Constructing digital twin water conservancy construction knowledge graph to mine the potential relationships between water conservancy construction objects can help the relevant personnel to optimize the water conservancy construction design scheme and decision-making process. Aiming at the interdisciplinary and complex knowledge structure of digital twin water conservancy construction， and the problems such as insufficient learning and low extraction accuracy of knowledge of general knowledge extraction models in water conservancy domain， a Digital Twin water conservancy construction Knowledge Extraction method based on Large Language Model （DTKE-LLM） was proposed to improve the accuracy of knowledge extraction. In this method， by deploying local Large Language Model （LLM） through LangChain and integrating digital twin water conservancy domain knowledge， prompt learning was used to fine-tune the LLM. In the LLM， semantic understanding and generation capabilities were utilized to extract knowledge. At the same time， a heterogeneous entity alignment strategy was designed to optimize the entity extraction results. Comparison experiments and ablation experiments were carried out on the water conservancy domain corpus to verify the effectiveness of DTKE-LLM. Results of the comparison experiments demonstrate that DTKE-LLM outperforms the deep learning-based BiLSTM-CRF （Bidirectional Long Short-Term Memory Conditional Random Field） named entity recognition model and the general Information extraction model UIE （Universal Information Extraction） in precision. Results of the ablation experiments show that compared with the ChatGLM2-6B （Chat Generative Language Model 2.6 Billion）， DTKE-LLM has the F1 scores of entity extraction and relation extraction improved by 5.5 and 3.2 percentage points respectively. It can be seen that the proposed method realizes the construction of digital twin water conservancy construction knowledge graph on the basis of ensuring the quality of knowledge graph construction.

An infrared small target method based on information compensation was proposed to address the problem that infrared small targets are prone to losing texture detail information during network iteration， which decreased accuracy of target localization and contour segmentation. Firstly， Image Feature Extraction （IFE） module was used to encode shallow details and deep semantic features of infrared image. Secondly， a Multi-level Information Compensation （MIC） module was constructed to perform information compensation to down-sampled features in the encoding stage by aggregating features from adjacent levels. Thirdly， Global Target Response （GTR） module was introduced to compensate the limitation of convolutional locality by incorporating global contextual information of feature map. Finally， Asymmetric Cross-Fusion （ACF） module was constructed to fuse shallow and deep features， thereby preserving texture and positional information during target decoding， thus achieving detection of infrared small targets. Experimental results of training and testing on publicly available NUAA-SIRST （Nanjing University of Aeronautics and Astronautics-Single-frame InfraRed Small Target） and NUDT-SIRST （National University of Defense Technology-Single-frame InfraRed Small Target） mixed datasets show that compared to methods such as UIUNet （U-Net in U-Net Network）， LSPM （Local Similarity Pyramid Modules）， and DNANet （Dense Nested Attention Network）， the proposed method achieves improvements of 9.2， 8.9， and 5.5 percentage points in Intersection over Union （IoU）， respectively， and 6.0， 5.4， and 3.1 percentage points in F1-Score， respectively. The above demonstrates that the proposed method enables accurate detection and effective segmentation of small targets in complex infrared background images.

To obtain more accurate molecular toxicity prediction results， a molecular toxicity prediction model based on meta Graph Isomorphism Network （GIN） was proposed， namely Meta-MTP. Firstly， graph isomorphism neural network was used to obtain molecular characterization by using atoms as nodes， bonds as edges， and molecules as graph structures. The pre-trained model was used to initialize the GIN to obtain better parameters. A feedforward Transformer incorporating layer-wise attention and local enhancement was introduced. Atom type prediction and bond prediction were used as auxiliary tasks to extract more internal molecular information. The model was trained through a meta learning dual-level optimization strategy. Finally， the model was trained using Tox21 and SIDER datasets. Experimental results on Tox21 and SIDER datasets show that Meta-MTP has good molecular toxicity prediction ability. When the number of samples is 10， compared to FSGNNTR （Few-Shot Graph Neural Network-TRansformer） model in all tasks， the Area Under the ROC Curve （AUC） of Meta-MTP is improved by 1.4% and 5.4% respectively. Compared to three traditional graph neural network models， Graph Isomorphism Network （GIN）， Graph Convolutional Network （GCN）， and Graph Sample and AGgrEgate （GraphSAGE）， the AUC of Meta-MTP improves by 18.3%-23.7% and 7.3%-22.2% respectively.

