How Last.fm Recommends New Artists and Songs: Full Guide

Written by: Kai Eldridge, Music Discovery Editor, OnesToWatch | Last updated: August 3, 2026

Last.fm recommends new artists and songs through a four-step engine built on scrobble data and collaborative filtering.

1. Scrobble Tracking: Every track a user plays is logged as a scrobble. These timestamped plays create a listening history that becomes the core behavioral dataset.
2. Collaborative Filtering via Musical Neighbors: Last.fm compares a user’s scrobble history against millions of other listeners to identify Musical Neighbors, which are accounts with the highest overlap in listening patterns. The system then surfaces artists those neighbors play frequently.
3. Artist Similarity Graph Construction: Co-listening data builds a weighted graph where artists are nodes and edges represent how often they appear together in listener histories. This graph becomes a navigable map of sonic relationships.
4. Recommendation Delivery and Underground Bias: The system relies on actual listening behavior rather than editorial curation or label spend. Niche and underground artists with dedicated listeners accumulate strong edges in the graph and appear in recommendations alongside mainstream acts.

Key Takeaways from Last.fm’s Recommendation System

• Last.fm’s recommendation engine uses scrobble data and collaborative filtering to identify Musical Neighbors and surface artists based on real listening behavior rather than editorial or paid placement.
• The system builds a weighted artist similarity graph from co-listening patterns, which gives niche and underground artists a structural advantage when they attract dedicated listener clusters.
• Key challenges include cold-start issues for new artists, data sparsity for users with limited histories, and popularity bias that can disadvantage long-tail talent.
• Emerging artists can improve discoverability by maintaining accurate metadata, releasing consistently, building engaged listener bases that scrobble regularly, and securing editorial coverage.
• For fans and artists seeking human-curated discovery alongside algorithmic tools, OnesToWatch provides editorial validation and coverage of emerging talent.

Core Concepts Behind Last.fm Recommendations

Four terms define how Last.fm’s recommendation engine functions.

A scrobble is a single logged play event. When a user listens to a track through a connected player or app, Last.fm records the artist name, track title, album, and timestamp. Accumulated scrobbles form a granular listening history that functions as an implicit rating system. The more a user scrobbles an artist, the stronger the signal of preference.

Collaborative filtering is the core algorithmic method. It predicts what a listener may enjoy by comparing their behavior to similar users or items, relying on historical signals such as listens, skips, likes, and ratings without analyzing audio content or metadata. Last.fm applies user-based collaborative filtering. It finds listeners whose scrobble histories most closely resemble a given user’s and recommends artists those listeners play.

Musical Neighbors is Last.fm’s term for the cohort of users identified as most similar to a given account based on scrobble overlap. The tighter the overlap, the higher the neighbor ranking. Higher-ranked neighbors carry more weight when the system generates recommendations.

These neighbor relationships then roll up into a broader network of artist connections. An artist similarity graph is the data structure that maps relationships between artists. Music recommendation systems model artists as nodes where edges represent relationships derived from listener behavior, playlist co-occurrence, and audio characteristics. Last.fm’s version of this graph is built almost entirely from co-listening behavior. When large numbers of users scrobble Artist A and Artist B in close temporal proximity, a weighted edge forms between them.

To see how this plays out in real careers, explore OnesToWatch’s Top Artists To Watch in 2026, which highlights artists already gaining traction in these graphs.

How Last.fm’s Scrobble Landscape Operates

Last.fm’s recommendation mechanics treat the scrobble as the atomic unit of preference data. Last.fm provides behavioral similarity data based on scrobbling and listening patterns, offering more open access than closed streaming platform graphs, though the data can be noisy or outdated.

The Musical Neighbors feature exposes this process directly to users. A listener can view their top neighbors, inspect those neighbors’ libraries, and receive artist recommendations derived from that cohort. This transparency is unusual among recommendation systems and gives users a degree of interpretability that closed platforms do not offer. This open structure also reveals an interesting characteristic of how the system surfaces music.

