Mostra: A Flexible Balancing Framework to Trade-off User, Artist and Platform Objectives for Music Sequencing

Abstract

We consider the task of sequencing tracks on music streaming platforms where the goal is to maximise not only user satisfaction, but also artist-and platform-centric objectives, needed to ensure long-term health and sustainability of the platform. Grounding the work across four objectives: Sat, Discovery, Exposure and Boost, we highlight the need and the potential to trade-off performance across these objectives, and propose Mostra, a Set Transformerbased encoder-decoder architecture equipped with submodular multi-objective beam search decoding. The proposed model affords system designers the power to balance multiple goals, and dynamically control the impact on one objective to satisfy other objectives. Through extensive experiments on data from a large-scale music streaming platform, we present insights on the trade-offs that exist across different objectives, and demonstrate that the proposed framework leads to a superior, just-in-time balancing across the various metrics of interest. CCS CONCEPTS• Information systems → Recommender systems.

Maximizing Submodular Functions for Recommendation in the Presence of Biases

“…Selecting the first item increases the utility of the selection g(1)−g(0) = 1. Further selecting the second item increases the utility by a smaller amount g(2)−g(1) ≈ 0.414, and further yet, selecting the third item increases the utility by an even smaller amount g(3) − g(2) ≈ 0.318. function considered by [14] when g 1 (x) = x and g 2 (x) = • • • = g m (x) = √ x for all x. Moreover, in Section A.1, we show that F captures functions used by [2,66].…”

“…We consider the following family of submodular functions that are studied in the context of content recommendation [14,23,57,66] and web search [2]. Definition 3.1 (A family of submodular functions).…”

“…Setting g j (x) = log(1 + x) for all x and j in the above definition, we get the submodular function tested by [57] on Amazon Music. With g 1 (x) = x and g j (x) = √ x for all x and j = 1, we get the submodular function used by [14] to measure the quality of song playlists. Further examples appear in Section A.1.…”

“…Subset selection arises in many web-based applications including different types of content recommendation systems [14,23,57,66] and search engines [2]. Generally speaking, in all of these applications, given a set of n items (e.g., posts, products, videos, or websites), the task is to select a subset S of size k that is the most valuable for the user.…”

“…They use the submodular function F (S) = m j=1 log 1 + i∈S W ij to capture the value of a set S of song recommendations. Another example is [14] who, roughly speaking, set W ij to encode the utility of song i for stakeholder j. They use F (S) = i∈S W i1 + m j=2 i∈S W ij to capture the value of a playlist S on Spotify.…”