Spotify Podcast: multitask learning for cold-start ads and promotion¶

Fidelity: core mechanism. Runs a shared low-rank multitask representation and cold-item bucketing.

Paper information¶

Field	Details
Paper	arXiv 2601.02306
Organization	Spotify
First public release	2026-01-05 (arXiv v1)
Authors' code	The paper did not supply official or author code when checked on 2026-08-09.
Adapter / method	`podcast-mtl`
Local reproduction code	`src/auto_research/reproductions/podcast_mtl/`

Original paper¶

Motivation and method¶

New podcasts lack ad click and stream data. The paper jointly trains ads and organic promotion, sharing user, content, and context representations to transfer signal from rich tasks to cold-start tasks. Imbalance handling and gradient balancing limit negative transfer.

flowchart LR
  X["User, content, and context"] --> S["Shared representation"]
  S --> A["Ads objective"]
  S --> P["Promotion objective"]
  A --> B["Gradient balance"]
  P --> B
  B --> C["Cold-start scoring"]

Key figure from the paper¶

Figure 1: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.

Core equations¶

\[ \mathcal L=\alpha\mathcal L_{\mathrm{ads}}+ (1-\alpha)\mathcal L_{\mathrm{promotion}},\qquad h=W_sx,\quad \hat y_t=W_th. \]

Paper results¶

Effective cost per stream decreases by up to 22%; podcast stream rate rises 18% to 24%, with stronger gains on cold content.

Local reproduction¶

Local comparison: The baseline is a single-task popularity/content ranker. The joint organic/promotion treatment changes NDCG@10 from 0.03540→0.02810, or -20.63% relative. Local negative transfer occurs.

This negative result is retained. The proxy tasks and weights need further evolve work. See the metrics.

auto-research reproduce --paper podcast-mtl --dataset-dir data --seed 42

Reproduction boundaries¶

MovieLens movies and genres stand in for podcasts and ad features. Spotify impressions, bids, and exposure-calibration logs are unavailable.