Graph Contrastive Learning (GCL) has proven effective in mitigating data sparsity and enhancing representation learning for recommendation. Yet, most GCL frameworks indiscriminately treat all non-anchor nodes as negatives during contrastive sampling, often leading to the false negative problem where semantically similar nodes are incorrectly repelled. Previous attempts to mitigate this issue rely on predetermined heuristics or local neighborhood mining, which struggle to reliably identify false negatives. More critically, they often overlook authentic user-item interactions for anchoring sample relationships. As a result, this paper presents MACRec, a Multi-View subspace-Alignment framework designed to Calibrate contrastive sampling in GCLbased Recommendation. MACRec comprises three core components: (1) a Multi-View Affinity (MVA) module that captures consistent semantic relations across multiple augmentations via self-expression modeling; (2) a Cross-Subspace Alignment (CSA) mechanism that leverages authentic useritem behavioral interactions to enforce semantic consistency across user and item subspaces; and (3) a Calibrationbased Contrastive Reweighting (CCR) strategy to dynamically down-weight potential false negatives during the contrastive learning process. Extensive experiments on three realworld benchmarks demonstrate that MACRec consistently improves performance across various augmentation backbones, achieving up to 14.55% relative gains.

In this paper, we introduce the IMPortance-awaReness maskIng NeTwork (IMPRINT), a novel approach to enhance the robustness of document retrieval systems against query variations, particularly those containing misspellings. Unlike previous models that treat all query components (words/features) equally, IMPRINT prioritizes the most important components while masking out less relevant ones. Specifically, we propose a Mutual Information-based measure to quantify component importance and integrate it into a dynamic masking mechanism that adjusts the retention probability of each component. Our method is evaluated on a combination of three benchmark datasets and three types of query variations. The experimental results show substantial performance gains compared to state-of-the-art models, achieving an average improvement of 1.2 absolute MRR@10 points in retrieval accuracy.

Graph Contrastive Learning (GCL) emerges as a powerful approach in recommendation systems, leveraging graph structures to learn effective representations. However, existing contrastive sampling strategies often introduce unintended biases, most notably, the misclassification of genuine positive samples as negatives, which undermines representation quality and overall recommendation performance. Accordingly, this paper revisits the conventional contrastive sampling and introduces Negative-Free Sampling for Graph Contrastive Learning (NFS). NFS adopts a two-stage sampling strategy that selectively identifies and utilizes only positive instances during training. By removing reliance on negative samples, it effectively mitigates misclassification bias and improves the semantic alignment between related representations. In addition, a comprehensive theoretical analysis is also provided to establish the robustness of NFS against representation collapse. Experimental results on three benchmarks demonstrate that NFS consistently outperforms or performs state-of-the-art methods, achieving up to a 14.2% relative improvement across evaluated datasets. In addition, a detailed ablation study is also provided to examine how exclusively leveraging positive samples contributes to the efficiency of GCL. The results further demonstrate the plug-and-play nature of the proposed method and its resilience to noisy data.

With the popularity of the digital human body, monocular three-dimensional (3D) face reconstruction is widely used in fields such as animation and face recognition. Although current methods trained using single-view image sets perform well in monocular 3D face reconstruction tasks, they tend to rely on the constraints of the a priori model or the appearance conditions of the input images, fundamentally because of the inability to propose an effective method to reduce the effects of two-dimensional (2D) ambiguity. To solve this problem, we developed an unsupervised training framework for monocular face 3D reconstruction using rotational cycle consistency. Specifically, to learn more accurate facial information, we first used an autoencoder to factor the input images and applied these factors to generate normalized frontal views. We then proceeded through a differentiable renderer to use rotational consistency to continuously perceive refinement. Our method provided implicit multi-view consistency constraints on the pose and depth information estimation of the input face, and the performance was accurate and robust in the presence of large variations in expression and pose. In the benchmark tests, our method performed more stably and realistically than other methods that used 3D face reconstruction in monocular 2D images.

