Recommendation algorithms, as an extremely important information filtering and pushing tool, have been widely applied in many fields such as e-commerce, social networks, and online audio-visual entertainment. In research in recent years, multimodal recommendation algorithms that integrate multimodal information of items (such as text, images, etc.) and users' historical interaction information have achieved good results. However, most of the existing multimodal recommendation algorithms adopt the decision - level fusion method, this approach makes it difficult to capture the complementary information between modalities. Moreover, the few recommendation algorithms that adopt feature - level fusion have also failed to achieve good recommendation results. In view of this, we propose a feature-level fusion method based on cross-attention mechanism to capture the correlations between various modalities. In addition, in order to alleviate the problems of data sparsity and weak self-supervised signals faced by graph representation learning, we introduce graph contrastive learning into the model. At the same time, considering the long-tail problem of data and in order to avoid the uneven learning representation caused by the difference in graph structure between head data and tail data in traditional graph enhancement methods, we propose a long-tail graph enhancement method. We have carried out a large number of experiments on two real datasets and compared with other advanced baseline models. The results show that our model has been significantly improved in multiple evaluation indicators.

The prevalence of Alzheimer's disease (AD) is rising annually, imposing a severe burden on patients and society. Therefore, assisted AD assessment is crucial. The decline in language function and the cognitive impairment it reflects are key external manifestations of AD. Many studies have utilized speech analysis to achieve convenient, non-invasive, and low-cost AD detection. Although state-of-the-art researches achieve high-precision AD detection using multimodal information, these studies often ignore interactions between different modalities and lack explanations for complex models. To address this, we propose a multi-task learning (MTL) AD assessment model that combines hybrid attention with multimodal representations. The model fuses audio, text, and expert knowledge to fully capture intra- and inter-modal interactions, achieving simultaneous AD detection and cognitive state prediction, along with comprehensive explainability analyses of the model and various modalities. Results show that the proposed method is sufficiently sensitive in assessing AD, achieving 89.58% accuracy and 91.67% recall for the classification task and a root mean square error of 4.31 for the regression task with good generalization performance. Multimodal representations with expert knowledge and MTL contribute to AD assessment performance. Explainability analyses indicate that, compared to healthy controls, AD patients exhibit slower speech rates, reduced syntactic complexity, and a greater tendency to use pause fillers and pronouns. Therefore, Our study validates the effectiveness of the proposed method, addressing trust issues in clinical practice for assisted decision-making and further advancing the development of speech as a promising biomarker for early AD screening and cognitive decline monitoring.

In medical image analysis, visual-language models facilitate zero-shot diagnosis by assessing the similarity between images and textual disease descriptions. However, challenges such as multi-center data variability and fine-grained lesion recognition constrain their full potential. To address these limitations, we propose a method that combines spatial domain adaptive preprocessing of images with the knowledge distillation capabilities of ChatGPT4o. Specifically, we developed a Spatial Domain Adaptive Correction module for multi-center medical images. This module analyzes the spatial pixel distribution and applies adaptive enhancement strategies to optimize image quality, thereby restoring details and enhancing edge features. Second, disease description texts are distilled using ChatGPT-4o and subsequently revised by radiology experts. These texts are further enhanced with lesion location and morphology information derived from large-scale chest imaging reports, and guiding words are added to emulate the case report style. Finally, text and image features are extracted using separate encoders, and prediction probabilities for various disease types are calculated based on cosine similarity. We validated our method on seven datasets containing three unseen lung diseases. The results demonstrate superior performance and significant clinical application potential in diagnosing unseen and rare diseases. The experimental code has been open-sourced at https://github.com/wtf633/Zero-Shot-Diagnosis-of-Unseen-Pulmonary-Diseases.

Overlapping community structure is a significant feature of large-scale complex networks. Some existing community detection algorithms cannot be applied to large-scale complex networks due to their high time or space complexity. Label propagation algorithms were proposed for detecting communities in large-scale networks because of their linear time complexity, however most of which can only detect non-overlapping communities, or the results are inaccurate and unstable. Aimed at the defects, we proposed an improved overlapping community detection algorithm, LPANNI (Label Propagation Algorithm with Neighbor Node Influence), which detects overlapping community structures by adopting fixed label propagation sequence based on the ascending order of node importance and label update strategy based on neighbor node influence and historical label preferred strategy. Extensive experimental results in both real networks and synthetic networks show that, LPANNI can significantly improve the accuracy and stability of community detection algorithms based on label propagation in large-scale complex networks. Meanwhile, LPANNI can detect overlapping community structures in large-scale complex networks under linear time complexity.