PLANET: A New Framework for Multimodal Graph Foundation Models Based on Divide-and-Conquer Strategy

PLANET is a multimodal graph foundation model framework accepted by ICML 2026. It uses a divide-and-conquer strategy to solve the core challenges of integrating graph neural networks (GNNs) with multimodal learning, offering new ideas for unified representation learning of complex relational data. This article will cover its background, core strategies, technical implementation, experimental validation, application prospects, and future directions.

In real-world complex systems, data often exists in graph forms (e.g., social networks, molecular structures, knowledge graphs), and nodes/edges carry multimodal information (text, images, time-series signals). Traditional GNNs excel at capturing topological structures but have limited ability to handle heterogeneous multimodal data; multimodal foundation models perform well in unimodal tasks but struggle to adapt to the non-Euclidean nature of graph structures, leading to a separation between "structural" and "semantic" aspects and limiting generalization capabilities.

The core of the PLANET framework is a divide-and-conquer strategy, which decomposes multimodal graph learning into subproblems and then integrates them:

1. Intra-modal divide-and-conquer: Train independent encoders for each modality to map to a latent space, avoiding interference from early fusion;
2. Structure-semantic divide-and-conquer: Parallel branches use graph attention mechanisms to capture topological patterns and Transformers to extract semantic features, respectively;
3. Hierarchical divide-and-conquer: A hierarchical aggregation strategy to capture node-level, subgraph-level, and full-graph-level features simultaneously.

PLANET adopts a modular design, with core components including:

• Multimodal encoders: Support text (BERT/RoBERTa), images (ViT/CLIP), and numerical features (MLP), with a unified interface for easy expansion;
• Graph structure learning module: GAT variant + cross-modal attention to enable interaction between multimodal representations;
• Divide-and-conquer fusion module: Supports multiple fusion strategies and adaptively selects paths via a gating mechanism;
• Pretraining and fine-tuning framework: Provides self-supervised task scripts (masked node prediction, edge prediction, etc.) and domain adaptation tools.

In the paper accepted by ICML 2026, PLANET was validated on multiple benchmark datasets:

• Node classification: Outperforms traditional GNNs on datasets like ogbn-arxiv, effectively using semantic information to improve accuracy;
• Link prediction: Joint structure-semantic representation reduces false positives and accurately models heterogeneous relationships;
• Cross-modal retrieval: After pretraining, it has zero-shot transfer capability, solving the cold-start problem.

PLANET can be applied in:

• Recommendation systems: Model user-item bipartite graphs + multimodal information to improve recommendation quality;
• Drug discovery: Process molecular graphs + chemical properties/spectra/text to accelerate new drug development;
• Knowledge graph enhancement: Integrate multimodal information of entities to enrich knowledge representation;
• Scientific computing: Adapt to graph-structured multimodal data in materials science and bioinformatics.

Insights: The divide-and-conquer strategy is more effective than end-to-end approaches, with advantages including reduced optimization difficulty, enhanced interpretability, and improved flexibility; the challenge lies in balancing module independence and information interaction. Future directions: Large-scale pretraining, efficient cross-modal alignment, causal reasoning capabilities, and domain-specific design (e.g., scientific computing, financial risk control).