Ocean4Rec: Offline LLM Profiles for VOD Reranking
The Ocean4Rec system introduces an offline LLM to create OCEAN personality profiles for video-on-demand content, allowing for reranking at the moment of request without needing live LLM interactions. This method categorizes items across five dimensions—Openness, Conscientiousness, Extraversion, Agreeableness, and Neuroticism—using metadata, while user profiles are developed through a time-decayed aggregation of recent clicks and deep links. When a request is made, Ocean4Rec combines pre-calculated item and user profiles with base recommender scores and catalog recency for numerical reranking. This process streamlines operations by eliminating repeated prompt construction and other tasks, enhancing throughput, managing tail latency, and ensuring reliable performance in high-demand latency-sensitive environments. The research is available on arXiv under ID 2605.27429.
Key facts
- Ocean4Rec uses an LLM offline to generate OCEAN profiles for VOD items.
- Items are scored on Openness, Conscientiousness, Extraversion, Agreeableness, Neuroticism.
- User profiles are built by time-decayed aggregation of recent clicks and deep-links.
- Request-time reranking joins precomputed profiles, base scores, and catalog recency.
- No LLM calls are made at request time.
- Aims to improve throughput, tail-latency, capacity isolation, and predictability.
- Published on arXiv with ID 2605.27429.
- Designed for high-volume latency-sensitive VOD services.
Entities
Institutions
- arXiv