Youquan Wang

Multi-behavior recommendation leverages diverse user interactions to capture personalized user preferences. To alleviate the data sparsity of the crucial target behavior, multi-behavior contrastive learning (MBCL) improves performance by enriching self-supervised signals from auxiliary behaviors. However, existing methods rely on heuristic augmentation techniques to construct contrastive views, which disrupt essential multi-behavior collaborative information; simultaneously, auxiliary behaviors, influenced by multiple confounders, easily introduce semantic confusion into the MBCL alignment process. Inspired by recent advances of diffusion models, we propose DiffMB, a novel diffusion contrastive learning framework. DiffMB first employs a Multi-Behavior Weight-Enhanced Graph Neural Network to capture behavioral heterogeneity through parallel encoding. At the core of DiffMB is a Hybrid Contrastive Learning (HCL) strategy that tackles this dilemma via two synergistic components. First, Target-Behavior Diffusion Contrastive Learning leverages a diffusion process to generate high-quality, semantically consistent views for the target behavior, thereby consolidating its semantics while avoiding the disruption of collaborative information. Second, Multi-Behavior Collaborative Contrastive Learning extracts valuable collaborative signals from auxiliary behaviors. Furthermore, a Dynamic Optimization Strategy is introduced to reconcile gradient conflicts in multi-task training. Extensive experiments on three datasets demonstrate that DiffMB significantly outperforms state-of-the-art baselines.

Vision-Language Pre-Training (VLP) aims to convert vision-language tasks into image-text matching challenges, necessitating robust interaction between visual and textual modalities. However, current research often faces a tradeoff: methods pursuing fine-grained alignment typically rely on computationally expensive object detectors and pre-defined annotations, while coarse-grained alignment methods overlook subtle visual details. To alleviate these issues, we present OMEGA, a C O mprehensive M ulti-l E vel G ranularity A lignment framework. OMEGA achieves sufficient cross-modal alignment without relying on external object detectors or annotations. Specifically, we first design a Text-Aware Patch Selection (TAPS) module, which dynamically constructs patch-level regions for fine-grained excavation based on informative token selection, effectively bypassing the need for bounding boxes. Second, to facilitate multi-grained information fusion, we introduce a Cross-Grained Alignment (CGA) module to learn modality-shared features across global and object levels. Experimental results demonstrate that OMEGA surpasses existing approaches in five pivotal vision-language tasks—image-text retrieval, visual question answering, visual reasoning, visual entailment, and image captioning, while maintaining superior inference efficiency compared to detector-based methods. We believe this study offers a promising object-free paradigm for scalable vision-language pre-training.

Vision-language-action (VLA) models map visual observations and language instructions to robot actions. Precise manipulation, however, depends on fine-grained visual cues remaining available as representations pass through the policy backbone. Our visual-token intervention analysis shows that pruning visual tokens at deeper layers produces smaller changes in action-prediction MSE. Motivated by this depth-dependent pattern, we introduce RelayVLA, which adds a parallel spatial pathway at an intermediate layer without changing the original VLA input interface. A frozen VGGT spatial foundation model extracts candidate spatial tokens from multi-view observations. Visual hidden states at the injection layer select the candidates most relevant to the current scene, which are then incorporated through gated cross-attention before action decoding. The pretrained VLM pathway remains intact, while later layers receive additional spatial cues. We evaluate RelayVLA on four simulation benchmarks spanning multi-task manipulation, long-horizon execution, controlled perturbations, and structured distribution shifts. RelayVLA obtains the highest reported aggregate score on each benchmark, with the largest gains on long-horizon tasks and under geometric or sensing perturbations. These results indicate that injecting VGGT-derived spatial features at intermediate layers can improve VLA control when tasks place greater demands on spatial reasoning.

