We present MarsViT-MAE, the first self-supervised Vision Transformer trained entirely on seven million Context Camera (CTX) crops, purpose-built for Martian geomorphology. The resulting 768-dimensional ViT outperforms ImageNet-pretrained ViT-B/16 and Earth-observation backbones of similar size, and is competitive with much larger web-scale models on the DoMars16k retrieval benchmark—while using an order of magnitude fewer parameters. A systematic cross-domain comparison further shows that models optimized for Earth scenes generalize more poorly to Mars than generic Internet models, underscoring the importance of domain-native pretraining. To translate these gains into practice, we deploy MarsViT-MAE in a lightweight web portal that performs planet-scale nearest-neighbor search over two million CTX embeddings in about two seconds, and supports sub-second relevance feedback via an in-browser logistic head. This transforms hours of manual browsing into an interactive loop suitable for tasks such as crater-freshness triage and dune-type mapping. By uniting a Mars-specific foundation model with a minimal retrieval stack, we show that near state-of-the-art planetary mapping is achievable on a single Milvus node. Code, weights, and the live service are available at: hf.co/jfang/mars-vit-base-mae.

Network interdiction problems, arising in critical applications from military strategy to disease control, involve a complex attacker-defender dynamic: one player optimizes a network-based objective, while the other strategically modifies the network to impede that objective. The inherent bi-level optimization and combinatorial nature of these problems pose a significant computational challenge, often rendering traditional exact solvers impractical and hindering the development of effective heuristics. While Graph Neural Networks (GNNs) have demonstrated promise in solving single-level combinatorial optimization problems on graphs, their direct application to bi-level interdiction problems remains limited. In this paper, we bridge this gap by introducing a novel approach that leverages the power of GNNs to learn Mixed-Integer Linear Programming (MILP) formulations of network interdiction problems. By representing the problem in this structured mathematical form, we empower a multipartite GNN with the representational capacity to effectively capture the complex interplay between the two players. This approach aligns the neural network with the underlying mathematical structure of interdiction problems, leading to improved performance. Through extensive experiments on two network interdiction tasks, we demonstrate the superiority of our proposed method over both baseline GNN models and traditional exact solvers, showcasing its potential for real-world applications.

In recent years, Graph Neural Networks (GNNs) have attracted considerable attention. However, the rapid emergence of diverse GNN models, each grounded in different theoretical foundations, complicates the model selection process, as these models are not easily understood within a unified framework. Initial GNNs were constructed using spectral theory, while others were developed based on spatial theory. This theoretical divergence makes direct comparisons difficult. Furthermore, the variety of models within each theoretical domain further complicates their evaluation. In this tutorial, we explore state-of-the-art GNNs and present a comprehensive framework that bridges the spatial and spectral domains, clarifying their interrelationship. This framework deepens our understanding of GNN operations. The tutorial delves into key paradigms, such as spatial and spectral methods, through a synthesis of spectral graph theory and approximation theory. We conduct an in-depth analysis of recent research advancements, addressing emerging issues like over-smoothing, using well-established GNN models to illustrate the universality of our framework.

This work develops a node-edge co-evolution model for attributed graph transformation, where both the node and edge attributes undergo changes due to complex interactions. Due to two fundamental obstacles, learning and approximating attributed graph transformation have not been thoroughly explored: 1) the difficulty of jointly considering four types of atomic interactions including nodes-to-edges, nodes-to-nodes, edges-to-nodes, and edges-to-edges interactions. 2) the difficulty of capturing iterative long-range interactions between nodes and edges. To solve these issues, we offer a novel and scalable equilibrium model, NEC∞, with node-edge message passing and edge-node message passing. Additionally, we propose an efficient optimization algorithm that is based on implicit gradient theorem and includes a theoretical analysis of NEC∞. The effectiveness and efficiency of the proposed model have been demonstrated through extensive experiments on synthetic and real-world data sets.

