Cross-Scenario Unified Modeling of User Interests at Billion Scale

TL;DR

This work tackles the fragmentation of user-interest modeling across diverse platform contexts by proposing RED-Rec, an LLM-enhanced hierarchical two-tower recommender that unifies user and item representations across homefeed, search, and ads. It introduces a 2-D dense mixing policy and scenario-aware querying to fuse cross-scenario signals efficiently, enabling near-real-time serving at billion-scale deployments. The authors validate the approach with a million-scale cross-scenario dataset (RED-MMU) and show substantial offline gains and production-level improvements, including a real-world online deployment affecting hundreds of millions of users. The study demonstrates that holistic cross-scenario modeling yields richer user representations, improved recall metrics, and meaningful business impact, establishing a practical pathway for unified personalization on large-scale UGC platforms.

Abstract

User interests on content platforms are inherently diverse, manifesting through complex behavioral patterns across heterogeneous scenarios such as search, feed browsing, and content discovery. Traditional recommendation systems typically prioritize business metric optimization within isolated specific scenarios, neglecting cross-scenario behavioral signals and struggling to integrate advanced techniques like LLMs at billion-scale deployments, which finally limits their ability to capture holistic user interests across platform touchpoints. We propose RED-Rec, an LLM-enhanced hierarchical Recommender Engine for Diversified scenarios, tailored for industry-level content recommendation systems. RED-Rec unifies user interest representations across multiple behavioral contexts by aggregating and synthesizing actions from varied scenarios, resulting in comprehensive item and user modeling. At its core, a two-tower LLM-powered framework enables nuanced, multifaceted representations with deployment efficiency, and a scenario-aware dense mixing and querying policy effectively fuses diverse behavioral signals to capture cross-scenario user intent patterns and express fine-grained, context-specific intents during serving. We validate RED-Rec through online A/B testing on hundreds of millions of users in RedNote through online A/B testing, showing substantial performance gains in both content recommendation and advertisement targeting tasks. We further introduce a million-scale sequential recommendation dataset, RED-MMU, for comprehensive offline training and evaluation. Our work advances unified user modeling, unlocking deeper personalization and fostering more meaningful user engagement in large-scale UGC platforms.

Figures (5)

• Figure 1: From fragmented signals to unified understanding. Users express consistent interests across diverse scenarios (left), generating rich behavioral sequences that span homefeed browsing, search queries, and ad interactions (middle). method synthesizes these cross-scenario signals using llm-powered hierarchical modeling to generate comprehensive user representations for context-aware recommendations (right). Vector graphics; zoom for details.
• Figure 2: Cross-scenario user interactions in the dataset dataset. Real platform interface examples showing user behavior progression across (a) homefeed content consumption, (b) search-based exploration, and (c) advertisement engagement. Each scenario captures diverse interaction types, including clicks, shares, comments, and purchases, illustrating the interconnected nature of cross-scenario interactions on ugc platforms.
• Figure 3: Comprehensive user engagement analytics across scenarios. Multi-faceted analysis of 10k sampled users showing interaction patterns across homefeed, advertisements, and search scenarios. The dashboard presents: (a) user distribution and interaction channels, (b) top content engagement by scenario, (c) behavioral pattern distributions, (d) duration-engagement correlations, (e) temporal activity patterns, (f) content lifecycle analysis, (g-h) cross-scenario user behavior comparisons. Analysis reveals distinct engagement patterns and temporal dynamics across different recommendation contexts.
• Figure 4: Overall framework of method.method employs a two-tower hierarchical architecture with multimodal item encoding and cross-scenario user modeling. The item encoder processes textual content (title, tags, OCR) and visual signals through unified embeddings. The user encoder incorporates a 2-D dense mixing policy to balance interactions across homefeed, advertisements, and search scenarios, followed by scenario-aware transformer blocks that capture evolving user interests. During training, positive and negative samples are drawn from a sampling pool to optimize contrastive objectives end-to-end. method generates unified representations suitable for cross-scenario recommendation tasks.
• Figure 5: Cross-scenario benefits and model scaling analysis. (a) Performance improvements from cross-scenario modeling using method, where different colored lines represent various scenario combinations (homefeed, ads, search) and the shaded regions highlight the performance gains achieved through unified cross-scenario learning compared to single-scenario baselines. (b) Scaling laws for model size versus Hit Rate performance, showing consistent improvements with increased parameter count across both LLaMA and Qwen model families, with corresponding serving throughput (samples per second) trade-offs indicated by the secondary axis.