Weihua Liu

Attributed graphs, which associate vertices or edges with attributes, are essential in applications like social network analysis, recommendation systems, bioinformatics, and healthcare analysis. However, their high-dimensional and non-Euclidean nature complicates tasks such as vertex classification, community detection, graph visualization, and graph embedding. Existing techniques often inadequately integrate structural and attribute data, and reconstruction-centric models can compromise graph topology. Additionally, many methods require extensive manual hyperparameter tuning. To address these challenges, we introduce Attributed Graph Clustering with Transitive Order convolutional autoencoder (AGCTO). AGCTO uses a Graph Transitive Convolutional AutoEncoder (GTCAE) framework that integrates structural and attribute information. A key innovation is the Graph ORder Distance (GORD), capturing topological relationships to enhance clustering performance. AGCTO's unified loss function combines GTCAE loss on the original graph, GTCAE loss on a simplified graph via GORD, and a topology-preserving loss. This ensures effective fusion of attribute and structure data and preserves graph topology. Evaluations on four real-world and two synthetic attributed graphs show AGCTO's superiority in clustering accuracy, normalized mutual information, F1-score, and Q-modularity. AGCTO offers a robust solution for attributed graph clustering, maintaining graph topology and delivering superior performance.

Sarcasm detection is a nuanced challenge in natural language processing, requiring deep understanding of textual and contextual cues. We present Sarcasm-GPT, a large language model-based model that integrates four key components: prompt template generation, retrieval-augmented generation, chain-of-thought generation, and a context fusion module. Together, these modules enrich contextual modeling, enable systematic reasoning, and seamlessly incorporate multimodal information, significantly boosting sarcasm detection accuracy. Extensive experiments, including ablation studies, confirm the complementary contributions of each module and demonstrate substantial performance gains over baselines. Additionally, our findings highlight the importance of context length in balancing interpretive accuracy and computational efficiency. The code is available at Sarcasm-GPT.

Deep learning has shown great promise in hyperspectral unmixing, especially the unmixing methods based on autoencoder networks are the most prevalent these days. Since most spectral mixing problems are nonlinear and cannot effectively utilize global context information, these methods have limited generalization ability under different ground features and scenarios. To address these limitations, an enhanced network based on spatial-spectral information attention is proposed. A two-channel attention mechanism is embedded within the convolutional autoencoder to acquire the feature dependencies. The spatial information extraction uses the dynamic large kernel block (DLK) to obtain the global spatial attention of the image. The DLK module uses multiple large kernels with different kernel sizes and dilation rates to capture multi-scale features, and the spectral information extraction uses Coordinate Attention to capture spectral correlation information. This can improve the quality of the endmember spectra and abundance maps. On real and synthetic data, this model is compared with several advanced unmixing methods, and the results show the effectiveness of this method.

In this paper, we aim to enhance the generalization ability of single-shot face anti-spoofing tasks in two aspects: 1) we establish a deep frequency factorization network (DFF-Net) to capture deep frequency information, enabling better discrimination of multiple styles of face spoofing, including both 2D and 3D methods; 2) aided by a sample redistribution strategy, we construct a unified anomaly metric-based supervision system to address unknown face attacks. Specifically, the proposed DFF-Net explicitly extracts spoofing cues from the high, low, and fusion frequency domains by embedding a deep frequency filter module into each frequency branch. Furthermore, the devised sample redistribution strategy enables the reorganization of learned features, making them more suitable for the three types of anomaly metric-based supervisions that correspond to the three frequency branches. This approach facilitates iterative learning of the importance of spoofing cues. Experimental results demonstrate that our proposed framework achieves state-of-the-art performance on both monocular and multi-spectral benchmark datasets.

In recent years, the use of action face anti-spoofing (FAS) tasks has become widespread in various identity security authentication scenarios. However, the current performance of action FAS is susceptible to the impact of lighting conditions and the scale of the action. Furthermore, due to the simplicity of facial action behavior, it is unable to withstand highly realistic 3D counterfeit face attacks. To address these issues, this paper proposes a spatio-temporal-based action FAS framework that combines dynamics and texture information around local keypoint areas of the face. To tackle the issue of action scale, we integrate a clip of face action video into a single face keypoint-based feature map, incorporating motion trajectories and motion history cues. Specifically, we develop two essential features in the spatiotemporal domain: Keypoints Trajectory Invariant Feature (KTIF) and Keypoints Motion History Feature (KMHF). These features are constructed by utilizing the absolute position of keypoints and the relative bias position of keypoints between frames. Additionally, to further enhance the security level of the action FAS task, we propose a novel representation called Keypoints Neighbourhood Texture Difference Feature (KNTDF). This representation utilizes an Encoder-Decoder module with center difference convolution (CDC) to effectively address the issue of 2D/3D spoofing at the image noise level. Finally, by embedding the Swin-transformer architecture with a channel attention module and skip fusion strategy, the above three deep representations are fused and further classified for both the face action classification (FAC) task and the face anti-spoofing task. The experimental results show that the proposed method performs well on both tasks. Especially for action FAS, compared to the majority of existing work, it demonstrates significant capabilities in overcoming lighting conditions and motion scales.

With recent advances in hyperspectral imaging sensors, subtle and concealed targets that cannot be detected by multispectral imagery can be identified. The most widely used anomaly detection method is based on the Reed–Xiaoli (RX) algorithm. This unsupervised technique is preferable to supervised methods because it requires no a priori information for target detection. However, two major problems limit the performance of the RX detector (RXD). First, the background covariance matrix cannot be properly modelled because the complex background contains anomalous pixels and the images contain noise. Second, most RX-like methods use spectral information provided by data samples but ignore the spatial information of local pixels. Based on this observation, this article extends the concept of the weighted RX to develop a new approach called an adaptive saliency-weighted RXD (ASW-RXD) approach that integrates spectral and spatial image information into an RXD to improve anomaly detection performance at the pixel level. We recast the background covariance matrix and the mean vector of the RX function by multiplying them by a joint weight that in fuses spectral and local spatial information into each pixel. To better estimate the purity of the background, pixels are randomly selected from the image to represent background statistics. Experiments on two hyperspectral images showed that the proposed random selection-based ASW RXD (RSASW-RXD) approach can detect anomalies of various sizes, ranging from a few pixels to the sub-pixel level. It also yielded good performance compared with other benchmark methods.