In recent years, Graph Contrastive Learning (GCL) has emerged as a key research direction for learning representations of unlabeled graph data, focusing on the self-supervised learning of efficient representations for both graphs and nodes. However, existing approaches typically assume feature homogeneity across different augmented views, overlooking the potential impact of inter-view feature differences, particularly weak features, on model performance. To address the problem of weak features and the feature differences between different enhanced views, this paper proposes the Feature Discrepancy-Aware Graph Contrastive Learning (FDAGCL) framework. Firstly, FDAGCL dynamically adjusts the importance of features through the feature importance decoupling mechanism, thereby effectively distinguishing strong features from weak view. Secondly, FDAGCL designs a multi-view map contrastive learning strategy to enhance the expression of strong features while simultaneously improving the learning of weak features through strong-strong view and strong-weak view map contrastive learning, thereby achieving information complementarity. To validate the effectiveness of our method, we conducted extensive empirical experiments on five datasets. The results demonstrate that FDAGCL exhibits significant superiority over the baseline methods.

Sentiment analysis is a critical task for natural language processing. Much research has been done for high-resource languages such as English and Chinese. However, Tibetan is an extremely low resource language with less reference information. According to the practical demands, this article proposes RPEPL, a Tibetan sentiment analysis method based on relative position encoding and prompt learning. First, word information is introduced to syllable sequences by converting the directed acyclic lattice into a squashed structure. Second, a relative position encoding is used to encode the position information of syllables and words. Third, the association relations and semantic information of tokens are identified by leveraging the multi-attention. Finally, the sentiment category of the Tibetan sentence is obtained through a prompt learning framework. Experimental results demonstrate that RPEPL significantly outperforms the baseline methods on the TUSA dataset and TNEC (Tibetan News Event Comments) dataset. Additionally, traditional recurrent neural networks cannot perform large-scale parallel computation and convolutional neural networks have difficulty modeling long-distance dependencies in Tibetan text, both of which are resolved using RPEPL. Furthermore, the use of multi-attention not only enriches the association relations between syllables and words but also enhances the understanding of sentence semantic and syntactic structure information, and improves the performance of Tibetan sentiment analysis.

Recently, memory-based methods have exhibited remarkable performance in Video Object Segmentation (VOS) by employing non-local pixel-wise matching between the query and memory. Nevertheless, these methods suffer from two limitations: 1) Non-local pixel-wise matching can result in the incorrect segmentation of background distractor objects, and 2) memory features with substantial temporal redundancy consume significant computing resources and reduce the inference speed. To address the limitations, we first propose a local attention mechanism to suppress background features, and we introduce a novel training framework based on contrast learning to ensure the network learns reliable and robust pixel-wise correspondence between query and memory. We adaptively determine whether to update the memory based on the variation of foreground objects. Next, we propose a dynamic memory bank, which utilizes a lightweight and differentiable soft modulation gate to determine the number of memory features to remove along the temporal dimension. This allows efficient and flexible management of memory features. Our network achieves competitive results (e.g., 92.1% on DAVIS 2016 val, 87.6%/81.3% on DAVIS 2017 val/test, 87.0% on YouTube-VOS 2018 val) compared with the state-of-the-art methods while maintaining a faster inference speed of 25+FPS. Moreover, our network demonstrates a favorable balance between performance and speed when dealing with the long-time video dataset.

In recent years, convolutional neural networks (CNNs) have achieved remarkable success in hyperspectral image (HSI) classification tasks, primarily due to their outstanding spatial feature extraction capabilities. However, CNNs struggle to capture the diagnostic spectral information inherent in HSI. In contrast, vision transformers exhibit formidable prowess in handling spectral sequence information and excelling at capturing long-range correlations between pixels and bands. Nevertheless, due to the information loss during propagation, some existing transformer-based classification methods struggle to form sufficient spectral-spatial information mixing. To mitigate these limitations, we propose a memory-augmented spectral-spatial transformer (MASSFormer) for HSI classification. Specifically, MASSFormer incorporates two efficacious modules, the memory tokenizer (MT) and the memory-augmented transformer encoder (MATE). The former serves to transform spectral-spatial features into memory tokens for storing prior knowledge. The latter aims to extend traditional multi-head self-attention (MHSA) operations by incorporating these memory tokens, enabling ample information blending while alleviating the potential depth decay in the model, and consequently improving the model’s classification performance. Extensive experiments conducted on four benchmark datasets demonstrate that the proposed method outperforms state-of-the-art methods. The source code is available at https://github.com/hz63/MASSFormer for the sake of reproducibility.

