DANCE: Dynamic, Available, Neighbor-gated Condensation for Federated Text-Attributed Graphs

Federated graph learning (FGL) enables collaborative training on graph data across multiple clients. With the rise of large language models (LLMs), textual attributes in FGL graphs are gaining attention. Text-attributed graph federated learning (TAG-FGL) improves FGL by explicitly leveraging LLMs to process and integrate these textual features. However, current TAG-FGL methods face three main challenges: \textbf{(1) Overhead.} LLMs for processing long texts incur high token and computation costs. To make TAG-FGL practical, we introduce graph condensation (GC) to reduce computation load, but this choice also brings new issues. \textbf{(2) Suboptimal.} To reduce LLM overhead, we introduce GC into TAG-FGL by compressing multi-hop texts/neighborhoods into a condensed core with fixed LLM surrogates (summaries/embeddings). However, this one-shot condensation is often not client-adaptive, leading to suboptimal performance. \textbf{(3) Interpretability.} LLM-based condensation further introduces a black-box bottleneck: summaries lack faithful attribution and clear grounding to specific source spans, making local inspection and auditing difficult. To address the above issues, we propose \textbf{DANCE}, a new TAG-FGL paradigm with GC. To improve \textbf{suboptimal} performance, DANCE performs round-wise, model-in-the-loop condensation refresh using the latest global model. To enhance \textbf{interpretability}, DANCE preserves provenance by storing locally inspectable evidence packs that trace predictions to selected neighbors and source text spans. Across 8 TAG datasets, DANCE improves accuracy by \textbf{2.33\%} at an \textbf{8\%} condensation ratio, with \textbf{33.42\%} fewer tokens per condensed node than TAG-FGL baselines.

Many real-world datasets can be represented as graphs whose nodes contain both structural connections and rich textual descriptions, such as citation networks, social platforms, and product-review systems. Federated graph learning allows multiple data owners to train a shared graph model without directly sharing their private local graphs or raw texts. However, when graph nodes contain long texts, existing methods often need to repeatedly process large amounts of textual information with language models, which can be computationally expensive. Moreover, once the textual information is compressed before training, it may become outdated as the federated model changes over communication rounds. Another practical concern is interpretability: users may need to know which local texts or neighboring nodes support a model prediction, especially in privacy-sensitive or regulated scenarios. This paper proposes \textbf{DANCE}, a dynamic condensation framework for federated learning on text-attributed graphs. Instead of using the full local graph and all node texts, DANCE selects a compact set of representative nodes, identifies the most informative neighboring evidence, and condenses relevant text chunks into shorter evidence summaries. As federated training progresses, DANCE periodically refreshes this condensed information using the latest global model, so that the retained evidence remains aligned with the evolving learning objective. It also reconstructs a lightweight graph among the selected nodes, allowing efficient message passing without reverting to the original large graph. A key feature of DANCE is that it keeps human-readable evidence traces locally on each client. These traces record which neighbors and text spans contribute to each condensed node, making the model's decisions easier to inspect while still avoiding the transmission of raw private data. Experiments on multiple text-attributed graph datasets show that DANCE improves node classification accuracy over strong federated graph learning baselines while substantially reducing token-processing costs. Overall, DANCE provides a practical way to make federated learning on text-rich graphs more efficient, adaptive, and auditable.