adding artist- and album-level modalities (Geiger et al., 18 Sep 2025). A plausible implication is that the dataset supports cross-level modeling in which artist and album representations are studied jointly with track-level signals.
The motivation for artist/album granularity also aligns with prior metadata-supervised learning research. In representation learning from factual metadata, artist information is treated as a broad stylistic concept, album information as a more specific level than artist, and track information as the most localized concept; joint learning across these levels improves overall performance (Lee et al., 2019). This suggests that artist and album are not merely administrative fields, but useful supervision targets with distinct inductive biases.
2. Modalities, scale, and entity-level contents
Music4All A+A contains 6,741 artists and 19,511 albums (Geiger et al., 18 Sep 2025). For both entity types, it provides metadata, genre labels, image representations, and textual descriptors (Geiger et al., 18 Sep 2025). The dataset thus differs from track-level corpora not only in unit of analysis but also in the semantic alignment of the modalities.
For artists, the metadata include name, sorted name, type, gender, country of origin, life span, user-generated tags, and relevant URLs such as fan pages, images, and Wikipedia links (Geiger et al., 18 Sep 2025). For albums, the metadata include album name, performing artist, release date, number of distinct listeners, number of listens, user-generated tags, and track listing information comprising track name, duration, and artist information (Geiger et al., 18 Sep 2025). Album cover images are provided, and artist images are collected separately (Geiger et al., 18 Sep 2025).
Textual descriptors for both artists and albums are derived from the English Wikipedia introductory paragraph, with fallback via Wikidata if needed (Geiger et al., 18 Sep 2025). The image pipeline is entity-specific: artist images are downloaded through Wikidata or Wikimedia Commons using artist-related URLs, whereas album cover images are obtained from Last.fm and, if unavailable, from Wikimedia Commons (Geiger et al., 18 Sep 2025).
| Entity type | Modalities | Examples of fields |
|---|---|---|
| Artist | metadata, genre labels, image representations, textual descriptors | name, sorted name, type, gender, country of origin, life span, tags, URLs |
| Album | metadata, genre labels, image representations, textual descriptors | album name, performing artist, release date, listeners, listens, tags, track listing information |
Because Music4All A+A is linked to Music4All-Onion, it can be extended with track-level audio, video clips, and user–item interaction data (Geiger et al., 18 Sep 2025). This multilevel coupling distinguishes it from datasets that only expose entity metadata without lower-level media.
3. Construction procedure and data curation
The dataset construction begins from the Music4All-Onion music catalog and enriches it through the Last.fm API, Wikidata API, Wikimedia Commons, and Wikipedia API (Geiger et al., 18 Sep 2025). Artist data extraction uses structured metadata from Last.fm and MusicBrainz relations, artist images from Wikidata and Wikimedia Commons, and textual descriptions from Wikipedia (Geiger et al., 18 Sep 2025). Album data extraction uses metadata and track listings from Last.fm, cover images from Last.fm or Wikimedia Commons, and album descriptions from Wikipedia (Geiger et al., 18 Sep 2025).
The storage format is entity-centric. Metadata, textual descriptors, and tags are stored in per-entity JSON files named by MBID as <MBID>.json, while images are stored as <MBID>_<index>.jpg (Geiger et al., 18 Sep 2025). This file organization is directly aligned with entity-level experimentation and simplifies multimodal loading pipelines.
The filtering policy is strict: artists and albums with incomplete information are excluded, and only entities for which genres, textual descriptions, and images are all available are retained (Geiger et al., 18 Sep 2025). This ensures multimodal completeness for the benchmarked tasks, though it also implies a selection effect toward entities with sufficiently rich metadata coverage.
A related but methodologically distinct line of work shows how music metadata can be operationalized as supervision targets rather than as descriptive covariates. In supervised representation learning from factual metadata, a Siamese-style CNN is trained on audio segments so that segments from the same artist or album become more similar than those from different artists or albums (Lee et al., 2019). That work uses artist and album metadata as “similarity concepts,” reinforcing the role of Music4All A+A as a resource for metadata-based supervision rather than only downstream classification.