In recent years, heterogeneous graph neural networks have become the mainstream approach for handling heterogeneous graph data. However, due to the sparsity of labels, most existing methods for heterogeneous graph neural networks typically employ a semi-supervised learning approach, which has certain limitations in practical applications. To address this issue, we propose a self-supervised heterogeneous graph representation learning method, namely Multi-scale Heterogeneous Graph Contrastive Learning (MHGCL). This approach decodes encoded information from two perspectives: meta-paths and network patterns, in a multi-scale fashion. It uses a loss function that maximizes the similarity between positive pairs at different scales and minimizes the similarity between negative pairs. This encourages related nodes and edges to be close to each other in the embedding space, while unrelated nodes and edges are pushed farther apart. Experimental results demonstrate that MHGCL comprehensively captures semantic information at different scales between nodes. It exhibits excellent performance in node classification tasks, validating its effectiveness in heterogeneous graph node embedding learning.

Abstract Currently, most researchers only utilize the network information or node characteristics to calculate the connection probability between unconnected node pairs. Therefore, we attempt to project the problem of connection probability between unconnected pairs into the physical space calculating it. Firstly, the definition of gravitation is introduced in this paper, and the concept of gravitation is used to measure the strength of the relationship between nodes in complex networks. It is generally known that the gravitational value is related to the mass of objects and the distance between objects. In complex networks, the interrelationship between nodes is related to the characteristics, degree, betweenness, and importance of the nodes themselves, as well as the distance between nodes, which is very similar to the gravitational relationship between objects. Therefore, the importance of nodes is used to measure the mass property in the universal gravitational equation and the similarity between nodes is used to measure the distance property in the universal gravitational equation, and then a complex network model is constructed from physical space. Secondly, the direct and indirect gravitational values between nodes are considered, and a novel link prediction framework based on the gravitational field, abbreviated as LPFGF, is proposed, as well as the node similarity framework equation. Then, the framework is extended to various link prediction algorithms such as Common Neighbors (CN), Adamic-Adar (AA), Preferential Attachment (PA), and Local Random Walk (LRW), resulting in the proposed link prediction algorithms LPFGF-CN, LPFGF-AA, LPFGF-PA, LPFGF-LRW, and so on. Finally, four real datasets are used to compare prediction performance, and the results demonstrate that the proposed algorithmic framework can successfully improve the prediction performance of other link prediction algorithms, with a maximum improvement of 15%.

Graph neural networks (GNNs) have shown significant performance in various practical applications due to their strong learning capabilities. Backdoor attacks are a type of attack that can produce hidden attacks on machine learning models. GNNs take backdoor datasets as input to produce an adversary‐specified output on poisoned data but perform normally on clean data, which can have grave implications for applications. Backdoor attacks are under‐researched in the graph domain, and almost existing graph backdoor attacks focus on the graph‐level classification task. To close this gap, we propose a novel graph backdoor attack that uses node features as triggers and does not need knowledge of the GNNs parameters. In the experiments, we find that feature triggers can destroy the feature spaces of the original datasets, resulting in GNNs inability to identify poisoned data and clean data well. An adaptive method is proposed to improve the performance of the backdoor model by adjusting the graph structure. We conducted extensive experiments to validate the effectiveness of our model on three benchmark datasets.

Currently, most existing link prediction algorithms simply study the interrelationships between node pairs without considering the interaction force and the higher order relationships between node pairs. In order to find a solution to this problem, the concept of the gravitational field is introduced in this article, and then, a novel algorithmic framework is proposed from the perspective of physics. The framework is applied to the classic link prediction algorithms to effectively enhance their prediction performance. First, the node contraction method is applied to measure the node importance, and a similarity-based link prediction algorithm is used to calculate the similarity values between node pairs. Second, the importance of nodes is introduced into the gravitational field model as the mass attribute, and the similarity values between node pairs are used as a distance metric between node pairs. Thereby, a gravitational field model of the complex network from the perspective of physics is established. Finally, the edges of the undirected complex network are assigned the weights, and a weighted local random walking-based link prediction algorithm is proposed. The link prediction method is adopted to evaluate the reasonableness and practical value of the gravitational field model. Experimental results show that most link prediction algorithms using the proposed algorithmic framework have got improvement with a minimum improvement of 2% and a maximum improvement of 33%; thus, the effectiveness and feasibility of the algorithm are verified.

Most word embedding models have the following problems: (1) In the models based on bag-of-words contexts, the structural relations of sentences are completely neglected; (2) Each word uses a single embedding, which makes the model indiscriminative for polysemous words; (3) Word embedding easily tends to contextual structure similarity of sentences. To solve these problems, we propose an easy-to-use representation algorithm of syntactic word embedding (SWE). The main procedures are: (1) A polysemous tagging algorithm is used for polysemous representation by the latent Dirichlet allocation (LDA) algorithm; (2) Symbols ‘+’ and ‘−’ are adopted to indicate the directions of the dependency syntax; (3) Stopwords and their dependencies are deleted; (4) Dependency skip is applied to connect indirect dependencies; (5) Dependency-based contexts are inputted to a word2vec model. Experimental results show that our model generates desirable word embedding in similarity evaluation tasks. Besides, semantic and syntactic features can be captured from dependency-based syntactic contexts, exhibiting less topical and more syntactic similarity. We conclude that SWE outperforms single embedding learning models.