Cross‑Domain Multi‑Objective Estimation and Fusion in Baidu Video Recommendation: Design, Modeling, and System Evolution

The presentation introduces Baidu Video's background and the shift to a unified immersive (vertical swipe) interaction across all video scenes, leveraging Baidu's large model to unify data and recommendation experiences.

The recommendation system aims to solve two core problems: a content selection mechanism that promotes high‑quality items and creators, and an exceptional user consumption experience that drives sustained growth.

A high‑level workflow is described: after audit, resources are stored in a unified forward index; the system performs cross‑domain recall, followed by coarse and fine ranking, multi‑objective fusion, and finally generates a video list using diversity, sequence modeling, and traffic allocation strategies.

Multi‑objective design starts with basic physical targets (e.g., playback duration, interactions) and extends to higher‑level satisfaction modeling, including search‑domain signals and long‑term value (LTV) attribution, where future consumption is linked back to the current video.

Long‑term value modeling treats future video consumption as an extension of the current video, attributing future satisfied views to the present item based on relevance and temporal distance.

Cross‑domain multi‑objective modeling addresses two challenges: transferring signals across heterogeneous domains and mitigating negative transfer among objectives. Baidu adopts a gated hierarchical architecture consisting of a common personalization network, a cross‑domain information extraction (GCG) network, and domain‑specific multi‑objective networks.

The multi‑objective fusion has evolved from manual prior‑knowledge fusion to Learning‑to‑Rank (LTR), then to a value‑based model, and currently to an Evolution Strategy (ES) approach that optimizes a reward combining session depth, duration, and interaction metrics.

Finally, the talk invites interested engineers to join Baidu's team (HR email: [email protected]) and thanks the audience.