Background The annotation of fashion images is a significantly important task in the fashion industry as well as social media and e-commerce. However, owing to the complexity and diversity of fashion images, this task entails multiple challenges, including the lack of fine-grained captions and confounders caused by dataset bias. Specifically, confounders often cause models to learn spurious correlations, thereby reducing their generalization capabilities. Method In this work, we propose the Deconfounded Fashion Image Captioning (DFIC) framework, which first uses multimodal retrieval to enrich the predicted captions of clothing, and then constructs a detailed causal graph using causal inference in the decoder to perform deconfounding. Multimodal retrieval is used to obtain semantic words related to image features, which are input into the decoder as prompt words to enrich sentence descriptions. In the decoder, causal inference is applied to disentangle visual and semantic features while concurrently eliminating visual and language confounding. Results Overall, our method can not only effectively enrich the captions of target images, but also greatly reduce confounders caused by the dataset. To verify the effectiveness of the proposed framework, the model was experimentally verified using the FACAD dataset.

Adversarial examples pose significant challenges for Natural Language Processing (NLP) model robustness, often causing notable performance degradation. While various detection methods have been proposed with the aim of differentiating clean and adversarial inputs, they often require fine-tuning with ample data, which is problematic for low-resource scenarios. To alleviate this issue, a Subspace Clustering based Adversarial Detector (termed SCAD) is proposed in this paper, leveraging a union of subspaces to model the clean data distribution. Specifically, SCAD estimates feature distribution across semantic subspaces, assigning unseen examples to the nearest one for effective discrimination. The construction of semantic subspaces does not require many observations and hence ideal for the low-resource setting.

Automated International Classification of Diseases (ICD) coding involves the automated assignment of diverse disease codes to clinical medical texts. It is considered as a multi-label classification task. Because most ICD codes are rare, the imbalanced distribution and small sample size issue make this task challenging. Inspired by the recent success of ChatGPT and prompt-based fine-tuning, this study proposes a model called ChatICD to address the issue of few-shot ICD coding. First, ChatGPT for data augumentation rephrases the descriptions of ICD codes into multiple samples. Then, ChatICD fine-tunes the pretrained model by generating prompt templates and label mapping words. We conduct an evaluation of ChatICD on benchmark datasets, namely MIMIC-III-50 and MIMIC-III-rare50. On the few-shot ICD coding task of MIMIC-III-rare50, ChatICD achieves macro-F1 and micro-F1 of 35.8% and 38.2% respectively, which is a 5.4% and 5.6% improvement over the current best model.

In recent years, deep learning-based software vulnerability detection methods have made significant progress. However, most existing methods focus on detecting vulnerabilities at the function-level or slice-level and cannot pinpoint the exact lines of code that cause the vulnerabilities. Program slicing can extract control and data dependency information from the code to assist deep learning models in detecting vulnerabilities. We propose a novel vulnerability detection model, PTLVD, which generates code gadgets(CGs) by slicing the program based on variables in the code, uses a transformer model for binary classification, and employs our proposed method Integrated Gradients Enhanced with Saliency(IGS) to locate the lines of code that are likely to cause vulnerabilities. IGS enhances the interpretability of the model by integrating the Integrated Gradients and Saliency methods. PTLVD employs an improved method of generating CGs to selectively remove irrelevant code statements, resulting in CGs that contain richer information and enhance the model’s performance. Additionally, during the preprocessing stage, PTLVD removes comments and standardizes code statements onto the same line, which effectively enhances the performance and vulnerability localization capabilities of the model. Experimental results show that, compared to state-of-the-art function-level and slice-level vulnerability detection models, PTLVD improves precision and F1 by 5.25% and 1.79%, respectively. In line-level prediction, Compared to the baseline method, PTLVD not only improved the Top-5 Accuracy by 1.61%, but also successfully reduced the Mean First Ranking by 5.08%.

Entities and relations extraction is one of the key task to build medical knowledge graph, which is of great significance to the development of medical artificial intelligence. However, overlapping triples are great challenge for biomedical entities and relations extraction. In order to improve the performance of biomedical entities and relations extraction, we propose a joint extraction of entities and relations method based on decomposition and recombination strategy to mine biomedical text. Our method decomposes entities and relations extraction task into three related sub-modules, which are entity tagging module, relation classification module and recombination matching module. Our main contributions are as follows: first, we introduce a decomposition and recombination end-to-end learning framework for joint entities and relations extraction. Second, we propose a bi-directional prediction method to deal with the overlapping triples problem. Finally, we propose the negative samples generation method to alleviate the error accumulation among these modules. The extensive experiments demonstrate that our method can improve the F1 score by 4.36%, 2.13% and 11.72% in ADE, DDI and BB biomedical corpus.