The vast presence of bots on Twitter requires reliable and accurate bot detection methods that differentiate legitimate bots from malicious ones. Despite the success of those methods, they fail to address the following challenges: (1) the huge size of datasets required to train a model to detect bots, (2) the constant evolution in bot accounts to evade automatic detection leads to scarcity in ground truth real-world datasets, and (3) the complexity in learning representations of a heterogeneous attributed network like Twitter. In this paper, we propose a novel framework, ADNET, to detect anomalies in Twitter-attributed networks using the least amount of labeled data. Specifically, we address the limitations of previous methods by proposing a topology-based active learning framework that uses a deep autoencoder to train the model and is able to handle large graphs better than previous methods. Our experimental results demonstrate that the proposed approach outperforms state-of-the-art methods in detecting anomalous bot accounts and reduces the annotation cost in Twitter attributed networks.

The COVID-19 pandemic has caused hate speech on online social networks to become a growing issue in recent years, affecting millions. Our work aims to improve automatic hate speech detection to prevent escalation to hate crimes. The first c hallenge i n h ate s peech r esearch i s t hat e xisting datasets suffer from quite severe class imbalances. The second challenge is the sparsity of information in textual data. The third challenge is the difficulty i n b alancing t he t radeoff b etween utilizing semantic similarity and noisy network language. To combat these challenges, we establish a framework for automatic short text data augmentation by using a semi-supervised hybrid of Substitution Based Augmentation and Dynamic Query Expansion (DQE), which we refer to as SubDQE, to extract more data points from a specific c lass f rom T witter. W e a lso p ropose the HateNet model, which has two main components, a Graph Convolutional Network and a Weighted Drop-Edge. First, we propose a Graph Convolutional Network (GCN) classifier, using a graph constructed from the thresholded cosine similarities between tweet embeddings to provide new insights into how ideas are connected. Second, we propose a weighted Drop-Edge based stochastic regularization technique, which removes edges randomly based on weighted probabilities assigned by the semantic similarities between Tweets. Using 3 different SubDQE-augmented datasets, we compare our HateNet model using eight different tweet embedding methods, six other baseline classification models, and seven other baseline data augmentation techniques previously used in the realm of hate speech detection. Our results show that our proposed HateNet model matches or exceeds the performance of the baseline models, as indicated by the accuracy and F1 score.

Existing COVID-19 prediction models focus on studying the dynamic nature of the virus spread by using pandemic-related temporal data. In this paper, we present a work that exclusively uses comprehensive socioeconomic factors to predict the high risk areas of COVID-19 infection based on fine-grained static spatial analysis. Moreover, the most and least influential socioeconomic factors on COVID-19 spread are identified. This paper uses a uniquely built dataset by combining local states’ cumulative COVID-19 statistics and their associated socioeconomic features on the zip code level. Further, the work solves the lack of data by augmentation. To evaluate the work, four case studies are conducted on Florida, Illinois, Minnesota, and Virginia. Experimental results show that the study provides accurate predictions with respect to ground truth data. By identifying high risk areas and socioeconomic factors, policymakers can use this study to take necessary measures to help disadvantaged communities.

Predicting individuals personality traits with their social media profile has proved to be feasible, but researchers recently have run into bottlenecks on further improving the prediction accuracy. One major limitation is that existing studies failed to consider context information in predicting social media users' behaviors. In this paper, we adopted the DIAMONDS situation theory in psychology to capture the context information in Facebook posts. To solve this issue, we proposed a novel situation-based feature interaction learning model. In this study, we extracted situation features according to the DIAMONDS lexicon and computed the interaction values between these situation features and the commonly used n-gram features at the post level. Features at the post level were aggregated up to the user level using the averaging strategy. A group lasso penalty was employed to enforce strong heredity in the model, which addressed the overfitting challenge introduced by the interaction features. Empirical tests on a large-scale data set have demonstrated the effectiveness of the proposed method.