Aiming at the problem of large numbers of mutants leading to high mutation testing cost， a Program Dependency based Mutant Generation （PDMG） strategy was proposed to select the mutation implementation objects satisfying certain constraint conditions for mutation generation. Firstly， program dependency graphs were generated based on data dependencies and control dependencies. Then， based on the mutation object selection strategy and program dependency graphs， the dependency statements were selected as mutation objects. Finally， the mutation operator was injected to the selected mutation objects in order to generate mutants. The proposed method was applied to mutation testing of 8 benchmark test programs. Experimental results show that compared with Random Selection （RS） and Mutation Operator Selection （MOS） strategies， PDMG strategy can reduce the mutants by 52.20% on average， improving the execution efficiency of mutation testing without reducing the effectiveness of mutation testing.

Air quality data， as a typical spatio-temporal data， exhibits complex multi-scale intrinsic characteristics and has abrupt change problem. Concerning the problem that existing air quality prediction methods perform poorly when dealing with air quality prediction tasks containing large amount of abrupt change， a Multi-Granularity abrupt Change Fitting Network （MACFN） for air quality prediction was proposed. Firstly， multi-granularity feature extraction was first performed on the input data according to the periodicity of air quality data in time. Then， a graph convolution network and a temporal convolution network were used to extract the spatial correlation and temporal dependence of the air quality data， respectively. Finally， to reduce the prediction error， an abrupt change fitting network was designed to adaptively learn the abrupt change part of the data. The proposed network was experimentally evaluated on three real air quality datasets， and the Root Mean Square Error （RMSE） decreased by about 11.6%， 6.3%， and 2.2% respectively， when compared to the Multi-Scale Spatial Temporal Network （MSSTN）. The experimental results show that MACFN can efficiently capture complex spatio-temporal relationships and performs better in the task of predicting air quality that is prone to abrupt change with a large magnitude of variability.

For the issue of existing models’ inability to accurately identify intricate and diverse patterns of gang fraud， a new practical credit card fraud detection model based on complex transaction graph was proposed. Firstly， the association transaction graph was constructed based on the original transaction information of the users， then the graph Transformer neural network module was employed to mine the gang fraud characteristics directly from the transaction network without cumbersome feature engineering. Finally， the high-precision detection of fraud transactions was realized by jointly optimizing the topological features and sequential transaction features by the fraud detection network. The credit card anti-fraud experiment results showed that the proposed model outperformed seven benchmark models in all evaluation indexes. The Average-Precision （AP） improved by 20% and the Area Under the ROC Curve （AUC） increased by an average of 2.7% over the best benchmark Graph Attention Network （GAT） model in transaction fraud detection tasks. These results indicate that the proposed model is effective in the detection of credit card fraud transactions.

Person search is one of the important research directions in the field of computer vision. Its research goal is to detect and identify characters in uncropped image libraries. In order to deeply understand the person search algorithms， a large number of related literature were summarized and analyzed. First of all， according to the network structure， the person search algorithms were divided into two categories： two-step methods and end-to-end one-step methods. The key technologies of the one-step methods， feature learning and measurement learning， were analyzed and introduced. The datasets and evaluation indicators in the field of person search were discussed， and the performance comparison and analysis of the mainstream algorithms were given. The experimental results show that， although the two-step methods have good performance， most of them have high calculation costs and take long time； the one-step methods can solve the two sub-tasks pedestrian detection and person re-identification， in a more efficient learning framework and achieve better results. Finally， the person search algorithms were summarized and their future development directions were prospected.