The underground bias emerges from the mechanics themselves. Niche audiences form tighter affinity clusters in recommendation graphs, allowing music to propagate faster within subgenre-specific listener communities and aiding underground artist discovery. An emerging artist with a small but highly engaged listener base accumulates strong co-listening edges within that cluster. This pattern causes the artist to surface reliably for users whose Musical Neighbors overlap with that cluster, even without editorial placement or promotional spend.

OnesToWatch complements this dynamic by providing human editorial validation for artists who are already gaining traction in these niche clusters. Fans and industry professionals gain a curated entry point into the same discovery pipeline.

To see how editorial and behavioral signals intersect, review OnesToWatch’s Top Artists To Watch in 2026 after exploring your own Musical Neighbors.

Key Trends in Scrobble-Driven Discovery (2025–2026)

Recent data shows that niche scene growth is accelerating in ways that validate scrobble-driven discovery models. UK underground rap streams on SoundCloud grew nearly 300% in 2025, exemplified by Fakemink’s UK monthly streams rising 696% and global streams rising 3830% over the past year. This pattern of explosive niche growth extends beyond rap. The Eclectic New Indie scene has seen streams up more than 2.5× in 2025 with an 89% Gen Z fanbase.

Underground and emerging artists are surfaced through interaction signals such as reposts, comments, follows, and social discovery rather than traditional genre classification, which aligns directly with how Last.fm’s scrobble-based graph weights edges through behavioral signals rather than editorial labels.

On the research side, a 2026 paper accepted at UMAP 2026 reports that artist-aware methods can more than double Recall and NDCG compared to content-only baselines in music recommendation. Systems that incorporate artist-level listening history, the same type of data Last.fm aggregates through scrobbles, substantially outperform approaches that rely on audio features alone.

To see which emerging artists from these underground scenes are already breaking through, explore OnesToWatch’s 2026 list.

Challenges and Friction Points for Emerging Artists

Three structural problems limit the accuracy of scrobble-driven collaborative filtering for emerging artists.

The cold-start problem affects any new artist or track with insufficient listening history. Cold artists with limited prior tracks account for a minority of new-item interactions in music recommendation datasets, so the problem is real but narrower than many assume. For most new tracks, the releasing artist already has some listening history that can anchor recommendations.

Data sparsity compounds the cold-start issue. When a user has a limited scrobble history, the Musical Neighbors algorithm has less data to work with. Sparse data produces lower-confidence neighbor matches and less precise recommendations. For emerging or underground artists facing the cold-start problem, accurate metadata submission enables content-similarity recommendations until sufficient user listening data accumulates.

Popularity bias is a documented failure mode in collaborative filtering systems. Mainstream GCN-based collaborative filtering methods suffer from high computational cost, over-smoothing, and poor generalization under sparse supervision signals. These conditions disproportionately disadvantage long-tail emerging artists whose interaction data is sparse by definition.

For editorial validation that helps artists overcome the cold-start problem, explore OnesToWatch’s curated 2026 artist predictions.

Collaborative Filtering Approaches Used in Music Discovery

Music recommendation systems deploy two primary variants of collaborative filtering, often in combination.

User-based collaborative filtering, the method underlying Last.fm’s Musical Neighbors, establishes similarity between listeners by matching one user to another based on overlapping historical activity like track ratings, likes, and listening time, then recommends music that similar users have engaged with.

Item-based collaborative filtering surfaces new artists by identifying tracks that many of the same users have interacted with positively, then recommending the nearest or most similar items based on those shared interaction patterns.

The artist similarity graph is the structural output of both approaches. Visualized as a network, it places artists as nodes with weighted edges connecting them. An artist with strong edges to multiple high-degree nodes sits near the center of a discovery cluster and surfaces frequently in recommendations.