Multi-Behavior Recommendation (MBR) aims to model personalized user preferences by integrating diverse interaction behaviors (e.g., page view, favorite, add to cart, purchase). However, latent confounders such as contextual influences and social relationships can obscure the true causal effects in real-world scenarios, thereby confounding the model’s prediction. Although existing MBR research extensively explores behavioral dependencies and heterogeneity, it frequently overlooks the impact of latent confounders, thereby limiting its ability to capture users’ genuine preferences. To address the limitations of existing methods, we identify two key challenges in MBR: (1) how to infer latent confounders, and (2) how to mitigate their influence across multi-behavior interactions. To this end, we propose Causal Variational Inference for Deconfounded (CVID) MBR. CVID employs a variational graph autoencoder to model latent uncertainty in multi-behavior interactions and introduces a confounder inference module to generate behavior-specific latent confounders via variational inference. In the conditional diffusion module, noise is progressively injected during the forward process to simulate the dynamic evolution of user preferences, while the reverse process leverages the inferred latent confounders to guide denoising through back-door adjustment, thereby recovering the true causal effects between multi-behavior interactions and the model’s prediction. Extensive experiments on public multi-behavior datasets demonstrate that CVID consistently outperforms state-of-the-art baselines in mitigating confounding effects and improving recommendation accuracy, validating its effectiveness and superiority.

Multivariate time series anomaly detection has shown potential in various fields, such as finance, aerospace, and security. The fuzzy definition of data anomalies, the complexity of data patterns, and the scarcity of abnormal data samples pose significant challenges to anomaly detection. Researchers have extensively employed autoencoders (AEs) and generative adversarial networks (GANs) in studying time series anomaly detection methods. However, relying on reconstruction error, the AE-based anomaly detection algorithm needs more effective regularization methods, rendering it susceptible to the problem of overfitting. Meanwhile, GAN-based anomaly detection algorithms require high-quality training data, significantly impacting their practical deployment. We propose a novel GAN based on a dual-discriminator structure to address these issues. The model first processes the data with the generator to obtain the reconstruction error and then calculates pseudo-labels to divide the data into two categories. One data category is input into the first discriminator, where a minor loss between the data and its reconstructed counterpart is better. The other data category is input into the second discriminator, where a larger loss between the data and its reconstructed counterpart is better. Through this process, the model can effectively constrain the generator, retaining information on normal data during data reconstruction while discarding information on abnormal data. After conducting experiments on multiple benchmark datasets, the proposed GAN based on a dual-discriminator structure achieved good results in anomaly detection, outperforming several advanced methods. Additionally, the model also performed well in practical transformer data.

The goal of partial multi-label learning is to induce a multi-label classifier from partial multi-label data where each instance is annotated with a number of candidate labels but only a subset of them are valid. Many of the existing studies either fail to fully utilize instance and label correlations to eliminate noisy labels or build an over-simplified multi-label classifier, both of which are unfavorable for the improvement of generalization performance. In this article, we put forward a novel model named P ml-ilc to learn a multi-label classifier from partial multi-label data. Specifically, P ml-ilc first encodes instances and labels into a compact semantic space and takes full advantage of instance and label correlations to eliminate noisy labels. Then, it induces a linear mapping from the feature space to the label space while exploiting label-specific features and instance correlations to facilitate the multi-label classifier learning process. Finally, the above two steps are combined into a joint optimization problem and an efficient alternating optimization procedure is developed to find a satisfactory solution. Extensive experiments show that P ml-ilc achieves superior performance on both real-world and synthetic partial multi-label datasets in terms of different evaluation metrics.

Fraudsters excel at concealing themselves by establishing connections with many trustworthy users to evade detectors. This cunning behavior increases the heterogeneity within the graph, rendering ineffective Graph Neural Network (GNN) methods based on homogeneity assumptions. To tackle this problem, several spectral GNN-based approaches have been introduced. However, existing methods fail to fully exploit frequency information, often overlooking signals of other frequencies in the graph, which diminishes the model's ability to capture node characteristics. Furthermore, many methods employ global filtering GNNs, neglecting the different frequencies of attention for fraudulent and benign nodes. To mitigate these shortcomings, we propose a new method called Node-Level Multi-Frequency Adaptive Graph Neural Network (NMFA). Specifically, our method is equipped with three multi-frequency filter groups—low-pass, band-pass, and high-pass—aimed at extracting comprehensive frequency information for node representation learning. Moreover, we devised a multi-frequency adaptive fusion mechanism based on the attention mechanism, enabling personalized filtering for each node. Experiments conducted on real datasets demonstrate that our method significantly improves fraud detection performance compared to the advanced GNN fraud detection methods.