Aiming at the problem of reliability degradation caused by the increase of the density and capacity of flash memory,an optimization scheme of refresh and copy combination strategy based on LDPC read delay is proposed.In general,the original strategy is to add LDPC code module to flash memory and use hard and soft decoding to correct data errors.The traditional refresh strategy is based on the original strategy,when the LDPC soft decoding fails to correct the error,the refresh strategy is used to correct the error.The scheme is based on the characteristics of LDPC soft decoding 7 quantitative levels,and takes this as the judgment condition,using the method of analysis and comparison to determine that the condition of refresh is level 3,the condition of the copy is level 5,and the two methods are reasonably applied in the LDPC soft decoding mode.Compared with the previous two strategies,the average response time of flash memory is reduced,and the read performance of flash memory is improved to a certain extent.Simulation is performed on the extended platform of the simulator disksim+ssd,and experimental results show that,the average response time of this scheme is 10% shorter than that of the original strategy,and its flash memory lifetime is prolonged compared to traditional refresh strategies.

Disease diagnosis prediction aims to use electronic health data to model disease progression patterns and predict the future health status of patients,and is widely used in assisting clinical decision-making,healthcare services and other fields.In order to further explore the valuable information in the medical records,a disease diagnosis prediction algorithm based on contrastive learning is proposed.Contrastive learning provides self-supervised training signals for the model by measuring the similarity between samples,which can improve the information capture ability of the model.The proposed algorithm excavates the common knowledge between similar patients through contrastive training,and enhances the ability of the model to learn patient representations.In order to capture more comprehensive common information,the information of similar groups of the target patient is further explored as auxiliary information to characterize the health status of the target patient.Experimental results on the public dataset show that compared with the Retain,Dipole,LSAN and GRASP algorithms,the proposed algorithm improves AUROC and AUPRC of the readmission prediction task by more than 2.9% and 8.1% respectively,and Recall@10 and MAP@10 of the diagnosis prediction task by 2.1% and 1.8%,respectively.

Anomaly detection is a hot topic that has been widely studied in the field of machine learning and plays an important role in industrial production,food safety,disease monitoring,etc.The latest anomaly detection methods mostly jointly train semi-supervised detection models based on a small number of available labeled samples and many unlabeled samples.However,these existing semi-supervised anomaly detection models mostly use deep learning frameworks.Due to the lack of enough feature information on low-dimensional data sets,it is difficult to learn accurate data boundaries,resulting in insufficient detection perfor-mance.To solve this problem,a dually encoded semi-supervised anomaly detection(DE-SAD)model is proposed.DE-SAD can make full use of a small amount of available labeled data and a large amount of unlabeled data for semi-supervised learning,and learn more accurate implicit manifold distribution of normal data through the dually encoded stage constraint,thus effectively magnifying the gap between normal data and abnormal data.DE-SAD shows excellent ano-maly detection performance on multiple anomaly detection datasets from different fields,especially on low-dimensional data,and its AUROC is up to 4.6% higher than the current state-of-the-art methods.

Event-based social network(EBSN) is a new kind of complex heterogeneous social network,the personalized event re-commendation in it has certain application value.In recent years,with the rapid development of EBSN,the problem of information overload for event recommendation has been solved by data mining technology.However,it will reduce the accuracy of event re-commendation by only using a single feature attribute or a small number of linear combinations for calculation,and predefining fixed weights.In addition,most approaches ignore the influence of user feedback information on subsequent recommendation.Aiming at the above problems,an event recommendation method fusing multi-factor features is proposed,which consists of two phases.In the query preprocessing phase,the events,historical users and their relationships in EBSN are abstracted as a directed he-terogeneous graph,and the feature information of nodes and edges is extracted for auxiliary storage.A relatively small candidate set is obtained by filtering invalid nodes and edges with the auxiliary data.According to the query context,the query semantics are transformed into the query graphs.In the online query phase,it combines the characteristics of potential friends,event-based collaborative filtering and users’ interests to recommend,and also receives feedback from users on whether they accept the event as a reference factor for subsequent recommendations.Large number of experiments on real datasets and simulated datasets verify the accuracy and user satisfaction of the proposed method in EBSN event recommendation.