Regular 3-SATisfiability （3-SAT） problem is an NP-hard problem. Studying alterations in the cluster structure of solutions to regular 3-SAT problem is to enhance the comprehension of the difficulty involved in problem determination and distribution of satisfiable solutions. However， existing analysis models only study a few discrete values near the cluster phase transition point. Under different constraint densities， there is lack of unified analysis model to describe the structural evolution of solution clusters. To solve this problem， a Phase transition Model of Solution cluster Structure （PMSS） was proposed. The main idea of this model is to obtain an initial solution of regular 3-SAT problem using WalkSAT algorithm and information propagation algorithm， construct a solution cluster of initial solutions by using random walks， and analyze the solution cluster. Modularity and community were used to measure community structure， and structural entropy was used to measure the complexity of the solution cluster structure. Experimental results show that PMSS can accurately analyze the structural evolution process of solution clusters， and the phase transition point of regular 3-SAT problem instances is between 13 and 14， which is consistent with the phase transition point obtained using Zchaff solver， further verifying the effectiveness of PMSS.

SATisfiability problem （SAT） is a NP complete problem， which is widely used in artificial intelligence and machine learning. Exact SATisfiability problem （XSAT） is an important subproblem of SAT. Most of the current research on XSAT is mainly at the theoretical level， but few efficient solution algorithms are studied， especially the stochastic local search algorithms with efficient verifiability. To address above problems and analyze some properties of both basic and equivalent coding formulas， a stochastic local search algorithm WalkXSAT was proposed for solving XSAT directly. Firstly， the random local search framework was used for basic search and condition determination. Secondly， the appropriate unsatisfiable scoring value of the text to which the variables belonged was added， and the variables that were not easily and appropriately satisfied were prioritized. Thirdly， the search space was reduced using the heuristic strategy of anti-repeat selection of flipped variables. Finally， multiple sources and multiple formats of examples were used for comparison experiments. Compared with ProbSAT algorithm， the number variable flips and the solving time of WalkXSAT are significantly reduced when directly solving the XSAT. In the example after solving the basic encoding transformation， when the variable size of the example is greater than 100， the ProbSAT algorithm is no longer effective， while WalkXSAT can still solve the XSAT in a short time. Experimental results show that the proposed algorithm WalkXSAT has high accuracy， strong stability， and fast convergence speed.

The privacy policy document declares the privacy information that an application needs to obtain， but it cannot guarantee that it clearly and fully discloses the types of privacy information that the application obtains. Currently， there are still deficiencies in the analysis of the consistency between actual sensitive behaviors of applications and privacy policies. To address the above issues， a method for analyzing the consistency between sensitive behaviors and privacy policies of Android applications was proposed. In the privacy policy analysis stage， a Bi-GRU-CRF （Bi-directional Gated Recurrent Unit Conditional Random Field） neural network was used and the model was incrementally trained by adding a custom annotation library to extract key information from the privacy policy declaration. In the sensitive behavior analysis stage， IFDS （Interprocedural， Finite， Distributive， Subset） algorithm was optimized by classifying sensitive API （Application Programming Interface） calls， deleting already analyzed sensitive API calls from the input sensitive source list， and marking already extracted sensitive paths. It ensured that the analysis results of sensitive behaviors matched the language granularity of the privacy policy description， reduced the redundancy of the analysis results and improved the efficiency of analysis. In the consistency analysis stage， the semantic relationships between ontologies were classified into equivalence， subordination， and approximation relationships， and a formal model for consistency between sensitive behaviors and privacy policies was defined based on these relationships. The consistency situations between sensitive behaviors and privacy policies were classified into clear expression and ambiguous expression， and inconsistency situations were classified into omitted expression， incorrect expression， and ambiguous expression. Finally， based on the proposed semantic similarity-based consistency analysis algorithm， the consistency between sensitive behaviors and privacy policies was analyzed. Experimental results show that， by analyzing 928 applications， with the privacy policy analysis accuracy of 97.34%， 51.4% of Android applications are found to have inconsistencies between the actual sensitive behaviors and the privacy policy declaration.

