Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Contextualizing Spotify's Audiobook List Recommendations with Descriptive Shelves

Gustavo Penha, Alice Wang, Martin Achenbach, Kristen Sheets, Sahitya Mantravadi, Remi Galvez, Nico Guetta-Jeanrenaud, Divya Narayanan, Ofeliya Kalaydzhyan, Hugues Bouchard

TL;DR

The paper addresses cold-start constraints in Spotify's audiobook catalog by introducing descriptive shelves generated from LLM-enriched metadata grounded in a domain taxonomy. The approach entails a pipeline that first enriches item metadata with descriptors via LLMs, then generates per-user descriptive shelves by ranking and diversifying descriptors and by ranking/filtering items. Two A/B tests reveal that descriptive shelves can boost discovery and engagement when personalization and placement are properly signaled, though the initial test suffered engagement issues due to UI limitations. The work demonstrates practical improvements for user exploration and for content creators, and outlines future work on richer explanations and deeper personalization in shelf construction.

Abstract

In this paper, we propose a pipeline to generate contextualized list recommendations with descriptive shelves in the domain of audiobooks. By creating several shelves for topics the user has an affinity to, e.g. Uplifting Women's Fiction, we can help them explore their recommendations according to their interests and at the same time recommend a diverse set of items. To do so, we use Large Language Models (LLMs) to enrich each item's metadata based on a taxonomy created for this domain. Then we create diverse descriptive shelves for each user. A/B tests show improvements in user engagement and audiobook discovery metrics, demonstrating benefits for users and content creators.

Contextualizing Spotify's Audiobook List Recommendations with Descriptive Shelves

TL;DR

The paper addresses cold-start constraints in Spotify's audiobook catalog by introducing descriptive shelves generated from LLM-enriched metadata grounded in a domain taxonomy. The approach entails a pipeline that first enriches item metadata with descriptors via LLMs, then generates per-user descriptive shelves by ranking and diversifying descriptors and by ranking/filtering items. Two A/B tests reveal that descriptive shelves can boost discovery and engagement when personalization and placement are properly signaled, though the initial test suffered engagement issues due to UI limitations. The work demonstrates practical improvements for user exploration and for content creators, and outlines future work on richer explanations and deeper personalization in shelf construction.

Abstract

In this paper, we propose a pipeline to generate contextualized list recommendations with descriptive shelves in the domain of audiobooks. By creating several shelves for topics the user has an affinity to, e.g. Uplifting Women's Fiction, we can help them explore their recommendations according to their interests and at the same time recommend a diverse set of items. To do so, we use Large Language Models (LLMs) to enrich each item's metadata based on a taxonomy created for this domain. Then we create diverse descriptive shelves for each user. A/B tests show improvements in user engagement and audiobook discovery metrics, demonstrating benefits for users and content creators.

Paper Structure

This paper contains 11 sections, 1 figure, 1 table.

Table of Contents

  1. Introduction
  2. Related Work
  3. Descriptive Shelves
  4. Descriptor Generation
  5. Shelf Generation
  6. Descriptor Ranking and Diversification
  7. Item Ranking and Filtering
  8. Results & Analysis
  9. First A/B test
  10. Second A/B test
  11. Conclusion

Figures (1)

  • Figure 1: Shelf generation pipeline: from a set of top-K recommendations and their respective descriptors, we rank the distinct set of descriptors of the recommendations to be used as shelf titles for the user based on their affinities and diversify the list by removing descriptors that are too similar to each other. Then we rank and filter the set of candidate items for each of the distinct descriptors, and finally, we decorate the shelves and display the top-N to the user.