[Paper Review] Liquid FM: Recommending Music through Viscous Democracy

Paolo Boldi, Corrado Monti+2 more|arXiv (Cornell University)|Mar 30, 2015

Complex Network Analysis Techniques4 references3 citations

TL;DR

Liquid FM proposes a music recommendation system based on viscous democracy, where users delegate trust to others for song recommendations instead of relying on collaborative filtering. It uses a weighted voting mechanism inspired by Katz centrality to compute influence, avoiding filter bubbles by making users consciously choose recommenders.

ABSTRACT

Most modern recommendation systems use the approach of collaborative filtering: users that are believed to behave alike are used to produce recommendations. In this work we describe an application (Liquid FM) taking a completely different approach. Liquid FM is a music recommendation system that makes the user responsible for the recommended items. Suggestions are the result of a voting scheme, employing the idea of viscous democracy. Liquid FM can also be thought of as the first testbed for this voting system. In this paper we outline the design and architecture of the application, both from the theoretical and from the implementation viewpoints.

Motivation & Objective

• To develop a music recommendation system that shifts responsibility for recommendations from algorithms to users by enabling trust delegation.
• To test the viscous democracy framework in a real-world application, particularly in the context of social networks and music discovery.
• To evaluate whether user-driven delegation reduces filter bubbles compared to traditional collaborative filtering approaches.
• To implement and deploy a functional Facebook-based application that collects real user data for viscous democracy evaluation.
• To explore the scalability and usability of viscous democracy in a practical, large-scale social recommendation setting.

Proposed method

• Users delegate one trusted friend (from their Facebook friend list) to act as a recommender for a specific music genre, forming a directed delegation graph.
• Each user can vote for up to three favorite songs, creating a bipartite voting graph between users and songs.
• The system computes a weighted influence score for each song using a viscous democracy algorithm based on Katz centrality, where trust decays over delegation chains via a decay factor α.
• The final recommendation score for a song s is computed as c(s) = k·p(s) + r(s), combining indirect influence (p(s)) and direct votes (r(s)), with k calibrated to balance both components.
• The system employs Redis for caching computed recommendation scores to improve performance and supports lazy loading for scalability.
• External integration with MusicBrainz ensures song validity and enables fast, local database access via a self-updating virtual machine.

Experimental results

Research questions

• RQ1 Can viscous democracy be effectively implemented in a real-world social media application for music recommendation?
• RQ2 Does user-driven delegation in viscous democracy reduce the risk of filter bubbles compared to traditional collaborative filtering?
• RQ3 How does the decay factor α in the viscous democracy model affect the distribution of influence and recommendation quality?
• RQ4 To what extent can user-generated delegation and voting data be used to produce meaningful and diverse music recommendations?
• RQ5 What are the practical challenges in scaling a viscous democracy-based recommendation system in a real social network?

Key findings

• The Liquid FM system successfully implements viscous democracy in a live Facebook application, demonstrating the feasibility of user-delegated trust in recommendation systems.
• The use of a decay factor α in the influence propagation ensures that indirect trust diminishes over longer delegation chains, preventing over-influence from distant users.
• The hybrid scoring model c(s) = k·p(s) + r(s) effectively balances direct votes and indirect influence, producing a ranked list of recommendations.
• The system leverages Redis for efficient caching of computed recommendations, enabling fast response times and supporting lazy evaluation for scalability.
• Initial data collection is ongoing, with delegation graphs currently containing only a few tens of delegations, indicating the need for viral growth to enable robust evaluation.
• External integration with MusicBrainz ensures data quality and enables fast, local access to a comprehensive music database for song validation and lookup.