An emerging artist with a small but dedicated listener base forms a tight local cluster. This cluster has lower overall degree but high edge weight within that subgraph, which means the artist surfaces reliably for users whose Musical Neighbors belong to that cluster. Graph-based methods consistently outperform matrix-based collaborative filtering across accuracy metrics, and graph structure alone is sufficient to handle sparse items in the long-tail catalog segment.

Hybrid models combine collaborative filtering with content-based signals. Integrating content-based multimodal features into sequential recommenders on LastFM-1K data improved Recall by up to 95% and NDCG by up to 79% over baselines.

To see how these technical approaches translate into real-world breakthroughs, review artists featured in OnesToWatch’s Top 30 for 2026.

Last.fm vs. Spotify for Niche Discovery

Dimension	Last.fm	Spotify	Source
Primary recommendation signal	Scrobble-based co-listening behavior across open user data	Co-listening behavior, playlist co-occurrence, audio embeddings	Vohnic Music
Related artist visibility	Open graph with user-inspectable Musical Neighbors	Approximately 20 related artists exposed, internal graph not visible	Vohnic Music
Long-tail / underground bias	Niche clusters propagate faster within subgenre-specific communities	Popularity bias documented, high-degree nodes tend to dominate recommendations	Music Tomorrow; Izimov & Naizabayeva 2026
Cold-start handling	Sparse for new artists with no prior scrobble history	Hybrid content plus collaborative filtering, artist-aware methods more than double Recall vs. content-only	ACARec, UMAP 2026

For a curated view of artists thriving in these different ecosystems, compare your listening to the names on OnesToWatch’s Top Artists To Watch in 2026.

Practical Checklist to Improve Scrobble-Based Discovery

Fans and artists can take concrete steps to improve the quality of Last.fm recommendations and increase discoverability within the scrobble graph.

For fans:

• Connect all listening apps and devices to Last.fm to maximize scrobble volume and produce a more accurate Musical Neighbors match.
• Scrobble consistently across genres rather than only within a primary genre. This broader listening history produces more diverse neighbor matches and exposes you to artists outside your usual clusters.
• Review and engage with Musical Neighbors’ libraries regularly to surface artists outside your existing clusters.
• Tag tracks and artists accurately using Last.fm’s community tagging system to strengthen subgenre edges in the similarity graph.

For emerging artists:

• Ensure all releases carry accurate metadata, including title, artist name, album, and genre tags, so that content-similarity recommendations can function until sufficient scrobble data accumulates.
• Build an engaged listener base that scrobbles consistently. Building an engaged and active community around your music remains the key to running a successful and sustainable music career.
• Release music at a consistent cadence. Consistent listening data helps algorithmic identity stabilize, allowing subsequent releases to benefit from the established audience cluster.
• Pursue editorial coverage on platforms like OnesToWatch to drive initial listener acquisition and seed the scrobble graph with a dedicated early audience.

Discover which artists have already applied these strategies in OnesToWatch’s Top 30 Artists To Watch in 2026.

Future Outlook for Scrobble-Based Recommendation

Scrobble-driven collaborative filtering remains an active area of innovation. The 2026 research landscape shows continued investment in improving its accuracy for the use cases where Last.fm has structural advantages: long-tail discovery, niche scene propagation, and emerging artist surfacing. The SaFeAU framework augments collaborative filtering with semantic factor routing to reduce false negative labeling and capture high-order signals without graph aggregation, outperforming state-of-the-art GCN-based methods across four sparse real-world datasets.

The integration of multimodal signals, including audio embeddings, lyric embeddings, and LLM-generated musicological metadata, into scrobble-based sequential recommenders represents the next evolution of the field. The enriched LastFM-1K benchmark supports future research on cold-start and long-tail music recommendation by providing audio feature coverage across 81 genres. As these hybrid systems mature, the behavioral foundation that Last.fm has accumulated over two decades becomes more valuable as the substrate on which richer multimodal models are trained.