Aiming at the issues in the existing prediction methods， such as only predicting b and y backbone fragment ions， as well as single model's difficulty in capturing the complex relationships within peptide sequences， a theoretical tandem mass spectrometry prediction method for peptide sequences based on Transformer and Gated Recurrent Unit （GRU）， named DeepCollider， was proposed. Firstly， through self-attention mechanism and long-distance dependencies， the deep learning architecture combining Transformer and GRU was used to enhance the modeling ability of relationship between peptide sequences and fragment ion intensities. Secondly， unlike the existing methods encoding peptide sequences to predict all b and y backbone ions， fragmentation flags were utilized to mark fragmentation sites within peptide sequences， thereby enabling the encoding of fragment ions at specific fragmentation sites and prediction of the corresponding fragment ions. Finally， Pearson Correlation Coefficient （PCC） and Mean Absolute Error （MAE） were employed as evaluation metrics to measure the similarity between predicted spectrometry and experimental spectrometry. Experimental results demonstrate that DeepCollider shows advantages in both PCC and MAE metrics compared to the existing methods limited to predicting b and y backbone fragment ions， such as pDeep and Prosit methods， with an increase of 0.15 in PCC value and a decrease of 0.005 in MAE value. It can be seen that DeepCollider not only predicts b， y backbone ions and their corresponding dehydrated and deaminated neutral loss ions， but also further improves the peak coverage and similarity of theoretical spectrometry prediction.

Multi-frame self-supervised monocular depth estimation constructs a Cost Volume （CV） based on the relationship between current frame and the previous frame， serving as an additional input source for the monocular depth estimation network. This approach provides a more accurate description of the temporal and spatial structure of scene videos. However， the cost volume becomes unreliable in the presence of dynamic objects or untextured regions in the scene. Overreliance on the unreliable information within the cost volume leads to a decrease in depth estimation accuracy. To tackle the issue of unreliable information in the cost volume， a multi-frame fusion module was designed to reduce the weights of unreliable information sources dynamically and mitigate the impact of unreliable information sources on the network. Besides， to handle the negative impact of unreliable information sources in cost volume on network training， a network was designed to guide the training of the depth estimation network， preventing the depth estimation network from overly depending on unreliable information. The proposed method achieves excellent performance on KITTI dataset， with absolute relative error， squared relative error， and Root Mean Square Error （RMSE） decreased by 0.015， 0.094， and 0.200， respectively， compared to the benchmark method Lite-Mono. In comparison to similar methods， the proposed method not only has higher precision， but also requires fewer computational resources. The proposed network structure makes full use of the advantages of multi-frame training， while avoiding the defects of multi-frame training （i.e.， the influence of cost volume uncertainty on the network）， and improves the model precision effectively.

Unmanned Aerial Vehicle （UAV） swarm path planning and task allocation are the cores of UAV swarm rescue applications. However， traditional methods solve path planning and task allocation separately， resulting in uneven resource allocation. In order to solve the above problem， combined with the physical attributes and application environmental factors of UAV swarm， the Ant Colony Optimization （ACO） was improved， and a Joint Parallel ACO （JPACO） was proposed. Firstly， the pheromone was updated by the hierarchical pheromone enhancement coefficient mechanism to improve the performance of JPACO task allocation balance and energy consumption balance. Secondly， the path balance factor and dynamic probability transfer factor were designed to optimize the ant colony model， which is easy to fall into local convergence， so as to improve the global search capability of JPACO. Finally， the cluster parallel processing mechanism was introduced to reduce the time consumption of JPACO operation. JPACO was compared with Adaptive Dynamic ACO （ADACO）， Scanning Motion ACO （SMACO）， Greedy Strategy ACO （GSACO） and Intersecting ACO （IACO） in terms of optimal path， task allocation balance， energy consumption balance and operation time on the open dataset CVRPLIB. Experimental results show that the average value of the optimal paths of JPACO is 7.4% and 16.3% lower than of IACO and ADACO respectively in processing small-scale operations. Compared with GSACO and ADACO， JPACO has the solution time reduced by 8.2% and 22.1% in large-scale operations. It is verified that JPACO can improve the optimal path when dealing with small-scale operations， and is obviously superior to the comparison algorithms in terms of task allocation balance， energy consumption balance， and operation time consumption when processing large-scale operations.

