International Journal of Machine Learning and Cybernetics最新文献

Federated learning is a distributed machine learning paradigm. Traditional federated learning performs well on the premise that all clients have the same learning ability or similar learning tasks. However, resource and data heterogeneity are inevitable among clients in real-world scenarios, leading to the situation that existing federated learning mechanisms cannot achieve high accuracy in short response time. In this study, an effective federated learning framework with cross-resource client collaboration, termed CEFL, is proposed to coordinate clients with different capacities to help each other, efficiently and adequately reflecting collective intelligence. Clients are categorized into different clusters based on their computational resources in the hierarchical framework. Resource-rich clusters use their knowledge to assist resource-limited clusters converge rapidly. Once resource-limited clusters have the ability to mentor others, resource-rich clusters learn from the resource-limited clusters in their favor to improve their own effectiveness. A cloud server provides tailored assistance to each cluster with a personalized model through an adaptive multi-similarity metric, in order for each cluster to learn the most useful knowledge. The experiments fully demonstrate that the proposed method not only has superior accuracy with significantly reduced latency but also improves the convergence rate compared to other state-of-the-art federated learning methods in addressing the problem of resource and data heterogeneity.

To address challenges in the digital economy’s landscape of digital intelligence, large language models (LLMs) have been developed. Improvements in computational power and available resources have significantly advanced LLMs, allowing their integration into diverse domains for human life. Medical LLMs are essential application tools with potential across various medical scenarios. In this paper, we review LLM developments, focusing on the requirements and applications of medical LLMs. We provide a concise overview of existing models, aiming to explore advanced research directions and benefit researchers for future medical applications. We emphasize the advantages of medical LLMs in applications, as well as the challenges encountered during their development. Finally, we suggest directions for technical integration to mitigate challenges and potential research directions for the future of medical LLMs, aiming to meet the demands of the medical field better.

In past decade, the android malware threats have been rapidly increasing with the widespread usage of internet applications. In respect of security purpose, there are several machine learning techniques attempted to detect the malwares effectively, but failed to achieve the accurate detection due to increasing number of features, more time consumption decreases in detection efficiency. To overcome these limitations, in this research work an innovative Hybrid dual path Bidirectional long short-term memory Kepler dynamic graph Convolutional Network (HBKCN) is proposed to analyze and detect android malwares effectively. First, the augmented abstract syntax tree is applied for pre-processing and extracts the string function from each malware. Second, the adaptive aphid ant optimization is utilized to choose the most appropriate features and remove irrelevant features. Finally, the proposed HBKCN classifies benign and malware apps based on their specifications. Four benchmark datasets, namely Drebin, VirusShare, Malgenome -215, and MaMaDroid datasets, are employed to estimate the effectiveness of the technique. The result demonstrates that the HBKCN technique achieved excellent performance with respect to a few important metrics compared to existing methods. Moreover, detection accuracies of 99.2%, 99.1%,99.8% and 99.8% are achieved for the considered datasets, respectively. Also, the computation time is greatly reduced, illustrating the efficiency of the proposed model in identifying android malwares.

The key to achieving an accurate and reliable traffic flow prediction lies in modeling the complex and dynamic correlations among sensors. However, existing studies ignore the fact that such correlations are influenced by multiple dynamic factors and the original sequence features of the traffic data, which limits the deep modeling of such correlations and leads to a biased understanding of such correlations. The extraction strategies for global features are less developed, which has degraded the reliability of the predictions. In this study, a novel multi-dynamic residual graph convolutional network with global feature enhancement is proposed to solve these problems and achieve an accurate and reliable traffic flow prediction. First, multiple graph generators are proposed, which fully preserve the original sequence features of the traffic data and enable layered depth-wise modeling of the dynamic correlations among sensors through a residual mechanism. Second, an output module is proposed to explore extraction strategies for global features, by employing a residual mechanism and parameter sharing strategy to maintain the consistency of the global features. Finally, a new layered network architecture is proposed, which not only leverages the advantages of both static and dynamic graphs, but also captures the spatiotemporal dependencies among sensors. The superiority of the proposed model has been verified through extensive experiments on two real-world datasets.

This paper aims to build the SVIHRD model for COVID-19 and it also simultaneously conduct stability and numerical analysis on the transmission of COVID-19. Here we do a mathematical analysis for the SVIHRD model, which involves positivity, boundedness, uniqueness, and proving both global and local stability. In the process of numerical simulation, we use real-world data for COVID-19 cases in Japan. An important feature presents in this paper, is that we replace the usual numerical solving technique for obtaining the parameters with a Physics Informed Neural Network (PINN). This PINN needs an order of time instances as input and the number of Susceptible (S), Vaccinated (V), Infected (I), Hospitalized (H), Recovered (R), and Death (D) people per time instances to learn specific parameters of the model using loss functions. We developed three different PINN setups-the baseline model, configuration-I, and configuration-II-to explore and optimize these parameters for modeling COVID-19 dynamics in Japan. During the validation process, we evaluated how well the learned parameters from these three PINN architectures predicted real infection data for the next two months. The baseline model, with four hidden layers and 32 neurons each, performed well with an (R^{2}) value of 0.8038 and a Wilcoxon signed-rank test p value of 0.001556, closely matching actual infection data. A sensitivity analysis of the baseline model’s parameters showed that the vaccination rate (sigma) is the most sensitive.