Meta-learning methods have been introduced into recommendation algorithms in recent years to alleviate the problem of cold start.The existing meta-learning algorithms can only improve the ability of the algorithm to deal with a set of statically distributed data sets(tasks).When faced with multiple data sets subject to non-stationary distribution,the existing models often have negative knowledge transfer and catastrophic forgetting problems,resulting in a significant decline in algorithm recommendation performance.This paper explores a recommendation algorithm based on variational continuous Bayesian Meta-learning(VC-BML).Firstly,the algorithm assumes that the meta parameters follow the dynamic mixed Gaussian model,which makes it have a larger parameter space,improves the ability of the model to adapt to different tasks,and alleviates the problem of negative knowledge transfer.Then,the number of task clusters in VC-BML is flexibly determined by Chinese restaurant process(CRP),which enables the model to store knowledge of different task distributions in different mixed components and invoke this know-ledge when similar tasks occur,helping to alleviate the catastrophic forgetting problem in traditional algorithms.To estimate the posterior probabilities of model parameters,a more robust structured variational reasoning method is used to approximate the posterior values to avoid forgetting knowledge.Finally,VC-BML outperforms the baselines on all four datasets with non-stationary distributions.Compared with point-based estimation methods,VC-BML improves the robustness of the model and helps alle-viate the catastrophic forgetting problem.

High dimensional data is often encountered in many data analysis tasks.Feature selection techniques aim to find the most representative features from the original high-dimensional data.Due to the lack of class label information,it is much more difficult to select suitable features in unsupervised learning scenarios than in supervised scenarios.Traditional unsupervised feature selection methods usually score the features of samples according to certain criteria in which samples are treated indiscriminately.However,these approaches cannot capture the internal structure of data completely.The importance of different samples should vary.There is a dual relationship between weight of sample and feature that will influence each other.Therefore,an unsupervised feature selection algorithm based on dual manifold re-ranking(DMRR) is proposed in this paper.Different similarity matrices are constructed to depict the manifold structures on samples and samples,features and features,and samples and features respectively.Then manifold re-ranking is carried out by combining the initial scores of samples and features.By comparing DMRR with three original unsupervised feature selection algorithms and two unsupervised feature selection post-processing algorithms,experimental results verify that importance information of different samples and the dual relationship between sample and feature are helpful to achieve better feature selection.

The phenomenon of missing data occurs frequently during the acquisition and transfer of data,and improper handling of missing data sets can adversely affect subsequent data mining efforts.In order to fill the missing data set more effectively,a method for data imputation based on compressed sensing is proposed for correlation data.First,the problem of missing data imputation is transformed into a sparse vector recovery problem under the compressed sensing framework.Second,a specialized sparse representation base is constructed for correlation data,so the data sparsity can be better realized.Finally,the fast iterative weighted thresholding algorithm(FIWTA) is proposed,which is refined based on the fast iterative shrinkage-thresholding algorithm (FISTA).The proposed algorithm adopts a new iterative weighted method and introduces a restart strategy,which greatly improves the convergence of the algorithm and the reconstruction accuracy of the data.Simulation results show that the proposed algorithm is able to fill the missing data efficiently,and both the reconstruction success rate and the reconstruction speed are improved compared with the traditional fast iterative shrinkage-thresholding algorithm.Meanwhile,even when the sparse transformation of the data is less effective,imputation of missing data sets can still be accomplished with better robustness.

The existing algorithms based on symmetric nonnegative matrix factorization(SymNMF) are mostly rely on initial data to construct affinity matrices,and neglect the limited pairwise constraints,so these methods are unable to effectively distinguish similar samples of different categories or learn the geometric features of samples.To solve the above problems,this paper proposes a block sparse symmetric nonnegative matrix factorization based on constrained graph regularization(CGBS-SymNMF).Firstly,the constrained graph matrix is constructed by prior information,which is used to guide the clustering indicator matrix to distinguish different clusters of samples with high similarity.Secondly,pairwise constraint propagation by semidefinite programming(PCP-SDP) is introduced to learn a new sample graph mapping matrix by using pairwise constraints.Finally,a dissimilarity matrix is constructed by cannot-link constraints,which is used to guide a block sparse regular term for enhancing the anti-noise capability of the model.Experimental results demonstrate a higher clustering accuracy and stability of the proposed algorithm.

With the technological development of intelligent transportation system and the surging spatio-temporal data in urban,the demand for public services is increasingly emphasized.As a vital part of urban transportation,public transportation also faces enormous challenges,and the spatio-temporal prediction task in transportation network is the core of the solutions for various traffic problems.Mobility pattern in traffic can reflect the travel behaviors of people and their rules.In most studies on traffic prediction task,the importance of mobility pattern is neglected.In view of the problem of existing work,a multi-pattern traffic prediction framework,MPGNNFormer,is proposed,in which based-graph neural network deep clustering method is used to extract mobility patterns of stations,and a Transformer-based spatio-temporal prediction model is designed to learn temporal dependence and spatial dependence of stations and to improve the computational efficiency.Then,a series of experiments are conducted on real bus dataset for evaluation and testing,including analysis of mobility patterns and comparison of prediction results.Finally,experimental results prove the efficacy of proposed method in the short and long-term traffic prediction of traffic networks,and its sca-lability is discussed.