Vol. 43 · No. 1

Federated learning facilitates collaborative data analysis among multiple participants while preserving user privacy. However, conventional federated learning approaches, typically employing weighted average techniques for model fusion, confront two significant challenges: 1. The inclusion of malicious models in the fusion process can drastically undermine the accuracy of the aggregated global model. 2. Due to the heterogeneity problem of devices and data, the number of client samples does not determine the weight value of the model. To solve those challenges, we propose a trustworthy model fusion method based on reinforcement learning (FedDRL), which includes two stages. In the first stage, we propose a reliable client selection mechanism to exclude malicious models from the fusion process. In the second stage, we propose an adaptive model fusion method that dynamically assigns weights based on model quality to aggregate the best global models. Finally, we validate our approach against five distinct model fusion scenarios, demonstrating that our algorithm significantly enhanced reliability without compromising accuracy.

Stock prediction is prominent in the field of Artificial Intelligence. Stock prediction problems are handled either as a regression or classification task. Studies in the literature have also shown success for hybrid learning to stock prediction. But little attention is paid to finding out the effect of spatial feature extraction/distortion over the temporal effect of the deep neural network and vice versa for the problem under study. The paper, therefore, proposes a hybrid long short-term memory (LSTM) network over a convolutional neural network (CNN) called LSTM-CNN as against the popular CNN-LSTM model. The daily price movement of the S & P 500 index data is utilized. A sliding window technique is considered to obtain a balanced data of 20-days window data from the S & P 500. The proposed stock prediction model is investigated further for an optimal set of hyperparameters using the Bayesian optimization (Bo) technique. In addition, the proposed model is compared with optimized CNN, LSTM, and CNN-LTSM models. The optimized LSTM-CNN model is found to outperform the other models with accuracy, precision, and recall values of 0.9741, 0.9684, and 0.9800, respectively. The proposed model is established to provide a better stock trend prediction.

Cloud Computing has immersed researchers in accessing the resources on-demand for deploying various applications by offering infinite services. But, as the demand for cloud resources is dynamic, it significantly affects the load on the system. Thus, this research emphasizes deploying a dynamic and autonomic load prediction framework. This paper proposes an Ensemble Time-Series Approach for Load Prediction (ETSA-LP), which integrates various time-series analysis techniques for predicting CPU and memory utilization. To evaluate the efficiency of the proposed approach, a series of experiments on Google and PlanetLab traces have been conducted in a real Cloud environment. The results were compared according to different performance metrics and models, the accuracy determined and the minimal error rate selected as the best among others. The proposed ensemble approach gives the best performance over the existing models showing the remarkable accuracy improvement and reducing the error rate and execution time.

With the increasing diversity and complexity of online social networks, effectively dividing communities presents a growing challenge. These networks are characterized by their large scale, sparse structure, and numerous isolated points. Traditional community detection methods lack consideration of node attribute information, thereby negatively impacting the accuracy of community detection. To address these challenges, this paper presents a novel Louvain-FTAS community detection algorithm that integrates topology and attribute structure. The proposed algorithm first selects attributes with positive effects to account for attribute heterogeneity. Subsequently, it utilizes a semi-local strategy to calculate topology similarity and information entropy to calculate attribute similarity. These values are combined to obtain the final node similarity matrix, which is then fed into the Louvain algorithm to maximize modularity and incorporate multi-dimensional attribute features to enhance community detection accuracy. The proposed model is evaluated through comparative experiments on two real datasets and artificial synthetic networks, demonstrating its rationality and effectiveness.

The detection of carbon deposit degree is of great significance to the maintenance of automobile engine. Due to issues with poor feature aggregation, inter-class similarity, and intra-class variance in carbon deposit data with a small number of samples, model-based discriminative approaches cannot be widely implemented in the market. In order to overcome this technical barrier, the article examines the impact of DCNNs (Deep Convolutional Neural Networks) level on the recognition effect of the degree of carbon deposit, introduces a dropout structure and data enhancement strategy to lower the risk of overfitting brought on by the small dataset, and suggests a recognition method based on the kernel of dual-dimensional multiscale-multifrequency information features to enhance the differentiation characteristic. After experimental testing, the accuracy of this method is 86.9

Process mining has received much attention in the field of business process management. Event logs that are generated from information systems can be correlated with the process models for conformance checking. The process models describe event activities at an abstraction level. However, hierarchical business processes, as a kind of typical complex process scenario, describe sub-processes invocation and multi-instantiation patterns. As existing conformance, checking approaches cannot identify sub-processes within hierarchical process models, they cannot be used for conformance checking of hierarchical process models. To handle this limitation, a definition of hierarchically alignment sequences is presented in this paper. Meanwhile, a novel conformance checking approach for hierarchical process models and event logs is proposed. The proposed method has been implemented within the ProM toolkit, which is an open-source process mining software. To evaluate the effectiveness of the proposed approach, both artificial and real-world event logs are utilized in a comparative analysis against existing state-of-the-art approaches.

A key problem domain inside Robotic Process Automation is the automatic discovery of workflow process schemes. Considering current process mining technologies, graph-based approaches dominate the industry. On the other hand, the conventional methods suffer from low time efficiency and varying accuracy. Machine learning-based methods can provide better efficiency, but they have significant limitations considering schema flexibility. The paper presents a novel neural network-based schema induction model for the discovery of event patterns containing parallel and optional sequences of different actors. This model can process more complex event graphs and situations than the conventional methods. The performed analysis and test results show the unique power of this approach in process schema mining.

Class imbalance problem in datasets can lead to biased classification decisions in favor of majority class samples. Additionally, class overlap can cause fuzzy classification boundaries, affecting the performance of classification algorithms. To address these issues, we propose a hybrid sampling method based on conditional tabular generative adversarial network (CTGAN) and K-nearest neighbor (KNN) algorithm. Firstly, we introduce an oversampling algorithm, named DB-CTGAN, based on CTGAN. This algorithm filters noisy and boundary samples using the density-based spatial clustering of applications with noise (DBSCAN) clustering algorithm and generates synthetic samples that conform to the real data distribution using CTGAN. Finally, we combine the expanded fraudulent samples generated by DB-CTGAN with the normal samples and use the KNN overlap undersampling algorithm to remove the samples in the overlap region, solving the class overlap problem. Experimental results show that compared with eight sampling methods using four standard classification models (Random Forest, Decision Tree, Support Vector Classification, and XGBoost), the proposed method significantly improves the F1, AUC, and G-mean metrics on five real datasets.

In order to improve the effect of intelligent fusion recommendation under the background of social network, this paper combines the fuzzy perception algorithm to research the intelligent fusion recommendation algorithm of social network. Moreover, this paper proposes a task offloading scheme that relies on V2V communication to utilize idle computing resources in a ``resource pool''. In addition, this paper formulates the computational task execution time as a min-max problem to reduce the storage overhead to optimize the total task execution time. Numerical results show that the proposed scheme greatly reduces the task execution time. The introduced particle swarm optimization algorithm also proves the convergence speed and accuracy of the optimization problem. The research verifies that the intelligent fusion recommendation algorithm for social network based on fuzzy perception has good social network data fusion effect and can effectively improve the effect of intelligent fusion recommendation.