Abstract
Learning to Rank (LTR) is fundamental to web search, yet integrating local text semantics with global graph-based relationships at scale remains challenging. We present MPGraf, a modular framework that combines Transformers and Graph Neural Networks (GNNs). MPGraf utilizes a "link rippling" strategy for graph construction and a "surgical" fine-tuning approach to mitigate distribution shifts. Evaluated on both public benchmarks and large-scale commercial search data, MPGraf demonstrates significant improvements in ranking performance, particularly for long-tail queries.
Introduction
In modern search engines, Learning to Rank (LTR) determines the relevance of webpages to user queries. Two dominant architectures excel in different areas: Transformers are adept at capturing fine-grained query-page text interactions, while Graph Neural Networks (GNNs) excel at modeling broader contextual relationships within query-page bipartite graphs. However, combining these approaches at "web-scale" is difficult due to differing data structures, computational overhead, and distribution shifts between pre-training and target datasets.
Methodology: The MPGraf Framework
MPGraf addresses these challenges through a three-stage modular pipeline.
1. Graph Construction with Link Rippling
To handle sparse interaction data, MPGraf employs "link rippling." This involves "Query-centered Expanding Ripples" to identify high-relevance pairs and "Webpage-centered Shrinking Ripples" for effective negative sampling. This ensures a robust graph structure even with limited human annotations.
2. Hybrid Graphformer Architecture
The framework is modular, supporting two integration strategies:
- Stacking Graphformer: GNN outputs are fed directly into a Transformer.
- Parallelizing Graphformer: GNN and Transformer modules run concurrently, with their features fused in later layers. This allows the model to simultaneously leverage local semantic content and global graph context.
3. Surgical Fine-tuning
To combat distribution shifts, a multi-stage "surgical" fine-tuning strategy is used. By selectively freezing and unfreezing modules, the model preserves general knowledge from pre-training while adapting to specific downstream tasks.
Experimental Evaluation
MPGraf was evaluated on standard benchmarks (MSLR-Web30K, MQ2007, MQ2008) and a large-scale commercial dataset from Baidu (15,000 queries, 770,000+ pairs).
Key Results
- Offline Performance: MPGraf consistently outperformed state-of-the-art baselines like XGBoost and LightGBM. The "Stacking" variant improved NDCG@10 by up to 1.74% on commercial data.
- Online Impact: Live deployment on a major search engine showed significant gains, particularly for "long-tail" queries (up to 7.50% improvement in GSB).
- Data Efficiency: The model remained robust even when labeled data was reduced to 5%, demonstrating its ability to "borrow" information across the graph.
Conclusion
MPGraf successfully bridges the gap between Transformers and GNNs for web-scale ranking. While global graph construction remains an engineering challenge, the modular architecture and surgical fine-tuning provide a scalable path forward. Future work will explore integrating Large Language Models (LLMs) to further enhance semantic understanding within this framework.
Summary updated