4. Genre label construction and label structure
Genre labels in Music4All A+A are derived from Last.fm user-generated tags (Geiger et al., 18 Sep 2025). The paper describes a two-step cleaning process. First, tags are normalized by converting to lowercase and trimming whitespace. Second, a tag is accepted only if the normalized form exactly matches a genre or style in a reference list based on Wikipedia’s “List of music genres and styles” (Geiger et al., 18 Sep 2025). The result is a cleaner multi-label genre setup.
The label space is large. Music4All A+A provides 659 genres at the artist level and 737 genres at the album level (Geiger et al., 18 Sep 2025). These counts are substantially larger than those reported for MuMu, with 250 genres, and MM-IMDb, with 26 genres (Geiger et al., 18 Sep 2025). The paper also emphasizes that Music4All A+A has a more balanced genre distribution than MuMu (Geiger et al., 18 Sep 2025).
The genre prediction setup is explicitly multi-label. Models output logits for each genre, apply a sigmoid per label, and predict a genre as present when the sigmoid output exceeds 0.5 (Geiger et al., 18 Sep 2025). The loss is binary cross-entropy with class-frequency-based weighting, and training uses early stopping based on validation F1 (Geiger et al., 18 Sep 2025). Splits are created with iterative stratification to obtain approximately 80% train and 20% test while preserving genre distributions (Geiger et al., 18 Sep 2025).
The use of genre labels derived from user-generated tags should not be conflated with factual metadata such as artist and album identity. Prior work distinguishes these signal types explicitly: semantic labels like genres require costly annotation, whereas artist, album, and track metadata are naturally attached to songs and can be used at scale for supervised representation learning (Lee et al., 2019). Music4All A+A contains both types of information, which makes it suitable for studying their interaction.
5. Benchmark tasks and empirical findings
The principal benchmark presented for Music4All A+A is multimodal multi-label genre classification for artists and albums (Geiger et al., 18 Sep 2025). The evaluated models include CLIP for unimodal baselines, specifically image-only and text-only, and the multimodal models MMBT, ViLT, and SBNet (Geiger et al., 18 Sep 2025). Evaluation uses sample-averaged F1 score (Geiger et al., 18 Sep 2025).
A central empirical result is that images are more informative than text for genre classification in this dataset (Geiger et al., 18 Sep 2025). For albums, CLIP image-only achieves 31.3, whereas CLIP text-only achieves 25.4. For artists, CLIP image-only achieves 26.0, whereas CLIP text-only achieves 23.9 (Geiger et al., 18 Sep 2025). The paper interprets this as evidence that album covers and artist photos encode genre cues more effectively than biographies or descriptions.
Among multimodal models, SBNet performs best on Music4All A+A, with 33.1 on albums and 29.4 on artists (Geiger et al., 18 Sep 2025). MMBT and ViLT perform worse than even some unimodal baselines on this dataset (Geiger et al., 18 Sep 2025). The same model family is also evaluated on MM-IMDb, where SBNet reaches 65.3, while CLIP text-only and visual-only reach 58.3 and 60.2 respectively (Geiger et al., 18 Sep 2025). The cross-domain comparison is not presented as direct score comparability, but as evidence that models competitive in the movie domain do not necessarily transfer well to music (Geiger et al., 18 Sep 2025).
| Setting | Albums | Artists |
|---|---|---|
| CLIP image-only | 31.3 | 26.0 |
| CLIP text-only | 25.4 | 23.9 |
| SBNet | 33.1 | 29.4 |
The dataset also supports missing-modality experiments. Availability percentages of 10%, 30%, 50%, 70%, 90%, and 100% are defined on the test set for a given modality, with nested subsets so that the 10% available subset is contained in the 30% subset, and so on (Geiger et al., 18 Sep 2025). Under this protocol, performance degrades as modality availability decreases. For albums, SBNet drops from 33.1 with both modalities present to about 13.0 when the image modality is missing, whereas performance remains around 30.5 when text is missing (Geiger et al., 18 Sep 2025). This indicates that image information is especially critical in the benchmarked genre setting.
