LFM-2b Music Listening Dataset

• LFM-2b is a public music-listening dataset capturing billions of streaming events over 15 years, offering insights into user behavior and local music trends.
• It provides curated benchmark subsets and supports top-K music recommendation tasks using metrics like MRR@10 and local-share bias analysis.
• Its niche user profile and metadata limitations highlight the conditional nature of conclusions regarding local music fairness and global representativeness.

Searching arXiv for papers mentioning LFM-2b and related dataset context. LFM-2b is a public music-listening dataset associated with Last.fm users and used as a benchmark in recommender-systems research on top-$K$ music recommendation, large-scale listening behavior, and the representation of local music. In recent arXiv discussion, it is treated less as a neutral proxy for general streaming behavior than as a dataset with a specific cultural and population profile: a 2024 reproducibility study concludes that LFM-2b is valuable and convenient, but that conclusions about local music promotion drawn from it alone may not generalize to commercial streaming populations or remain stable under modest experimental changes (Matrosova et al., 2024).

1. Corpus scale and benchmark subsets

The reproducibility literature distinguishes between the full LFM-2b corpus and the smaller experimental subsets used for recommender evaluation. The cited construction of LFM-2b reports approximately 2 billion listening events, spanning over 15 years, from about 120,000 users. A later study of local-music bias reused a much smaller subset: about 14 million interactions from 2018 and 2019, involving 13,000 users in 20 countries, selected for having at least 100 users and artists who had collectively created at least 1,000 tracks (Matrosova et al., 2024).

Configuration	Reported scale	Scope
Full LFM-2b	approximately 2 billion listening events; over 15 years; about 120,000 users	Public Last.fm-based corpus
Lesota et al. subset	about 14 million interactions; 2018 and 2019; 13,000 users	20 countries; filtered by minimum user and track-count criteria
Three-country subset in the reproducibility study	over 3 million listening events; around 100,000 music tracks	France: 254 users; Germany: 805 users; Brazil: 1064 users

For the cross-dataset comparison with Deezer, the reproducibility study further restricted the Lesota subset to France, Germany, and Brazil, because those countries were available in both settings and were important Deezer markets. The resulting three-country subset was obtained from the original authors via private communication. This layered reuse is important for interpretation: many claims attributed to “LFM-2b” in the local-music literature are in fact claims about a filtered, temporally restricted, country-restricted subset rather than about the full corpus.

2. Task formulation and bias metric

Within the local-music line of work, LFM-2b supports a top-$K$ music recommendation task in which “local music” means music from the user’s own country. The evaluation protocol reported for both LFM-2b and Deezer uses top-10 track recommendation, MRR@10 for ranking quality, 10\% randomly selected users masked during training for validation, and a bias analysis over recommendation lengths $K \in \{10, 15, 20, \dots, 100\}$ (Matrosova et al., 2024).

The local-share baseline for a user $u$ is defined as the proportion of listened streams that are local:

$$L(u) = \frac{N_{\text{local}(u)}}{N_{\text{listened}(u)}}.$$

For a recommender $\mathrm{MRS}_K$ returning $K$ tracks, the local share of the recommendation list is:

$$L_{\mathrm{MRS}_K}(u) = \frac{N_{\text{local},\mathrm{MRS}_K}(u)}{K}.$$

The central quantity is the average deviation between recommended and listened local share:

$$\mathrm{Bias}_{\mathrm{MRS}_K} = \frac{1}{M} \sum_{u \in \mathcal{U}} \left( L_{\mathrm{MRS}_K}(u) - L(u) \right),$$

where $M$ is the number of users. Positive values indicate that recommendations contain more local music than users organically listen to; negative values indicate less local music than users organically listen to.

This formulation makes the interpretation of LFM-2b unusually sensitive to the dataset’s user population. Because bias is defined relative to observed listening behavior rather than against an external notion of fairness, any peculiarity in the LFM-2b listening baseline directly affects the sign and magnitude of measured recommender bias.

3. Population profile and representativeness

A major conclusion of the reproducibility study is that LFM-2b users are not representative of global music-streaming audiences. The study explicitly characterizes Last.fm users as tending to be active on the internet and social media, as not evenly distributed across countries, and as unlikely to reflect the diverse audience of music streaming services. It also relies on descriptions from the LFM-2b paper and related literature that portray the user base as music enthusiasts with a strong interest in collecting and organizing their music libraries; this, in turn, may skew behavior toward less mainstream genres and older catalog listening. The user base is also described as predominantly English-speaking, which introduces population bias (Matrosova et al., 2024).

These properties matter especially for “local music” studies. If a dataset overrepresents globally oriented, English-speaking, niche, or enthusiast listeners, then the empirical baseline $K$0 is already culturally specific. A plausible implication is that LFM-2b may be well suited for analyzing recommendation behavior in a historically important Last.fm-like listening regime, while being less reliable as a sole evidential basis for claims about music streaming populations at large.

