Bridging the Semantic-Collaborative Gap: An Asymmetric Graph Architecture for Cold-Start Item Recommendation

2026-06-06 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Expert, extended

Summary

Tubi's production retrieval system introduces Shallow-RHS, an asymmetric graph architecture designed to solve cold-start item recommendation for new content and devices. This model formulates cold-start as an inductive graph-completion problem on a temporal bipartite device–content graph, leveraging the Kumo GNN platform. The Left-Hand Side (LHS) device tower uses watch-history message passing to capture collaborative signals, while the Right-Hand Side (RHS) content tower is intentionally shallow, encoding content solely from intrinsic features like metadata and LLM-based semantic embeddings (e.g., from OpenAI). This forces the content encoder to map intrinsic features into a collaborative-filtering-aware embedding space. For device cold-start, cohort-based embeddings are constructed from demographic features. Large-scale online A/B tests at Tubi demonstrated consistent relative improvements, including a +0.42% global Total View Time (TVT) lift and increased cold-title promotion speed by 13%.

Key takeaway

For Machine Learning Engineers building recommendation systems facing cold-start challenges, you should consider adopting an asymmetric graph architecture like Shallow-RHS. This approach allows you to generate immediate, collaborative-filtering-aware embeddings for new content using only intrinsic features, and for new devices via demographic cohorts. This strategy significantly improves engagement metrics and content promotion speed, enabling effective recommendations even without historical interaction data.

Key insights

An asymmetric graph architecture can bridge the semantic-collaborative gap for cold-start recommendations by aligning intrinsic features with behavioral signals.

Principles

• Content embeddings must derive solely from intrinsic features for cold-start.
• Asymmetric model design forces feature encoders into a CF-aware space.
• Missing graph representations can be reconstructed from side-information priors.

Method

Formulate cold-start as an inductive graph-completion problem on a temporal bipartite device-content graph. Use an asymmetric two-tower architecture (Shallow-RHS) where the content tower is shallow and feature-only, and the device tower uses message passing. Train with temporal softmax link prediction loss.

In practice

• Generate cold content embeddings using a learned feature-only encoder.
• Use approximate nearest-neighbor search for warm surrogate neighbors.
• Assign new devices to demographic cohorts for initial embeddings.

Topics

• Cold-Start Recommendation
• Graph Neural Networks
• Collaborative Filtering
• Two-Tower Architecture
• Temporal Link Prediction
• Tubi

Best for: AI Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Architect

Related on AIssential

• Bridging the Semantic-Collaborative Gap: An Asymmetric Graph Architecture for Cold-Start Item Recommendation — Shallow-RHS uses an asymmetric graph architecture to generate cold-start item embeddings from intrinsic features, bridging semantic and collaborative signals.
• Cracking the Cold Start Problem — Hybrid recommender systems combining collaborative filtering, embeddings, and bandit learning improve cold start and fairness.
• How Recommendation System Works on Youtube — YouTube's recommendation system uses a two-stage deep learning approach to personalize content at massive scale.
• SilverTorch: Index as Model — A New Retrieval Paradigm for Recommendation Systems — Unifying recommendation retrieval into a single neural network significantly boosts efficiency, quality, and development velocity.
• Inside ByteDance’s Monolith: The Engine Powering Smarter, Faster Content Feeds — Monolith optimizes real-time recommendation systems via collisionless embeddings and continuous online training.
• HSTU From Scratch in PyTorch - A complete Walkthrough — HSTU fuses item and action embeddings with time-aware attention for robust sequential recommendation.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.