6. Methodological relevance beyond genre classification
Although Music4All A+A is introduced through multimodal genre classification, the surrounding literature indicates broader methodological uses. In metadata-based representation learning, artist and album labels can supervise audio embeddings without requiring subjective semantic annotation (Lee et al., 2019). That work trains a Siamese-style CNN with anchor, positive, and negative samples, and compares artist-only, album-only, track-only, and joint artist + album + track models (Lee et al., 2019). The joint model is built by simply adding three loss functions from the three similarity concepts and sharing model parameters for all of them (Lee et al., 2019). The reported interpretation is that artist captures broader stylistic identity, album is more specific than artist, and track is the most localized and segment-sensitive concept (Lee et al., 2019). This provides a direct rationale for using Music4All A+A as a supervision resource for embedding learning.
Music4All-style artist and album metadata have also been used in long-tail genre classification through bag-level learning. In MATT, each bag is defined by a unique album ID plus artist ID pair, grouping all music segments that share that album and artist into the same bag under a Multi-instance Learning formulation (Liu et al., 2022). The model uses a multi-instance attention mechanism to weight the most informative segments in a bag, and this bag construction is explicitly motivated by the assumption that segments with the same album ID and artist ID share the same genre label (Liu et al., 2022). Although the reported experiments are on the FMA medium subset rather than Music4All A+A, the metadata organization is directly aligned with an artist-and-album entity view.
A different extension concerns artist similarity. A graph neural network approach trained with triplet loss combines artist content features with graph topology derived from known artist–artist similarity relations, embedding artists into a vector space for retrieval (Korzeniowski et al., 2021). This work is artist-centric rather than album-centric, but it is relevant to Music4All A+A because the dataset’s artist entities, metadata, and multimodal linkage can support artist similarity search, recommendation, and hybrid content-plus-structure models.
More recent multimodal retrieval work expands the role of metadata even further. CLaMP 3 constructs a shared representation space across audio, symbolic music, and multilingual text, explicitly using titles and artist names as retrieval keys in a metadata-enriched music–text pipeline (Wu et al., 14 Feb 2025). Album is not an explicit field in the main M4-RAG table, but the approach shows how textual metadata can act as a bridge between modalities (Wu et al., 14 Feb 2025). This suggests a path for adapting Music4All A+A to cross-modal retrieval tasks using artist names, titles, and potentially album names where available.
7. Scope, limitations, and research significance
Music4All A+A fills a specific gap in MIR benchmark design by shifting the unit of analysis from tracks to artists and albums while preserving linkage to track-level multimodal resources (Geiger et al., 18 Sep 2025). Its significance lies in enabling several lines of inquiry within a common framework: multimodal genre classification at higher semantic granularity, missing-modality robustness studies, recommendation informed by track-level interaction data, and metadata-based supervision using naturally occurring factual labels.
At the same time, the benchmarked results reveal nontrivial limitations. Several multimodal models that perform well on MM-IMDb struggle on Music4All A+A, indicating weak cross-domain generalization and a substantial domain dependence of multimodal genre modeling (Geiger et al., 18 Sep 2025). The heavy reliance on images in the reported classification experiments also shows that modality utility is asymmetric and task-specific (Geiger et al., 18 Sep 2025). A plausible implication is that future models for Music4All A+A will need more music-specific fusion strategies rather than straightforward transfer of architectures from other multimodal domains.
The dataset also inherits the constraints of its curation choices. Filtering for entities with genres, textual descriptions, and images improves multimodal completeness but may bias the retained set toward better-documented artists and albums (Geiger et al., 18 Sep 2025). Genre labels are based on cleaned Last.fm user-generated tags rather than expert annotation (Geiger et al., 18 Sep 2025). This does not invalidate the benchmark, but it means that label construction reflects platform-specific tagging practices.
Future uses suggested in the source material include extending experiments with audio, using user–item interaction data, exploiting weighted genre tags for regression, and employing the dataset for recommendation and hierarchical MIR (Geiger et al., 18 Sep 2025). Additional work in the broader Music4All ecosystem shows that artist-linked historical engagement signals can dramatically improve predictive tasks such as popularity estimation, although those results are track- and artist-centered rather than explicitly album-centered (Choudhary et al., 6 Dec 2025). This suggests that Music4All A+A is best understood not only as a fixed genre classification benchmark, but as a multilevel substrate for entity-centric MIR research in which artist, album, and track information can be combined in principled ways.