Social bot detection aims to precisely identify and mitigate the effects of bots spreading misinformation and manipulating public opinion by analyzing users’ textual content and network relationships on social media. Many existing methods rely on graph-based modalities, and while some incorporate multiple modalities, they often struggle to effectively capture the intricate interrelations within multimodal data. To address this limitation, we propose DMIBot, a dynamic multimodal interaction framework that comprehensively integrates metadata, text, and graph modalities. The framework utilizes a relational-aware graph aggregation module to enhance user node representations and a dynamic mixture of experts module to selectively engage the most appropriate expert for each sample. Furthermore, we introduce a synergistic interaction module that employs layer-wise recursive deep interaction mechanisms for efficient and effective integration of the three modalities. Experiments demonstrate that DMIBot outperforms nine benchmark models on two public datasets, achieving state-of-the-art performance in bot detection. Additional evaluations confirm its capabilities in ensuring the framework’s effectiveness and robustness in detecting Twitter bots.

Abstract New knowledge builds upon existing foundations, which means an interdependent relationship exists between knowledge, manifested in the historical records of the scientific system for hundreds of years. By leveraging natural language processing techniques, this study introduces the Scientific Concept Navigator, an embedding‐based navigation model to infer the “knowledge pathway” from the research trajectories of millions of scholars. We validate that the learned representations effectively delineate disciplinary boundaries and capture the intricate relationships between diverse concepts. Utility of the navigation space is showcased through multiple applications. Firstly, we demonstrate the multi‐step analogy inferences between concepts from various disciplines. Secondly, we formulate the cross‐domain conceptual dimensions of knowledge, observing the distributional shifts of 19 disciplines along these conceptual dimensions, including “Theoretical” to “Applied,” and “Societal” to “Economic,” highlighting the evolution of functional attributes across diverse domains. Lastly, by analyzing the knowledge network structure, we find that knowledge connects with shorter global pathways, and interdisciplinary concepts play a critical role in enhancing accessibility. Our framework offers a novel approach to mining knowledge inheritance pathways from extensive scientific literature, which is of great significance for understanding scientific progression patterns, tailoring scientific learning trajectories, and accelerating scientific progress.

Emotion recognition has become a focal research area in brain-computer interfaces, aiming to leverage artificial intelligence for diagnosing and treating clinical disorders. Neuropsychological studies indicate a close relationship between the activities of different brain functional areas and emotions. Consequently, we introduce a Local-Global Graph Attention Network (LG-GAT), a model inspired by neurology that learns the intra-regional and inter-regional activities of the brain by studying the local-global graph representations of electroencephalography (EEG). The LG-GAT mainly consists of two modules, namely the temporal convolution layer and kernel-level attention fusion (KAF) module, the feature graph building (LTG) module. The KAF module is to learn specific task-related features from EEG signals and to fuse the features learned from different kernels. The LTG module utilizes graph attention networks combined with local and global graphical representations with neurophysiological significance to model complex relationships within and between functional areas of the brain. To demonstrate LG-GAT’s efficacy, it was evaluated on three public datasets: DEAP, DREAMER, and SEED. Furthermore, LG-GAT’s performance was compared with state-of-the-art (SOTA) methods. The results indicate that LG-GAT surpasses these methods, validating the enhancement of emotion recognition performance through the integration of neuroscientific prior knowledge into neural network design.

Electroencephalogram (EEG) emotion recognition has become a central research focus within brain-computer interface (BCI) studies, where efficient emotion classification models can enable intelligent emotion regulation. In this paper, we propose a novel model based on three-dimensional convolutional neural networks (3D-CNN), residual networks (ResNet), and graph attention networks (GAT), named 3DCR-GAT, for EEG emotion recognition. The main advantage of the 3DCR-GAT model is its ability to simultaneously process multi-channel EEG signals, capturing complex spatiotemporal dependencies and local features within a unified network framework. Additionally, the ResNet employs skip connections to address the vanishing gradient problem during deep network training, improving the stability and expressiveness of the model. The attention mechanism within the model dynamically adjusts feature weights, captures global features, and enhances the accuracy of EEG emotion classification. Experimental results on public dataset, demonstrate that 3DCR-GAT significantly outperforms traditional methods in classification accuracy and feature extraction capability. Ablation studies reveal that each component of the model contributes substantially to its overall performance. Specifically, our model achieve an average accuracy of 92% on SEED for multi-state emotion recognition tasks. The 3DCR-GAT model represents an advanced and effective tool for EEG emotion classification, with broad potential applications.

Videos such as movies or TV episodes usually need to divide the long storyline into cohesive units, i.e., scenes, to facilitate the understanding of video semantics. The key challenge lies in finding the boundaries of scenes by comprehensively considering the complex temporal structure and semantic information. To this end, we introduce a novel Context-Aware Transformer (CAT) with a self-supervised learning framework to learn high-quality shot representations, for generating well-bounded scenes. More specifically, we design the CAT with local-global self-attentions, which can effectively consider both the long-term and short-term context to improve the shot encoding. For training the CAT, we adopt the self-supervised learning schema. Firstly, we leverage shot-to-scene level pretext tasks to facilitate the pre-training with pseudo boundary, which guides CAT to learn the discriminative shot representations that maximize intra-scene similarity and inter-scene discrimination in an unsupervised manner. Then, we transfer contextual representations for fine-tuning the CAT with supervised data, which encourages CAT to accurately detect the boundary for scene segmentation. As a result, CAT is able to learn the context-aware shot representations and provides global guidance for scene segmentation. Our empirical analyses show that CAT can achieve state-of-the-art performance when conducting the scene segmentation task on the MovieNet dataset, e.g., offering 2.15 improvements on AP.

Estimated time of arrival (ETA) is a very important factor in the transportation system. It has attracted increasing attentions and has been widely used as a basic service in navigation systems and intelligent transportation systems. In this paper, we propose a novel solution to the ETA estimation problem, which is an ensemble on tree models and neural networks. We proved the accuracy and robustness of the solution on the A/B list and finally won first place in the SIGSPATIAL 2021 GISCUP competition.