For fans and artists who want discovery that operates ahead of these algorithmic curves, human-curated platforms remain essential. OnesToWatch has covered over 850 artists across the past decade, with alumni including Billie Eilish, Chappell Roan, Olivia Rodrigo, and Doechii. This track record of early-signal detection, the same capability scrobble-based systems attempt to automate, demonstrates the continued value of human curation.

To explore current editorial picks that may soon dominate recommendation graphs, visit OnesToWatch’s Top Artists To Watch in 2026.

FAQ

Does Last.fm recommend underground music?

Yes. Last.fm’s scrobble-driven collaborative filtering has a structural bias toward underground and niche music because it weights recommendations by actual listening behavior rather than editorial placement or promotional spend. When a small but highly engaged listener base consistently scrobbles an underground artist, that artist accumulates strong co-listening edges within their subgenre cluster. Users whose Musical Neighbors belong to that cluster will encounter the artist in recommendations regardless of mainstream visibility. Niche audiences form tighter affinity clusters in recommendation graphs, which allows music to propagate faster within subgenre-specific listener communities than on platforms optimized for broad catalog engagement.

What is a scrobble and why does it matter for recommendations?

A scrobble is a logged play event recorded by Last.fm each time a user listens to a track through a connected app or device. It captures the artist name, track title, album, and timestamp. Accumulated scrobbles form a detailed listening history that functions as an implicit preference dataset. The more frequently a user scrobbles an artist, the stronger the signal of affinity. This data feeds directly into the Musical Neighbors algorithm, which compares scrobble histories across millions of users to identify listeners with the highest behavioral overlap and generate artist recommendations from their combined libraries. Without scrobbles, the recommendation engine has no behavioral signal to work with.

How is Last.fm’s recommendation method different from Spotify’s?

Last.fm relies primarily on open scrobble-based co-listening data contributed by its user community and exposes the Musical Neighbors mechanism directly to users, which makes the recommendation logic partially interpretable. Spotify combines co-listening behavior with playlist co-occurrence and audio embeddings derived from tempo, energy, acousticness, and instrumentation, and its internal similarity graph is not visible to users. The public interface typically exposes only around 20 related artists with no explanation of the connections. Spotify’s system is more sophisticated in handling cold-start scenarios through hybrid content and collaborative filtering methods, but it carries a documented popularity bias in which high-degree nodes dominate recommendations. Last.fm’s open behavioral graph tends to surface niche and underground artists more reliably for listeners who already have relevant Musical Neighbors.

How can an emerging artist improve their discoverability on Last.fm?

An emerging artist can improve Last.fm discoverability through several concrete actions. Accurate and complete metadata on all releases, including artist name, album title, and genre tags, enables content-similarity signals to function before sufficient scrobble data accumulates. Building a dedicated listener base that scrobbles consistently is the most direct way to generate the co-listening edges that drive the similarity graph. Releasing music at a consistent cadence allows time for algorithmic identity to stabilize and an artist’s audience cluster to become well defined. Pursuing editorial coverage on platforms like OnesToWatch can seed the initial listener base needed to establish meaningful scrobble volume and begin forming Musical Neighbor connections with relevant listener cohorts.

Conclusion: Why Last.fm Still Matters for Discovery

Last.fm recommends new artists and songs through a scrobble-driven collaborative filtering engine that identifies Musical Neighbors, constructs a weighted artist similarity graph from co-listening behavior, and delivers recommendations with a structural bias toward niche and underground talent. The four-step process of scrobble tracking, Musical Neighbor identification, similarity graph construction, and recommendation delivery produces results that differ meaningfully from mainstream platform algorithms, particularly for listeners and artists operating outside the commercial mainstream.

The 2025–2026 research landscape confirms that behavioral listening data remains the most powerful signal in music recommendation and that artist-aware and graph-based methods are extending its accuracy further into long-tail discovery scenarios. For fans who want to go beyond what any algorithm surfaces, and for emerging artists who want a curated editorial pipeline alongside their algorithmic presence, OnesToWatch provides a human-curated complement to scrobble-based discovery.