Graph Anomaly Detection (GAD) aims to identify nodes deviating from normal behavior patterns in graphs. However, prevailing graph autoencoder (GAE)-based GAD methods face challenges including class imbalance, homophily trap and efficiency issues. Moreover, revisiting the GAE-based framework reveals that most existing approaches overlook data quality and treat all data indiscriminately. While curriculum learning has demonstrated low-quality data is detrimental to model training, our preliminary experiments verify that removing such data from training sets can enhance model performance and significantly reduce training time. Building on this insight, we design a dual curriculum learning strategy to optimize the traditional GAE-based anomaly detection framework, addressing the aforementioned challenges. Specifically, we propose a non-learning local inconsistency analysis method to calculate node difficulty scores, then use pacing functions to partition curricula based on these scores. In the first training stage, we pre-train the graph encoder with these curricula to capture simple yet typical graph features. In the second stage, an anti-curriculum strategy trains a new decoder to enhance sensitivity to anomalous features, before performing anomaly detection on the trained GAE framework. Experiments on five real-world datasets demonstrate our method's advanced detection performance and higher efficiency.
Multi-modal crisis event detection is a critical yet challenging task due to its reliance on (i) rich external knowledge and contextual awareness, (ii) deep reasoning across sparse, subtle, and often implicit clues, and (iii) the high variability and diversity inherent in real-world crisis scenarios. To address these challenges, we propose a novel multi-modal detection framework, Crisis-Gated Expert Fusion with Attention-Assembled Reasoning (CG-MoEAR), which integrates the complementary strengths of four core modules. First, a Contextual Cognitive Chain Enhancer, built on large vision-language models, leverages multi-modal prompts to integrate external knowledge and improve situational understanding. Second, a Crisis-Aware Heterogeneous Dual-Stream Encoder combines a modality-consistent CLIP backbone with modality-specific encoders, along with a learnable memory module that captures latent crisis-domain semantics. Third, we design an Attentional-Assembled Reasoning Network that performs multi-hop inference over structured semantic representations using serial-parallel attention mechanisms, effectively uncovering deep, fine-grained clues. Finally, a Crisis-Gated Mixture-of-Experts Aggregator dynamically selects the most relevant expert pathways based on event semantics, achieving adaptive and context-aware decision-making. Extensive experiments on two benchmark datasets—NED and CrisisMMD—demonstrate that the proposed model consistently outperforms existing single-modal and multi-modal baselines, achieving state-of-the-art results in both accuracy and robustness. These results highlight the potential of structured reasoning and expert-aware fusion for complex crisis event detection. This work also provides practical value for real-time emergency response, public opinion monitoring, and disaster management systems.
Link forecasting is a fundamental extrapolation task for temporal knowledge graphs (TKGs), which are collections of real-world facts enriched with temporal information. It aims at inferring future facts from past facts in a TKG. To improve the accuracy in link forecasting, we propose a novel formalism for temporal logical rules named TWC-rules in this paper, which introduces time window constraint to avoid inferring wrong new facts. In addition, we find that existing link forecasting methods either rely on embedding-based models with low interpretability, or apply pipelines probably with error propagation to learn temporal logical rules. To avoid error propagation while keeping high interpretability, we accordingly develop an end-to-end approach to learning TWC-rules, underpinned by a so-called TWC-Rule Learning Model or TRLM for short to simulate the inference of TWC-rules. We theoretically show that a set of TWC-rules can be encoded by a certain parameter assignment of TRLM. To improve both the training and inference efficiencies of TRLM, we propose a tensorization optimization of TRLM based on batched sparse tensor operations. Experimental results on ten benchmark datasets demonstrate that the introduction of time window constraint enables TRLM to outperform state-of-the-art (SOTA) rule-based approaches on timestamp-based TKGs and achieve competitive performance against SOTA embedding-based approaches in link forecasting. Notably, TRLM scales to more than ten million facts, achieving a new SOTA on all TGB2.0 benchmark datasets.
Knowledge graph completion (KGC) aims to predict missing relationships between pairs of entities. Recently, Message Passing Neural Networks (MPNNs) have been employed in KGC and demonstrated superior ability in embedding learning. Current methods, however, face two major dilemmas. One focuses on learning complex entity embedding while neglecting the importance of relation embedding learning. The other is that some KGC methods conduct relation embedding learning through a neural network. However, this learning procedure limits learning independence from neighboring entities. To address this issue, we propose a Relation Entity Graph Message Passing mechanism (REGMP) to model the learning of entity-dependent relation embedding. Firstly, REGMP constructs a novel relation entity graph, which transforms relation edges into nodes with new edges and directions. The re-constructed relation entity graph builds a direct linkage between entities and relations during embedding learning. Besides, to conduct the embedding learning on the relation entity graph, we propose a Relation Entity Graph Message-passing mechanism to exploit local and fine-grained information. Moreover, to ease the inevitable over-smoothing problem, we develop a reaction-diffusion message-passing network. Experimental results on three knowledge graph completion datasets, including WN18RR, FB15k-237, and Kinship demonstrate the superiority of REGMP.
In many real-world applications, data are collected as a continuous stream, which exhibits an evolving distribution over time. How to design algorithms to track these evolving data distributions with provable guarantees, particularly in terms of the generalization ability, remains a formidable challenge. To handle this crucial problem, we propose a novel framework called Adaptive Model Rectification (AMR), which enables the model to effectively track the evolving data distributions. Specifically, we propose Encoding Feature Marginal Distribution Information (EFMDI) to extend optimal transport beyond its conventional limitations in characterizing environmental changes, thereby enabling effective model reuse across evolving data distributions. EFMDI enhances the reusability and evolvability of the AMR framework, allowing it to better accommodate distributional shifts over time. Furthermore, to enhance the model interpretability, we not only analyze the generalization ability of the local step in the evolution process, but also investigate the generalization ability associated with the entire classifier trajectory of the evolution based on the Fisher-Rao distance. To illustrate, we develop two concrete instantiations of the proposed AMR framework: T-AMR, a kernel-based approach tailored for tabular data, and I-AMR, a network-based approach designed for image data. Both variants are developed to address stream classification under evolving data distributions. Extensive experiments on both synthetic and real-world evolving data streams validate the effectiveness and practical applicability of the AMR framework.
