[Paper Review] Local Music Event Recommendation with Long Tail Artists

Douglas Turnbull, Luke Waldner|arXiv (Cornell University)|Sep 7, 2018

Music and Audio Processing15 references3 citations

TL;DR

This paper proposes a local music event recommendation system that leverages Latent Semantic Analysis (LSA) to model artist and tag similarities, enabling effective recommendations for long-tail, low-footprint local artists. It introduces a Music Event Graph and demonstrates that even minimal digital footprint data enables accurate recommendations, with AUC scores of up to 0.79 for expert users using early/late fusion strategies.

ABSTRACT

In this paper, we explore the task of local music event recommendation. Many local artists tend to be obscure long-tail artists with a small digital footprint. That is, it can be hard to find social tag and artist similarity information for many of the artists who are playing shows in the local music community. To address this problem, we explore using Latent Semantic Analysis (LSA) to embed artists and tags into a latent feature space and examine how well artists with small digital footprints are represented in this space. We find that only a relatively small digital footprint is needed to effectively model artist similarity. We also introduce the concept of a Music Event Graph as a data structure that makes it easy and efficient to recommend events based on user-selected genre tags and popular artists. Finally, we conduct a small user study to explore the feasibility of our proposed system for event recommendation.

Motivation & Objective

• To address the challenge of recommending local music events featuring obscure, long-tail artists with limited digital footprints.
• To develop a scalable, hybrid recommendation system that combines content-based tags and collaborative filtering via artist similarity.
• To reduce popularity bias in recommendations so that both popular and obscure local artists are fairly represented.
• To evaluate the system’s feasibility through a user study with local music experts and assess performance using AUC metrics.
• To explore the use of latent semantic analysis for sparse, real-world music event data where traditional content-based methods fail.

Proposed method

• Construct a 4-partite Music Event Graph that links genre tags, popular artists, event artists, and events for efficient retrieval.
• Apply Latent Semantic Analysis (LSA) to embed artist and tag data into a shared low-dimensional latent feature space.
• Use truncated SVD to compute LSA embeddings from sparse co-occurrence matrices of artists and tags in event listings.
• Implement early and late fusion strategies: average, cluster, or interleave user preference vectors to generate personalized event artist rankings.
• Apply k-means clustering to user preferences with k set to the rounded natural log of the number of preferences.
• Rank event artists using cosine similarity scores or average/interleaved ranking fusion to combine multiple user preference signals.

Experimental results

Research questions

• RQ1 Can LSA effectively model artist similarity and tag relationships even when artists have minimal digital footprints?
• RQ2 How does the performance of the recommendation system vary when relying on genre tags versus artist preferences alone?
• RQ3 To what extent can the system reduce popularity bias in event recommendations, ensuring visibility for long-tail artists?
• RQ4 How do different fusion strategies (early vs. late, average vs. interleave) impact recommendation accuracy for local music event discovery?
• RQ5 Can a user study with local music experts provide reliable ground truth for evaluating local event recommendation systems?

Key findings

• The system achieved an AUC of 0.79 (±0.09) when using artist preferences alone with the 'none / avg. cosine' fusion strategy, indicating strong recommendation accuracy.
• For genre preferences alone, the AUC was 0.69 (±0.16), showing that tag-based recommendations are less accurate than artist-based ones but still meaningful.
• When combining both artist and genre preferences, the best-performing approach ('none / avg. cosine') achieved an AUC of 0.74 (±0.12), demonstrating the value of hybrid signals.
• The 'cluster / interleave' strategy achieved an AUC of 0.75 (±0.09) for combined preferences, showing that clustering user preferences can maintain high performance.
• Less frequent attendees (monthly) showed significantly lower performance (AUC 0.61 ± 0.15), suggesting that user familiarity with the local music scene is critical for reliable evaluation.
• The popularity baseline achieved only 0.53 AUC, confirming that the proposed system significantly outperforms simple popularity-based ranking.