Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Let's Get It Started: Fostering the Discoverability of New Releases on Deezer

Léa Briand, Théo Bontempelli, Walid Bendada, Mathieu Morlon, François Rigaud, Benjamin Chapus, Thomas Bouabça, Guillaume Salha-Galvan

TL;DR

The paper addresses the cold-start problem for new music releases on Deezer, which hampers discoverability due to limited prior usage data. It introduces CF-Cold-Start, a 3-layer neural network that predicts future item embeddings for new albums using metadata and initial usage, updated every four hours, and employs approximate nearest-neighbor search to generate personalized recommendations. It further extends this with TS-CF-Cold-Start, a contextual bandit variant using Thompson Sampling to inject exploration and diversify recommendations, with embeddings as priors and online inference via ONNX. Online A/B tests show notable improvements in exposure and engagement for new releases, indicating that shifting from editorial-only to personalized, dynamically updated ranking enhances discoverability, while raising considerations for fairness and suggesting future work to separate discovery from unmissable content and to integrate releases with other personalized features like Flow.

Abstract

This paper presents our recent initiatives to foster the discoverability of new releases on the music streaming service Deezer. After introducing our search and recommendation features dedicated to new releases, we outline our shift from editorial to personalized release suggestions using cold start embeddings and contextual bandits. Backed by online experiments, we discuss the advantages of this shift in terms of recommendation quality and exposure of new releases on the service.

Let's Get It Started: Fostering the Discoverability of New Releases on Deezer

TL;DR

The paper addresses the cold-start problem for new music releases on Deezer, which hampers discoverability due to limited prior usage data. It introduces CF-Cold-Start, a 3-layer neural network that predicts future item embeddings for new albums using metadata and initial usage, updated every four hours, and employs approximate nearest-neighbor search to generate personalized recommendations. It further extends this with TS-CF-Cold-Start, a contextual bandit variant using Thompson Sampling to inject exploration and diversify recommendations, with embeddings as priors and online inference via ONNX. Online A/B tests show notable improvements in exposure and engagement for new releases, indicating that shifting from editorial-only to personalized, dynamically updated ranking enhances discoverability, while raising considerations for fairness and suggesting future work to separate discovery from unmissable content and to integrate releases with other personalized features like Flow.

Abstract

This paper presents our recent initiatives to foster the discoverability of new releases on the music streaming service Deezer. After introducing our search and recommendation features dedicated to new releases, we outline our shift from editorial to personalized release suggestions using cold start embeddings and contextual bandits. Backed by online experiments, we discuss the advantages of this shift in terms of recommendation quality and exposure of new releases on the service.
Paper Structure (7 sections, 3 figures)

This paper contains 7 sections, 3 figures.

Table of Contents

  1. Introduction
  2. Listening to New Releases on Deezer
  3. Improving the Discoverability of New Releases
  4. Cold Start Embedding Representations of New Releases
  5. Carousel Personalization with Contextual Bandits
  6. Online Experimental Evaluation on Deezer
  7. Conclusion and Future Work

Figures (3)

  • Figure 1: Interface of the "New releases for you" carousel on the website version of Deezer, recommending a personalized shortlist of recently released albums to each user.
  • Figure 2: Overview of the CF-Cold-Start new album recommendation framework from Section \ref{['coldstartsection']}, available for online requests in our production environment on Deezer.
  • Figure 3: Overview of the TS-CF-Cold-Start new album recommendation framework from Section \ref{['bandit']}, available for online requests in our production environment on Deezer.