Research project conflict risk detection aims to distinguish potential conflicts from multiple aspects including personnel allocation, resource distribution, and content overlap, etc. Existing research on project conflict detection faces three critical challenges: the difficulty in capturing semantic information from unstructured project documentation, limited model expressiveness in processing heterogeneous project relationships, and challenges in modeling diverse conflict patterns across interconnected project elements. To address these challenges, we first construct a research project heterogeneous graph with three types of nodes (i.e., projects, personnel, and resources) and five types of relationships capturing various project interactions. Besides, we propose a novel Semantic-augmented Heterogeneous grAph neural network for project conflict risk DEtection (SHADE) framework, equipped with three specially designed modules: 1) a semantic-augmented module leveraging large language models to extract fine-grained content representations from project descriptions, 2) a multi-view risk detection module with adaptive fusion to capture conflicts from diverse perspectives, and 3) a self-supervised contrastive learning module to enhance the discriminative power between positive and negative patterns. Extensive experiments on the real-world research project heterogeneous graph demonstrate that our proposed framework SHADE significantly outperforms state-of-the-art methods.

Sensitive content detection in financial reimbursement documents is fundamental to corporate compliance management. Traditional keyword-based approaches suffer from high false-positive rates due to their inability to capture contextual semantics. In this paper, we propose TripletAudit, a context-aware framework for sensitive content detection that leverages triplet learning, incorporating a hierarchical architecture combining BERT for semantic embedding, Bi-Lstm for sequential modeling, and multi-head attention for context understanding. Our triplet learning mechanism effectively distinguishes between sensitive and non-sensitive contexts through anchor-positive-negative samples. Experiments on real-world reimbursement datasets demonstrate that TripletAudit achieves state-of-the-art performance with 94.03% accuracy and 93.24% F1-score, significantly outperforming baseline methods in financial compliance scenarios.

In recent years, single-label intent recognition has faced significant challenges in handling short and semantically sparse user inputs. Traditional methods often treat intent labels as flat categories, neglecting their inherent hierarchical relationships and limiting model performance. To address these issues, we propose HiIntent , a novel zero-shot intent detection framework that integrates hierarchical semantic modeling with a collaborative generation-discriminative mechanism. HiIntent first constructs a structured label hierarchy through a two-stage process: a large language model (LLM) generates semantic abstractions of intent labels, which are then evaluated and refined by a discriminative module to ensure coherence and correctness. This is followed by a similarity-driven convergence strategy that enhances intra-class consistency and inter-class separability using multi-metric similarity calculations. Finally, a contrastive prompt construction method leverages the learned label hierarchy to generate enriched semantic descriptions for each intent, improving representation learning and facilitating accurate classification even in zero-shot scenarios. Extensive experiments on both general-purpose (CLINC-150) and domain-specific (RFMR) datasets demonstrate that HiIntent consistently outperforms existing approaches across multiple evaluation metrics. Ablation studies and hyperparameter analyses further validate the effectiveness of each component in the proposed framework.

Graph anomaly detection has attracted considerable interest due to the wide use of graph structure data. Several GNN-based anomaly detection methods discover anomalies through the powerful node representation ability of GNNs. However, real-world graphs are typically rarely labeled, which leads these deep learning methods to face the challenge of under-fitting. A fundamental question here is: can anomalies in graphs be detected with few annotations? In this paper, we propose a novel semi-supervised anomaly detection method in graphs based on Consistency Exploitation (ConsE). First, ConsE adopts a consistency-based neighbor sampler, which ensures the consistency of a central node and its neighbors on attributes and categories during the aggregation process through attribute similarity and soft pseudo-labels. Afterward, ConsE encourages the consistency of node representations generated by a dedicated neighbor sampler and a generic neighbor sampler to improve its robustness in complex neighborhoods. Experimental results on real-world datasets demonstrate that our model significantly outperforms several state-of-the-art baseline methods.

BACKGROUND: Cancer subtype classification is helpful for personalized cancer treatment. Although, some approaches have been developed to classifying caner subtype based on high dimensional gene expression data, it is difficult to obtain satisfactory classification results. Meanwhile, some cancers have been well studied and classified to some subtypes, which are adopt by most researchers. Hence, this priori knowledge is significant for further identifying new meaningful subtypes. RESULTS: In this paper, we present a combined parallel random forest and autoencoder approach for cancer subtype identification based on high dimensional gene expression data, ForestSubtype. ForestSubtype first adopts the parallel RF and the priori knowledge of cancer subtype to train a module and extract significant candidate features. Second, ForestSubtype uses a random forest as the base module and ten parallel random forests to compute each feature weight and rank them separately. Then, the intersection of the features with the larger weights output by the ten parallel random forests is taken as our subsequent candidate features. Third, ForestSubtype uses an autoencoder to condenses the selected features into a two-dimensional data. Fourth, ForestSubtype utilizes k-means++ to obtain new cancer subtype identification results. In this paper, the breast cancer gene expression data obtained from The Cancer Genome Atlas are used for training and validation, and an independent breast cancer dataset from the Molecular Taxonomy of Breast Cancer International Consortium is used for testing. Additionally, we use two other cancer datasets for validating the generalizability of ForestSubtype. ForestSubtype outperforms the other two methods in terms of the distribution of clusters, internal and external metric results. The open-source code is available at https://github.com/lffyd/ForestSubtype . CONCLUSIONS: Our work shows that the combination of high-dimensional gene expression data and parallel random forests and autoencoder, guided by a priori knowledge, can identify new subtypes more effectively than existing methods of cancer subtype classification.