Temporal Graph Neural Networks (TGNNs) have become indispensable for analyzing dynamic graphs in critical applications such as social networks, communication systems, and financial networks. However, the robustness of TGNNs against adversarial attacks, particularly sophisticated attacks that exploit the temporal dimension, remains a significant challenge. Existing attack methods for Spatio-Temporal Dynamic Graphs (STDGs) often rely on simplistic, easily detectable perturbations (e.g., random edge additions/deletions) and fail to strategically target the most influential nodes and edges for maximum impact. We introduce the High Impact Attack (HIA), a novel restricted black-box attack framework specifically designed to overcome these limitations and expose critical vulnerabilities in TGNNs. HIA leverages a data-driven surrogate model to identify structurally important nodes (central to network connectivity) and dynamically important nodes (critical for the graph's temporal evolution). It then employs a hybrid perturbation strategy, combining strategic edge injection (to create misleading connections) and targeted edge deletion (to disrupt essential pathways), maximizing TGNN performance degradation. Importantly, HIA minimizes the number of perturbations to enhance stealth, making it more challenging to detect. Comprehensive experiments on five real-world datasets and four representative TGNN architectures (TGN, JODIE, DySAT, and TGAT) demonstrate that HIA significantly reduces TGNN accuracy on the link prediction task, achieving up to a 35.55% decrease in Mean Reciprocal Rank (MRR) - a substantial improvement over state-of-the-art baselines. These results highlight fundamental vulnerabilities in current STDG models and underscore the urgent need for robust defenses that account for both structural and temporal dynamics. Code and Data are available at https://github.com/ryandhjeon/hia.

While deep-learning-enabled recommender systems demonstrate strong performance benchmarks, many struggle to adapt effectively in real-world environments due to limited use of user-item relationship data and insufficient transparency in recommendation generation. Traditional collaborative filtering approaches fail to integrate multifaceted item attributes, and although Factorization Machines account for item-specific details, they overlook broader relational patterns. Collaborative knowledge graph-based models have progressed by embedding user-item interactions with item-attribute relationships, offering a holistic perspective on interconnected entities. However, these models frequently aggregate attribute and interaction data in an implicit manner, leaving valuable relational nuances underutilized.This study introduces the Knowledge Graph Attention Network with Information Fusion (KGIF), a specialized framework designed to merge entity and relation embeddings explicitly through a tailored self-attention mechanism. The KGIF framework integrates reparameterization via dynamic projection vectors, enabling embeddings to adaptively represent intricate relationships within knowledge graphs. This explicit fusion enhances the interplay between user-item interactions and item-attribute relationships, providing a nuanced balance between user-centric and item-centric representations. An attentive propagation mechanism further optimizes knowledge graph embeddings, capturing multi-layered interaction patterns. The contributions of this work include an innovative method for explicit information fusion, improved robustness for sparse knowledge graphs, and the ability to generate explainable recommendations through interpretable path visualization. The implementation and datasets for this study are publicly available1.

As a special type of Knowledge Graph (KG), Continual Knowledge Graph Learning (CKGL) plays a pivotal role in various areas such as recommendation systems, search engines, and personalized services, where knowledge dynamically evolves. A challenging problem in CKGL is catastrophic forgetting, where models forget previously learned knowledge upon being trained on new data. To overcome the challenge, this study proposes Flexible Memory Rotation (FMR), a dual-level regularization technique that focuses on both parameter level and structural level. Our idea is inspired by the natural human learning process, which tends to memorize correctly learned knowledge, leverage the learned to acquire new knowledge, and refine incorrectly learned knowledge with newly obtained information. Commonly, existing regularization-based methods fail to mimic this human nature by having a fixed constraint strategy for all model parameters. To this end, the proposed FMR offers flexible constraints based on qualities of learned knowledge evaluated by the Fisher Information Matrix (FIM). Additionally, we identified a limitation of FIM in CKGL, which is the assumption of independence of time steps does not always hold. To overcome this, FMR rotates the parameter space to diagonalize the FIM. This work has four major contributions: 1) develop a novel regularization technique, FMR, a flexible regularization technique, 2) reveal the unexpected failure of FIM in CKGL and provides an easy remedy via parameter space rotation, 3) the comparison experiments on four benchmark datasets designed for CKGL demonstrates improvement over various state-of-the-art (SOTA) CKGL models, and 4) a comprehensive ablation study investigates each component of the proposed model. The source code is available at https://github.com/lijingzhu1/FMR.

In this paper, we study a not well-investigated but important transfer learning problem termed Distant Domain Transfer Learning (DDTL). This topic is closely related to negative transfer. Unlike conventional transfer learning problems which assume that the source domain and the target domain are more or less similar to each other, DDTL aims to make efficient transfers even when the domains or the tasks are completely different. As an extreme example in image classification, there are only a sufficient amount of unlabeled images of watches, airplanes, and horses in the source domain, and the target domain only has a small set of labeled human face images. Previously, a few instance-based distant domain transfer algorithms were proposed to deal with this type of binary distant domain image classification problems. Yet most existing algorithms are very task-specific and they are only good at binary classification tasks. In this study, we propose a novel feature-based distant domain transfer learning algorithm, which requires only a tiny set of labeled target data and unlabeled source data from completely different domains. Instead of selecting intermediate instances, we introduced Distant Feature Fusion (DFF), a novel feature selection method, to discover general features cross distant domains and tasks by using convolutional autoencoder with a domain distance measurement as a feature extractor. As the novelty of this study, it can effectively handle both distant domain mutil-class image classification and binary image classification problems. More importantly, it has achieved up to 19% higher classification accuracy than "non-transfer" algorithms, and up to 9% higher than existing distant transfer algorithms.