The reproducibility study therefore does not reject LFM-2b as a benchmark. Rather, it repositions the dataset as informative but non-representative for certain fairness questions. That distinction is central to current use of LFM-2b in recommender-systems research.

4. Local-music consumption patterns observed on LFM-2b

The strongest empirical contrast reported for LFM-2b concerns the prevalence and distribution of local music. Using MusicBrainz labels and considering labeled tracks only, the Deezer comparison dataset shows substantially more local music than LFM-2b. For Brazil, LFM-2b has 2.5× fewer local streams than Deezer. The difference is not only in the mean: on Deezer, user-level local-share histograms span the full spectrum from users who listen to no local music to users who listen to only local music, whereas on LFM-2b relatively few users listen to a majority of local music (Matrosova et al., 2024).

The study also presents a qualitative comparison for France. In the LFM-2b France subset, top tracks are described as dominated by older releases from the 1980s–2010s, mostly English-language, and concentrated in genres such as indie rock, alternative rock, and trip hop. Examples given include Portishead, Radiohead, The Verve, Franz Ferdinand, and The xx. By contrast, the Deezer France top tracks are described as more recent, more French-language, and more oriented toward rap, pop, and EDM.

These observations are used in the literature as evidence that LFM-2b captures a more niche, historically oriented, and Anglophone-leaning listener population than a commercial streaming platform. This suggests that the dataset’s most distinctive empirical value lies not in demographic representativeness, but in documenting a specific listening culture at scale.

5. Dependence of recommender findings on experimental setting

The local-music literature around LFM-2b is notable for showing that conclusions can be both reproducible and fragile. In the original global-training setting with MusicBrainz labels and $K$1, the reproducibility study reports broadly similar conclusions to the earlier LFM-2b work: NeuMF tends to have negative local bias, while ItemKNN is more favorable to local music, with positive bias in Brazil and Germany and slightly negative but small bias in France. These conclusions remain broadly consistent across varying $K$2 in that specific global setting (Matrosova et al., 2024).

However, the same study shows that the headline interpretation is conditional rather than universal. On Deezer, under the same nominal global-training setup and at $K$3, all models have positive average bias. Even within LFM-2b, conclusions can reverse when the setup changes: with country-specific training at $K$4, NeuMF becomes positively biased in Germany and ItemKNN becomes negatively biased in Brazil. For some settings, NeuMF changes sign as $K$5 increases, being negative for $K$6 and positive for $K$7.

The study attributes this instability to several factors: recommendation list size $K$8, global versus country-specific training, training randomness, and metadata choice. Bias values were averaged over 20 runs with $K$9 standard deviation intervals, and the paper explicitly notes that in some LFM-2b cases NeuMF shows both positive and negative bias across runs. The recommenders examined were ItemKNN and NeuMF, both in RecBole implementations; ItemKNN used cosine similarity on the user-track train interaction matrix with shrink parameter 0, while NeuMF used Adam, learning rate 0.001, batch size 512, dropout 0.1, binary cross-entropy, embedding dimension 64, and up to 300 epochs. In this literature, LFM-2b therefore functions not merely as a static benchmark but as a test case for the sensitivity of fairness conclusions to modeling and evaluation choices.

6. Metadata limitations, access issues, and name disambiguation

A central caveat in LFM-2b-based local-music research is that “local” is mediated by metadata rather than observed directly. The reproducibility study emphasizes that assigning a country to an artist is inherently ambiguous, with competing notions such as country of activity, country of origin, and MusicBrainz country labels. It further argues that excluding unlabeled tracks, as done in the original LFM-2b local-music study, can discard a large fraction of data, distort estimated local-share baselines, and introduce label bias. The broader claim is that metadata incompleteness is not random: it may be systematically worse for less popular artists, certain locations, certain genres, and communities underrepresented among annotators (Matrosova et al., 2024).

The study also notes a reproducibility complication external to modeling: LFM-2b was recently taken down because of license issues, despite having previously been publicly available. This is one reason the authors publicly released an anonymized Deezer dataset and full Python source code as an alternative or complement for studying local music recommendation and fairness.

A separate source of confusion is terminological. An unrelated microscopy paper, "PNR: Physics-informed Neural Representation for high-resolution LFM reconstruction" (Zhao et al., 2024), discusses a new light-field microscopy dataset built around two-photon synthetic aperture microscopy (2pSAM), but the paper explicitly does not mention a dataset named “LFM-2b” anywhere in its main text, figures, tables, or references. Any association between that microscopy dataset and the music recommendation dataset is therefore unsupported by the paper itself. In current scholarly usage, “LFM-2b Dataset” refers to the music-listening benchmark, not to the 2pSAM light-field microscopy data.

Taken together, these strands of evidence position LFM-2b as an important but highly conditional benchmark. It remains useful for studying music recommendation, fairness, and large-scale listening behavior, yet for questions about local music representation it is best interpreted as a dataset with a narrow and historically specific listening regime, substantial dependence on metadata choices, and limited evidential force when used in isolation.