Vol. 45 · No. 1

To address issues such as data sparsity, overfitting, and the inability to fully extract latent information in traditional recommendation methods, a new insurance product recommendation algorithm is proposed that combines knowledge graphs improved by matrix factorization with deep neural networks (DNN). First, to tackle the issue of sparse data in existing insurance products, an improved knowledge graph recommendation algorithm based on matrix factorization (FunkSVD) is proposed. By training the data through matrix factorization the problem of data sparsity is mitigated. After refining the customer-product matrix, knowledge graph triples are constructed based on customer characteristics and insurance product features such as insured age and coverage. Feature extraction and customer preference prediction are carried out using an alternating learning approach within a multi-task knowledge graph framework. Then, to alleviate issues like local optima and vanishing gradients during recommendation, DNN is applied for further recommendation. A fully connected layer is constructed, and forward propagation and backpropagation algorithms are used to train customer features and product matrices, predicting customer purchasing behavior and generating recommendations. Finally, comparative experimental results show that, compared to other recommendation algorithms such as collaborative filtering and DNN, the proposed algorithm improves accuracy, recall, F1 score, and other metrics. This algorithm not only speeds up recommendations but also improves recommendation quality.

Coreference resolution, the task of identifying expressions in a text that refer to the same entity, is a critical component in various natural language processing applications. This paper presents a novel end-to-end neural coreference resolution system, utilizing the CorefUD 1.1 dataset, which spans 17 datasets across 12 languages. The proposed model is based on the standard end-to-end neural coreference resolution system. We first establish baseline models, including monolingual and cross-lingual variations, and then propose several extensions to enhance performance across diverse linguistic contexts. These extensions include cross-lingual training, incorporation of syntactic information, a Span2Head model for optimized headword prediction, and advanced singleton modeling. We also experiment with headword span representation and long-document modeling through overlapping segments. The proposed extensions, particularly the heads-only approach, singleton modeling, and long document prediction, significantly improve performance across most datasets. We also perform zero-shot cross-lingual experiments, highlighting the potential and limitations of cross-lingual transfer in coreference resolution. Our findings contribute to the development of robust and scalable coreference systems for multilingual coreference resolution. Finally, we evaluate our model on the CorefUD 1.1 test set and surpass the best model from the CRAC 2023 shared task of comparable size by a large margin.

Deep generative models are widely used in recommendation systems because of their ability to deal with uncertainty by learning inherent data distribution. Among deep generative models, Generative Adversarial Networks (GAN) perform well in recommendation tasks. However, existing Collaborative Filtering (CF) recommendation algorithms based on GAN generally have problems such as mode collapse and training instability, which are further aggravated by sparse and noisy recommendation data. To solve these problems, a Collaborative Diffusion Generative Adversarial Networks (CDGAN) framework for recommendation systems is proposed in this paper. Specifically, CDGAN framework is mainly composed of three parts: feature encoder, diffusion generator, and self-attention discriminator. The feature encoder extracts the rating information and side information to obtain potential feature vectors to alleviate the data sparsity problem. The diffusion generator simulates the complex nonlinear mode of the user-item interaction matrix through forward diffusion and reverse denoising to reconstruct the user-item interaction matrix to alleviate the problem of data noise. The self-attention discriminator obtains the user's specific behaviors and preferences through the self-attention mechanism to improve the discriminator's discriminating ability. Furthermore, a corresponding CDGAN recommendation algorithm is designed based on our proposed CDGAN framework. Comprehensive experiments are conducted on three real-world recommendation datasets. The experimental results indicate that, when compared with multiple representative recommendation models, the proposed CDGAN model achieves superior performance in the evaluation metrics Precision and Recall on all datasets, thereby proving its effectiveness.

Accurate load forecasting is the premise of efficient and stable operation of integrated energy systems. For multiple integrated energy systems that have insufficient energy consumption data but similar energy consumption behavior, federated learning can establish a higher accuracy multi-load forecasting model for each system without disclosing data privacy. However, the existing federated learning methods cannot fully utilize common and individual characteristics in the energy consumption data of different nodes (that is, integrated energy systems), which obviously affects their prediction accuracy. In view of this, we propose a multi-time scale multi-load federated forecasting method based on mode decomposition (MD-MMFF). Firstly, a multivariate empirical mode decomposition method is introduced to decompose the energy consumption data of each node into two types, i.e., regular components and irregular components. Each node uses the local LSTM model to learn the regular components and predict their outputs. For the irregular components, a multi-load federated forecasting training mechanism based on knowledge distillation is proposed, and the corresponding multi-load forecasting model is jointly established for each node. Then, the predicted values of regular components and irregular components are integrated to obtain the final multivariate load forecasting results. Experimental results show that compared with the existing multi-load forecasting algorithms, the proposed MD-MMFF method can obtain higher accuracy multi-source load forecasting results.