To address the problem of difficulty in identifying user plagiarism in social networks and to protect the rights of original authors while holding users accountable for plagiarism actions， a plagiarism identification scheme for social network users under blockchain was proposed. Aiming at the lack of universal tracing model in existing blockchain， a blockchain-based traceability information management model was designed to record user operation information and provide a basis for text similarity detection. Based on the Merkle tree and Bloom filter structures， a new index structure BHMerkle was designed. The calculation overhead of block construction and query was reduced， and the rapid positioning of transactions was realized. At the same time， a multi-feature weighted Simhash algorithm was proposed to improve the precision of word weight calculation and the efficiency of signature value matching stage. In this way， malicious users with plagiarism cloud be identified， and the occurrence of malicious behavior can be curbed through the reward and punishment mechanism. The average precision and recall of the plagiarism detection scheme on news datasets with different topics were 94.8% and 88.3%， respectively. Compared with multi-dimensional Simhash algorithm and Simhash algorithm based on information Entropy weighting （E-Simhash）， the average precision was increased by 6.19 and 4.01 percentage points respectively， the average recall was increased by 3.12 and 2.92 percentage points respectively. Experimental results show that the proposed scheme improves the query and detection efficiency of plagiarism text， and has high accuracy in plagiarism identification.

Aiming at the forgetting and underutilization of the text information of image in image captioning methods， a Scene Graph-aware Cross-modal Network （SGC-Net） was proposed. Firstly， the scene graph was utilized as the image’s visual features， and the Graph Convolutional Network （GCN） was utilized for feature fusion， so that the visual and textual features were in the same feature space. Then， the text sequence generated by the model was stored， and the corresponding position information was added as the textual features of the image， so as to solve the problem of text feature loss brought by the single-layer Long Short-Term Memory （LSTM） Network. Finally， to address the issue of over dependence on image information and underuse of text information， the self-attention mechanism was utilized to extract significant image information and text information and fuse then. Experimental results on Flickr30K and MS-COCO （MicroSoft Common Objects in COntext） datasets demonstrate that SGC-Net outperforms Sub-GC on the indicators BLEU1 （BiLingual Evaluation Understudy with 1-gram）， BLEU4 （BiLingual Evaluation Understudy with 4-grams）， METEOR （Metric for Evaluation of Translation with Explicit ORdering）， ROUGE （Recall-Oriented Understudy for Gisting Evaluation） and SPICE （Semantic Propositional Image Caption Evaluation） with the improvements of 1.1，0.9，0.3，0.7，0.4 and 0.3， 0.1， 0.3， 0.5， 0.6， respectively. It can be seen that the method used by SGC-Net can increase the model’s image captioning performance and the fluency of the generated description effectively.