Industrial applications and optimization problems in reality often involve multiple objectives. Due to the high dimensionality of objective space in many-objective optimization problems (MaOPs), the ability of traditional evolution operators to search the optimal region and generate promising offspring sharply decreases. Besides, as the number of objectives increases, it becomes difficult to balance the convergence and diversity of the population. Considering all these facts, this paper proposes a mutation individual position detection strategy. It estimates both individual fitness and diversity contributions, and assigns appropriate positions to individuals in the mutation operator through individual ranking. Then, by introducing an external population to adjust the weight vectors, its maintenance process takes into account the matching information between the population and the weight vectors. By comparing five representative algorithms, numerical experiments have shown that the algorithm can obtain a well distributed final solution set on optimization problems of various objective scales. Moreover, it also demonstrates advantages in generating excellent offspring individuals and balancing the overall performance of the population. In summary, the algorithm has competitiveness in solving MaOPs.

Knowledge graph (KG) based recommender systems have shown promise in improving accuracy and interpretability. They reveal the intrinsic relationships of knowledge through the associations and paths between entities for personalized recommendations. However, existing approaches do not adequately consider the high-order connections between neighboring nodes in the relational graph, resulting in a lack of sufficient capture of structured information. In this paper, we propose a knowledge-enhanced recommendation model via dynamic co-attention and high-order connectivity (DCHC) to address this issue. First, we construct a hybrid graph by aligning users and items in the user-item bipartite graph with entities in the KG. As a result, we are able to simultaneously consider the interaction between users and items as well as the entity information in the KG, thereby gaining a more comprehensive understanding of user behavior and interests. Second, we explicitly model the high-order connections between entities through the hybrid structured graphs in an end-to-end manner. Therefore, we not only explored the complex interactive relationships between entities but also ensured the accurate capture of structural information in the graph. Third, we employ a dynamic co-attention mechanism to enhance the representation of users and items, effectively exploiting the potential correlation between them. We therefore effectively exploited the potential correlation between users and items and successfully integrating these relationships into their representations. Extensive experiments conducted on three benchmarks demonstrate that DCHC outperforms state-of-the-art KG-based recommendation methods.

Most Siamese algorithms take little account of the information interaction between the target and search region, leading to tracking results that are often disturbed by the cumulative effect of target-like distractors between layers. In this paper, we propose a reciprocal interlayer-temporal discriminative target model for robust visual tracking. Firstly, an interlayer target-aware enhancement model is constructed, which establishes pixel-by-pixel correlation between the template and search region to achieve interlayer feature information interaction. This alleviates the cumulative error caused by the blindness of the target to search region during feature extraction, enhancing target perception. Secondly, to weaken the impact of interference, a temporal interference evaluation strategy is designed. It utilizes the interframe candidate propagation module to build associations among multi-candidates in the current frame and the previous frame. Then, the similar distractors are eliminated by using object inference constraint, so as to obtain a more accurate target location. Finally, we integrate the interlayer target-aware enhancement model and temporal interference evaluation strategy into the Siamese framework to achieve reciprocity for robust target tracking. Experimental results show that our tracking approach performs well, especially on seven benchmark datasets, including OTB-100, TC-128, DTB, UAV-123, VOT-2016, VOT-2018 and GOT-10k.

Product Knowledge Graphs (PKGs) play a crucial role in enhancing e-commerce system performance by providing structured information about entities and their relationships, such as complementary or substitutable relations between products or product types, which can be utilized in recommender systems. However, relation labeling in PKGs remains a challenging task due to the dynamic nature of e-commerce domains and the associated cost of human labor. Recently, breakthroughs in Large Language Models (LLMs) have shown surprising results in numerous natural language processing tasks, especially in the in-context learning (ICL). In this paper, we conduct an empirical study of LLMs for relation labeling in e-commerce PKGs, investigating their powerful learning capabilities in natural language and effectiveness in predicting relations between product types with few-shot in-context learning. We evaluate the performance of various LLMs, including PaLM-2, GPT-3.5, and Llama-2, on benchmark datasets for e-commerce relation labeling tasks. We use different prompt engineering techniques to examine their impact on model performance. Our results show that LLMs can achieve competitive performance compared to human labelers using just 1–5 labeled examples per relation. We also illustrate the bias issues in LLMs towards minority ethnic groups. Additionally, we show that LLMs significantly outperform existing KG completion models or classification methods in relation labeling for e-commerce KGs and exhibit performance strong enough to replace human labeling. Beyond empirical investigations, we also carry out a theoretical analysis to explain the superior capability of LLMs in few-shot ICL by comparing it with kernel regression.

While deep learning has made significant progress in many applications including fault diagnosis, its relatively high computational cost and long training time seriously limits its applicability in some areas. To address these challenges, lightweight neural networks, such as the randomly weighted networks like the random vector functional link (RVFL) with a non-iterative training mechanism, have been proposed. In the RVFL model, the initialization of weights plays a crucial role in determining model performance. Therefore, this paper investigates the impact of different random parameter distributions on RVFL model performance in bearing fault diagnosis. Specifically, we propose a weight generation strategy that approximately follows uniform or normal distributions, and through a case study, we compare the effects of these distributions on the model. Subsequently, we conduct an experimental analysis on a publicly available bearing anomaly detection dataset. The experimental results demonstrate that the choice of distribution affects the model’s accuracy, with the normal distribution showing slightly better performance than the uniform distribution in this application scenario. These findings provide some guidelines for selecting appropriate parameter distributions for bearing fault diagnosis using RVFL networks.