Human emotion plays a critical role in the purpose of communication, with the correct detection of emotion, one can analyze the human feeling without even asking, as it is non-verbal communication method. Human emotion through facial expression is one of the highlighted topics of study because of its wide range of application from robotics, security, artificial intelligence, marketing and health monitoring where human-computer interaction (HCI) is a key ingredient. To tackle the challenges that arise in accurate human facial emotion recognition, a modified version of the EfficientNet-B3 model is proposed, which shows promising results for emotion detection using the FER-2013 dataset. By leveraging the universal FER-2013 dataset, which contains more than 35 000 grey-scale human facial images, the proposed model aims to improve recognition performance. The modified architecture incorporates EfficientNet-B3 as the base model, with an additional batch normalization layer followed by a dense layer and an output layer. The model has been trained on the facial dataset while considering these crucial factors. Performance evaluation of the model is conducted using confusion matrix, precision, recall, accuracy, and F1 score as performance metrics. Remarkably, the proposed model achieved an impressive accuracy of 92

Cloud computing primarily focuses on providing secure and reliable access to geographically distributed resources. Cloud computing is a parallel and distributed system that simplifies the virtualization of distributed computing. Job scheduling is one of the most important services in cloud computing. Job scheduling is used to schedule users' jobs to allocate suitable resources in a cloud environment. Recently, several efficient job scheduling algorithms have been proposed for cloud computing. The purpose of designing these algorithms is to achieve time and cost optimization for job execution. On the other hand, it is necessary to provide a model to assess the reliability of the different job scheduling algorithms. Unfortunately, no accurate and complete model has been proposed to evaluate the reliability of these algorithms. Scheduling failures are inevitable in cloud computing due to the harsh deployment environment, resource migration, non-compliance time or cost, etc. Therefore, providing a reliability model is necessary for successful scheduling algorithms. In this paper, a reliability model based on a discrete-time Markov chain is incorporated for the famous scheduling algorithms MCT, MET, BCO, HCOC, and PPO. We propose an approach to evaluating the reliability of scheduling algorithms based on time and cost constraints. The proposed approach is modeled by a Discrete Time Markov Chain, and the results show the reliability of scheduling algorithms in different states.

Information systems play an important role in realizing business processes. Process mining uses event logs produced by information systems to construct the models of business processes. Then, the business process model can be used to verify and improve business processes. To represent a changing business process, some structures of the process model need to be replaced with other structures. However, existing model repair methods do not consider replacing the structures of the process model. This work proposes a logical Petri net-based approach to repair models based on the order relations among activities. First, we construct two relation sets of an event log and a process model, respectively. Comparing the two relation sets, the differences between a process model and an event log are collected in a deviation set. According to the deviation set, the model is repaired by constructing concurrent structures. Finally, we perform some experiments to illustrate the effectiveness and correctness of the proposed approach. The results show that the proposed method produces a more precise and simpler repaired model compared to state-of-the-art methods.

To address the issue of low recall rates in detecting small faces within crowded scenes, this paper conducts an analysis of the primary reasons behind this challenge and introduces a real-time face detection system named CF-YOLO (Crowded-Face-YOLO). The study identifies a crucial factor contributing to this problem, which is the insufficient provision of positive samples for small faces during the training phase by conventional face detectors. To tackle this limitation, a Sa-SimOTA strategy is proposed to enhance the availability of positive samples for small targets. Additionally, in the post-processing stage, the utilization of the non-maximum suppression (NMS) algorithm for assigning optimal bounding boxes to detected faces is discussed. The traditional fixed threshold employed in the NMS algorithm for decision-making often results in the loss of small face detection boxes in crowded scenarios. To alleviate this issue, a Soft-Face-NMS algorithm is introduced, which incorporates facial feature variables into the Soft-NMS algorithm for weighted processing, facilitating the selection of face boxes with higher confidence in overlapping regions. Furthermore, to augment the feature extraction capabilities of the YOLO backbone, an EMA+ attention module is proposed, and modifications are made to the network structure of YOLOv7 to enhance the extraction of more effective features conducive to small face detection. The proposed model demonstrates impressive accuracy rates of 97.3

Event logs of business systems are often dirty owing to recording errors or system conventions, and this dirty data can definitely affect the quality of the event logs. Consequently, many anomalous behaviors that deviate from pre-established process models are frequently seen in event logs. To properly assess and enhance business processes, it is imperative to identify and correct such anomalous behaviors. When the missing or anomalous location is known, the current repair methods have a high success rate; however, when the missing or anomalous location is unknown, they do not function well. Therefore, a formalized solution using an activity feature graph is proposed for event logs where uncertainty arises from missing, duplicate, and replaced occurrences. Several prediction and repair models are used in conjunction with the viewpoints of control flow and data flow to identify and address anomalous behaviors in event logs. In addition to fixing exceptions at the case level, the proposed repairing approach can also correct exceptions at the attribute level, or feature repairing. Process analysts are provided with an activity feature graph to examine and refine the repair outcomes. Finally, four event logs with various anomaly ratios are used to confirm the method's viability.

Privacy-preserving process mining (PPPM) is a technology that analyses and optimises processes while safeguarding sensitive information. Earlier research on PPPM mainly focused on the data perspective, employing techniques such as noise insertion and data generalisation to protect the sensitive personal information of process executors. However, these studies overlooked the behavioural relationships between activities in the process. Attackers can exploit domain knowledge and certain behavioural information of executors to carry out re-identification attacks, leading to the leakage of personal sensitive information. To address this issue, a behaviour anonymous method of business process based on a log skeleton is proposed. This method begins with the individual process model of the executor, clustering based on the similarity between models and utilising K-anonymity and the log skeleton technology to achieve cluster division and the construction of behaviour constraint sets. Furthermore, the generation of privacy models is standardised by the behaviour constraint set to achieve global behaviour anonymization. To evaluate the effectiveness of this method, experiments were conducted using multiple real and synthetic datasets. The experimental results indicate that this method significantly outperforms the comparison methods.