Image-based virtual try-on aims to transfer a clothes onto a person while preserving both person's and cloth's attributes. However, the existing methods to realize this task require a target clothes, which cannot be obtained in most cases. To address this issue, we propose a novel user-friendly virtual try-on network (UF-VTON), which only requires a person image and an image of another person wearing a target clothes to generate a result of the person wearing the target clothes. Specifically, we adopt a knowledge distillation scheme to construct a new triple dataset for supervised learning, propose a new three-step pipeline (coarse synthesis, clothing alignment, and refinement synthesis) for try-on task, and utilize an end-to-end training strategy to further refine the results. In particular, we design a new synthesis network that includes both CNN blocks and swin-transformer blocks to capture global and local information and generate highly-realistic try-on images. Qualitative and quantitative experiments show that our method achieves the state-of-the-art virtual try-on performance.

Context: Current image inpainting methods show great effects in different applications such as image editing, object removal, art creation and soon, but lack of editability of the inpainting results and convincing unsupervised features.Objective: To improve the existing methods, an optimized framework for image inpainting purpose is proposed based on hierarchical variational auto-encoder (VAE) as well as some optimization strategies.Method: Firstly, the VAE is used to extract the distribution of the features of the masked image in different scales, however, it will cause the distribution offset of extracted features which is unfavorable for image inpainting.Therefore, an optimal strategy that sampling the effective feature and invalid feature separately to avoid the offset of feature distribution of the masked image is integrated into the framework.To further improve the formulation of the proposed framework, the same encoder is used to realize the conversion from two domains to the same domain, which is a benefit to enhance the extraction of effective feature regions.In addition, we also introduce the cycle consistency constraints and GAN constraints into the framework to supervise the inpainting process.Result: Experimental results on the available image dataset demonstrate the effectiveness and superiority of the proposed framework.

The recognition of subway driver behavior is an important way for early warning of public safety. The current models of behavior recognition focus on action recognition of target objects in large-scene, which are difficult to apply for the subway driver behavior recognition directly because of space-time constraints. RepC3D model is proposed for recognizing subway driver behaviors in the paper. The model fuse the features of C3D model and RepVGG model. Firstly we preprocess the dataset by cutting the subway driver operation video into short videos, then the preprocessed dataset is adopted as the input of RepC3D model and is downsampled with the multiscale convolution layers of the main network VGG, which is used to extract the effective features of the driver's action behavior. Next, as the feature tranning network,RepC3D model identify and classfy the behaviors of the subway driver from the videos. The experimental result shows that the RepC3D model is btteetter than the C3D model and RepVGG model in terms of recognition accuracy, false detection rate, and missed detection rate, the recognition efficiency is also improved. The dataset is available at https://github.com/wtazyy/Datasets.git.

The approaches based on laboratory experiments to explore the interactions between viruses and their hosts are costly and time-consuming. Due to advances in high-throughput technologies, recent computational methods to predict virus-host interaction have attracted increasing attention. But these methods couldn’t effectively utilize the heterogeneous information of the virus-host interaction network. In this paper, we propose a computational method to predict potential virus-host interaction via network fusion and graph mining, named VHINFGM. Different from existing methods, VHINFGM constructs two different heterogeneous networks from the existing interaction network through similarity network fusion and graph embedding technique. Then, VHINFGM introduces two kinds of meta-path scores to extract features from each heterogeneous graph. Based on this graph mining approach, a mixed feature vector for two heterogeneous networks can be obtained, which can be used as the input of a classifier to predict potential interactions. VHINFGM is verified on four datasets, it can be found that VHINFGM outperforms the state-of-the-art methods. 5 out of the top 10 virus-host interaction reported by VHINFGM has been validated in the biological experiments. Besides, VHINFGM predicts 5 new virus-host relationships, which could guide further research.

Image-based virtual try-on systems with the goal of transferring a target clothing item onto the corresponding region of a person have received great attention recently. However, it is still a challenge for the existing methods to generate photo-realistic try-on images while preserving non-target details(Fig. 1). To resolve this issue, we present a novel virtual try-on network, DP-VTON. First, a clothing warping module combines pixel transformation with feature transformation to transform the target clothing. Second, a semantic segmentation prediction module predicts a semantic segmentation map of the person wearing the target clothing. Third, an arm generation module generates arms of the reference image that will be changed after try-on. Finally, the warped clothing, semantic segmentation map, arms image and other non-target details (e.g. face, hair, bottom clothes) are fused together for try-on image synthesis. Extensive experiments demonstrate our system achieves the state-of-the-art virtual try-on performance both qualitatively and quantitatively.1