Modeling Artist Preferences of Users with Different Music Consumption Patterns for Fair Music Recommendations

Dominik Kowald, Elisabeth Lex, Markus SchedlarXiv 2019 · arXiv

TL;DR

Standard collaborative filtering algorithms often fail niche music listeners by disproportionately recommending mainstream tracks. This paper introduces $BLL_u$, a cognitive-inspired approach based on human memory decay (recency and frequency) that models individual artist preferences more fairly and accurately, particularly for non-mainstream listeners.

What Problem Does This Paper Address?

Modern music recommender systems rely heavily on Collaborative Filtering (CF), an approach that thrives on dense data and shared user behavior. Because mainstream music dominates streaming platforms, CF algorithms naturally favor popular artists. Consequently, users with unorthodox or niche tastes—termed "low-mainstream" listeners—receive poor, biased recommendations.

The authors address this popularity bias to create a fairer recommendation framework. The goal is to accurately model artist preferences across different music consumption patterns (from niche to mainstream) without defaulting to the lowest common denominator of popular music.

Methods

The authors propose $BLL_u$, an approach that adapts the Base-Level Learning (BLL) equation from the ACT-R cognitive architecture. Instead of relying on what other users like (which introduces mainstream bias), $BLL_u$ models an individual's preference based on their own cognitive patterns of memory retention and reinforcement.

The model quantifies the utility of an artist to a specific user by balancing two factors:

  1. Frequency: How often the user has listened to the artist in the past.
  2. Recency: How recently the user interacted with the artist, modeled using a time-dependent power-law decay.

By capturing these individual temporal drifts, the model personalizes recommendations using only the user's own history, inherently insulating the process from mainstream noise.

Datasets

The study utilizes the massive, publicly available LFM-1b dataset, which contains the listening histories of over 120,000 Last.fm users, totaling 1.1 billion individual listening events over nine years.

To evaluate fairness, the authors calculated a "mainstreaminess" ($MS$) factor for users and sampled three distinct groups of 1,000 users:

  • LowMS: Listeners of unorthodox, niche music ($MS = 0.125$).
  • MedMS: Average listeners centered around the median ($MS = 0.379$).
  • HighMS: Mainstream music listeners ($MS = 0.688$).

The dataset characteristics for these groups are outlined below:

Dataset statistics for the three user groups

Key Insights and Ideas

  • Cognitive Architectures Map to Digital Habits: Human memory decay (how we forget and reinforce information) mirrors how our music preferences drift over time.
  • Isolation Breeds Fairness: By evaluating a user's history independently of the crowd, the system completely bypasses the data sparsity issues that cause collaborative filtering to fail for niche listeners.
  • The "Just Now" Effect: For extremely short-term predictions (predicting the next 1 or 2 artists), pure recency beats complex modeling, highlighting how potent immediate context is in listening sessions.

Key Results

The authors evaluated $BLL_u$ against four baselines: global popularity (TOP), collaborative filtering ($CF_u$), user-specific popularity ($POP_u$), and user-specific recency ($TIME_u$). The evaluation used a time-based split, withholding the most recent 1% of listening events for testing.

  • Superior Accuracy: $BLL_u$ achieved the highest precision and recall across all three user groups when predicting a standard slate of top $k=1\dots20$ artists.
  • Empowering Niche Listeners: The performance boost was most pronounced for the LowMS group, where standard collaborative filtering ($CF_u$) performed poorly.
  • The Recency Trade-off: The pure recency baseline ($TIME_u$) outperformed $BLL_u$ only when predicting a very small number of artists ($k \le 2$). Beyond that, $BLL_u$’s blend of frequency and power-law decay proved superior.

The precision/recall curves across the three groups highlight these trends:

Figure 1: Recall/precision plots of the baselines and our BLLu approach for the three user groups LowMS, MedMS, and HighMS, and for k = 1 . . . 20 predicted artists. We see that BLLu provides the best prediction accuracy results for all user groups.

Limitations and Future Work

  • Simplistic Mainstreaminess Metric: The current definition of "mainstreaminess" is relatively basic. The authors plan to explore more sophisticated metrics, such as rank-order correlation or Kullback-Leibler (KL) divergence.
  • Hybrid Integration: Rather than using $BLL_u$ in isolation, the authors suggest integrating these cognitive preference values as a contextual dimension into matrix factorization models to combine the strengths of collaborative filtering and cognitive modeling.
  • Granularity: The current model operates strictly at the artist level; extending this time-aware cognitive approach to individual tracks or genres remains an open area for exploration.

Summary updated

Abstract

Music recommender systems have become central parts of popular streaming platforms such as Last.fm, Pandora, or Spotify to help users find music that fits their preferences. These systems learn from the past listening events of users to recommend music a user will likely listen to in the future. Here, current algorithms typically employ collaborative filtering (CF) utilizing similarities between users' listening behaviors. Some approaches also combine CF with content features into hybrid recommender systems. While music recommender systems can provide quality recommendations to listeners of mainstream music artists, recent research has shown that they tend to discriminate listeners of unorthodox, low-mainstream artists. This is foremost due to the scarcity of usage data of low-mainstream music as music consumption patterns are biased towards popular artists. Thus, the objective of our work is to provide a novel approach for modeling artist preferences of users with different music consumption patterns and listening habits.
BibTeX
@misc{kowald2019modeling,
  author = {Dominik Kowald and Elisabeth Lex and Markus Schedl},
  title = {{Modeling Artist Preferences of Users with Different Music Consumption Patterns for Fair Music Recommendations}},
  year = {2019},
  eprint = {1907.09781},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1907.09781},
}

Loading related content...

Figure