Multi-modal entity alignment aims to identify equivalent entities between two multi-modal Knowledge graphs by integrating multi-modal data, such as images and text, to enrich the semantic representations of entities. However, existing methods may overlook the structural contextual information within each modality, making them vulnerable to interference from shallow features. To address these challenges, we propose MyGram, a \textbf{m}odalit\textbf{y}-aware \textbf{gra}ph transformer with global distribution for \textbf{m}ulti-modal entity alignment. Specifically, we develop a modality diffusion learning module to capture deep structural contextual information within modalities and enable fine-grained multi-modal fusion. In addition, we introduce a Gram Loss that acts as a regularization constraint by minimizing the volume of a 4-dimensional parallelotope formed by multi-modal features, thereby achieving global distribution consistency across modalities. We conduct experiments on five public datasets. Results show that MyGram outperforms baseline models, achieving a maximum improvement of 4.8\% in Hits@1 on FBDB15K, 9.9\% on FBYG15K, and 4.3\% on DBP15K.

Inductive reasoning aims to infer missing knowledge for unseen entities and relations. Existing methods exhibit limited generalization capabilities due to their dependence on localized structural patterns and inadequate handling of graph imbalance. To address these challenges, we propose a novel Hi erarchical Mod eling with Graph Perturbation-Enhanced Network (HiMod), which effectively integrates hierarchical relation modeling with a dynamic perturbation mechanism to enhance the generalization ability of inductive reasoning models. HiMod leverages a hierarchical relation modeling mechanism that maps specific relations to higher-level general concepts within a global semantic framework. This allows for capturing semantic commonalities across relations, enabling robust reasoning for unseen queries. Simultaneously, a dynamic perturbation mechanism is introduced to adjust perturbation strength based on node importance and graph sparsity, facilitating deeper exploration of the latent semantic space and mitigating the effects of graph imbalance. Extensive experiments on three benchmark inductive knowledge graph reasoning datasets demonstrate that HiMod achieves the most significant MRR improvements among the four split versions, with 11.17% on WN18RR, 4.61% on FB15K-237, and 5.47% on NELL-995. Our code is available at https://github.com/HubuKG/HiMod .

Concept prerequisite relation learning is used to identify dependency relations between knowledge concepts, which helps learners choose effective learning paths. Currently, most of the mainstream methods utilise deep learning algorithms to capture the prerequisite relations between concepts through supervised or semi-supervised learning. However, these methods are highly dependent on labelled data, which is scarce and costly to annotate in reality. To address this problem, we propose a framework calledWeaklySupervisedEnhancedConceptPrerequisiteRelationLearning (WSECPRL). Specifically, we first generate an enhanced concept pseudo-relation graph without labeled data using the pre-trained language model and the large knowledge base as auxiliary information. Second, we propose an improved variational graph auto-encoder model to correctly determine the concept prerequisite relations. We incorporate a multi-head attention mechanism to enhance the representation learning capability of weakly supervised learning. The model reconstructs a directed graph into multiple undirected graphs by splitting the adjacency matrix and determines the direction of the concept prerequisite relation based on the strength of the dependency relation between concepts. Finally, experimental results on several publicly available datasets demonstrate the effectiveness of our proposed framework, with WSECPRL outperforming existing baseline models in terms of F1 scores and AUC.