Current deep multi-view clustering methods have the following shortcomings： 1） When feature extraction is carried out for a single view， only attribute information or structural information of the samples is considered， and these two types of information are not integrated. Thus， the extracted features cannot fully represent latent structure of the original data. 2） Feature extraction and clustering were divided into two separated processes， without establishing the relationship between them， so that the feature extraction process cannot be optimized by the clustering process. To solve these problems， a Deep Fusion based Multi-view Clustering Network （DFMCN） was proposed. Firstly， the embedding space of each view was obtained by combining autoencoder and graph convolution autoencoder to fuse attribute information and structure information of samples. Then， the embedding space of the fusion view was obtained through weighted fusion， and clustering was carried out in this space. And in the process of clustering， the feature extraction process was optimized by a two-layer self-supervision mechanism. Experimental results on FM （Fashion-MNIST）， HW （HandWritten numerals）， and YTF （YouTube Face） datasets show that the accuracy of DFMCN is higher than those of all comparison methods； and DFMCN has the accuracy increased by 1.80 percentage points compared with the suboptimal CMSC-DCCA （Cross-Modal Subspace Clustering via Deep Canonical Correlation Analysis） method on FM dataset， the Normalized Mutual Information （NMI） of DFMCN is increased by 1.26 to 14.84 percentage points compared to all methods except for CMSC-DCCA and DMSC （Deep Multimodal Subspace Clustering networks）. Experimental results verify the effectiveness of the proposed method.

To solve the problem that information hiding algorithms based on neural style transfer do not solve the embedding problem of color images， a color image information hiding algorithm based on style transfer process was proposed. Firstly， the advantages of feature extraction of Convolutional Neural Network （CNN） were utilized to extract the semantic information of the carrier image， the style information of the style image and the feature information of the color image， respectively. Then， the semantic content of images and different styles were fused together. Finally the embedding of color image was completed while performing the style transfer of the carrier image through the decoder. Experimental results show that the proposed algorithm can integrate the secret image into the generated stylized image effectively， making the secret information embedding behavior indistinguishable from the style change behavior. Under the premise of maintaining the security of the algorithm， the proposed algorithm has the hiding capacity increased to 24 bpp， and the average values of Peak Signal-to-Noise Ratio （PSNR） and Structural SIMilarity （SSIM） reached 25.29 dB and 0.85 respectively， thereby solving the color image embedding problem effectively.

Concerning of the curse of dimensionality caused by too high data dimension and redundant information， a high-dimensional Attribute Reduction algorithm based on Similarity and Difference Matrix （ARSDM） was proposed. In this algorithm， on the basis of discernibility matrix， the similarity measure for samples in the same class was added to form a comprehensive evaluation of all samples. Firstly， the distances of samples under each attribute were calculated， and the similarity of same class and the difference of different classes were obtained based on these distances. Secondly， a similarity and difference matrix was established to form an evaluation of the entire dataset. Finally， attribute reduction was performed， i.e.， each column of the similarity and difference matrix was summed， the feature with the largest value was selected into the reduction in proper order， and the row vector of the corresponding sample pair was set to the zero vector. Experimental results show that compared with the classical attribute reduction algorithms DMG （Discernibility Matrix based on Graph theory）， FFRS （Fitting Fuzzy Rough Sets） and GBNRS （Granular Ball Neighborhood Rough Sets）， the average classification accuracy of ARSDM is increased by 1.07， 6.48， and 8.92 percentage points respectively under the Classification And Regression Tree （CART） classifier， and increased by 1.96， 11.96， and 12.39 percentage points under the Support Vector Machine （SVM） classifier. At the same time， ARSDM outperforms GBNRS and FFRS in running efficiency. It can be seen that ARSDM can effectively remove redundant information and improve the classification accuracy.

By using complex pre-training targets and a large number of model parameters， Pre-Training Model （PTM） can effectively obtain rich knowledge from unlabeled data. However， the development of the multimodal PTMs is still in its infancy. According to the difference between modals， most of the current multimodal PTMs were divided into the image-text PTMs and video-text PTMs. According to the different data fusion methods， the multimodal PTMs were divided into two types： single-stream models and two-stream models. Firstly， common pre-training tasks and downstream tasks used in validation experiments were summarized. Secondly， the common models in the area of multimodal pre-training were sorted out， and the downstream tasks of each model and the performance and experimental data of the models were listed in tables for comparison. Thirdly， the application scenarios of M6 （Multi-Modality to Multi-Modality Multitask Mega-transformer） model， Cross-modal Prompt Tuning （CPT） model， VideoBERT （Video Bidirectional Encoder Representations from Transformers） model， and AliceMind （Alibaba’s collection of encoder-decoders from Mind） model in specific downstream tasks were introduced. Finally， the challenges and future research directions faced by related multimodal PTM work were summed up.

