Spotify Research presents a unified multi-task learning (MTL) framework that jointly optimizes podcast ads and promotions within a single model, replacing previously siloed systems. The model shares a common encoder for user, content, and context representations, with task-specific prediction heads for streams, clicks, likes, and follows. Two key techniques—directional loss masking and source-balanced sampling—prevent negative transfer between channels. Offline ablations show the 5-task MTL model achieves +4.5% Promotions AP and +50.2% Ads AP over the promotions-only baseline. Online A/B tests across 180+ markets showed ~18% higher impression-to-stream rates, ~20% lower cost-per-stream, and ~18% more total streams. Cold-start gains were even larger for emerging creators, with up to ~60% improvement in impression-to-stream rate for low-popularity podcasts. The unified approach also reduces engineering complexity and accelerates new product launches.

8m read timeFrom research.atspotify.com
Post cover image
Table of contents
One Podcast Ecosystem, Many ObjectivesThe Core Idea: A Unified Multi-Objective Podcast ModelFrom Siloed Models to a Joint ArchitectureJoint Ads–Promotions ModelingControlling Transfer Between ChannelsThe Role of Ancillary ObjectivesOffline EvaluationOnline A/B Test ResultsCold-Start Behavior by PopularityKey TakeawaysWhere We’re Heading

Sort: