Fusion of Graph-Based Ensembles Using Fuzzy Integrals for Persian Learning to Rank

The rapid growth of Persian-language Web content has created an urgent need for effective Learning to Rank (LTR) systems. However, developing such systems for Persian presents significant challenges, including Persian’s morphological complexity, its unique Web link topology, and the scarcity of annotated resources relative to dominant languages such as English. Existing methods often rely on hand-crafted features, apply graph techniques in isolation, or employ simple ensemble fusion, lacking a unified framework to integrate structural semantics and handle model uncertainty. This paper proposes PersianRank, a novel LTR framework that bridges these gaps by modeling query-document relationships as a weighted bipartite graph to extract graph-based structural features. It applies deep neural networks to predict these graph features for unseen query-document pairs, enabling feature augmentation during the testing phase. A diverse ensemble of base rankers is then trained on the enriched feature set, and their predictions are fused via the Choquet fuzzy integral, which performs a non-linear, uncertainty-aware aggregation. Evaluated on the dotIR dataset, PersianRank achieves state-of-the-art performance, improving P@۱ by ۱۷.۱% relative to the best baseline (MDPRank) and demonstrating notable top-rank accuracy through effective graph-based feature space expansion and uncertainty-aware model fusion. The framework is especially effective for top-rank retrieval, a critical aspect in user-centric search systems.

کلیدواژه ها:

Learning to rank ، Persian Web Retrieval ، Graph-Based Features ، Ensemble Fusion ، Fuzzy Integral ، Uncertainty-Aware Ranking