Hyperintent improves efficiency in reasoning on changing text graphs
HyperIntent: Tracking Intent Shifts for Efficient Reasoning on Dynamic Textual Graphs
Graphics
Summary
Many online interactions happen over time and create changing connections between pieces of text. The authors found that these changes often repeat themselves, making it hard for language models to focus on important updates. They created HyperIntent, a method that notices when the main goal or intention behind these changes actually shifts and updates reasoning only then. This helps the language model work faster and use less data while still making good predictions about future connections.
What this means in practice
- •For web interaction developers: Identify key changes in user behavior patterns on websites by tracking shifts in interaction intents over time.
- •For recommendation system engineers: Improve predictions of future user-item interactions in dynamic recommendation graphs by updating models only when underlying user intent shifts.
Authors
Yilin Yao, Bin Lu, Xinhui Li, Meng Jin, Xinbing Wang
Abstract
Dynamic web interactions continuously evolve as new edges arrive, offering large language models (LLMs) great potential for reasoning over edge-level temporal semantics. However, continuously arriving edges are often repetitive, obscuring the key structures that truly reflect temporal changes and limiting the efficiency of LLMs. A key insight is that beneath the dense stream of edge arrivals lies a sparse evolution of temporal intent, enabling LLMs to be responsively triggered only when such intent shifts, while otherwise inheriting the previous state. To this end, we propose a hypergraph-based framework named HyperIntent, which enhances dynamic graph reasoning by parsing differentiated intents, quantifying their temporal discrepancies, and performing shift-triggered semantic updating. Experimental results on destination node retrieval and future link prediction demonstrate that HyperIntent consistently achieves the best overall performance, with average relative Hits@1 gains of 34.0% and 25.4% over vanilla TGAT and DyGFormer, respectively. Meanwhile, HyperIntent reduces LLM token consumption by 54.4%-95.9% across six datasets while achieving leading predictive performance. In-depth studies further confirm the necessity of shift triggering for identifying structural and semantic temporal changes. The code will be made publicly available upon completion of the review process.