Collaborative Filtering （CF） algorithm can realize personalized recommendation on the basis of the similarity between items or users. However， data sparsity has always been one of the challenges faced by CF algorithm. In order to improve the prediction accuracy， a CF algorithm based on Collaborative Training and Boosting （CFCTB） was proposed to solve the problem of sparse user-item scores. First， two CFs were integrated into a framework by using collaborative training， pseudo-labeled samples with high confidence were added to each other’s training set by the two CFs， and Boosting weighted training data were used to assist the collaborative training. Then， the weighted integration was used to predict the final user scores， and the accumulation of noise generated by pseudo-labeled samples was avoided effectively， thereby further improving the recommendation performance. Experimental results show that the accuracy of the proposed algorithm is better than that of the single models on four open datasets. On CiaoDVD dataset with the highest sparsity， compared with Global and Local Kernels for recommender systems （GLocal-K）， the proposed algorithm has the Mean Absolute Error （MAE） reduced by 4.737%. Compared with ECoRec （Ensemble of Co-trained Recommenders） algorithm， the proposed algorithm has the Root Mean Squared Error （RMSE） decreased by 7.421%. The above rasults verify the effectiveness of the proposed algorithm.

Considering the lack of effective trend feature descriptors in existing methods， financial technical indicators such as Vertical Horizontal Filter （VHF） and Moving Average Convergence/Divergence （MACD） were introduced into power data analysis. An anomaly detection algorithm and a load forecasting algorithm using financial technical indicators were proposed. In the proposed anomaly detection algorithm， the thresholds of various financial technical indicators were determined based on statistics， and then the abnormal behaviors of user power consumption were detected using threshold detection. In the proposed load forecasting algorithm， 14 dimensional daily load characteristics related to financial technical indicators were extracted， and a Long Shot-Term Memory （LSTM） load forecasting model was built. Experimental results on industrial power data of Hangzhou City show that the proposed load forecasting algorithm reduces the Mean Absolute Percentage Error （MAPE） to 9.272%， which is lower than that of Autoregressive Integrated Moving Average （ARIMA）， Prophet and Support Vector Machine （SVM） algorithms by 2.322， 24.175 and 1.310 percentage points， respectively. The results show that financial technical indicators can be effectively applied to power data analysis.

In order to solve the problems of large amount of network computation and difficulty in distinguishing falling-like behaviors in fall detection algorithms， a fall detection algorithm based on joint point features was proposed. Firstly， based on the current advanced CenterNet algorithm， a Depthwise Separable Convolution-CenterNet （DSC-CenterNet） joint point detection algorithm was proposed to accurately detect human joint points and obtain joint point coordinates while reducing the amount of backbone network computation. Then， based on the joint point coordinates and prior knowledge of the human body， the spatial and temporal features expressing the fall behavior were extracted as the joint point features. Finally， the joint point feature vector was input into the fully connected layer and processed by Sigmoid classifier to output two categories： fall or non-fall， thereby achieving the fall detection of human targets. Experimental results on UR Fall Detection dataset show that the proposed algorithm has the average accuracy of fall detection under different state changes reached 98.00%， the accuracy of distinguishing falling-like behaviors reached 98.22% and the fall detection speed of 18.6 frame/s. Compared with the algorithm of the original CenterNet combining with joint point features， the algorithm of DSC-CenterNet combining with joint point features has the average detection accuracy increased by 22.37%. The improved speed can effectively meet the realtime requirement of the human fall detection tasks under surveillance video. This algorithm can effectively increase fall detection speed and accurately detect the fall state of human body， which further verifies the feasibility and efficiency of fall detection algorithm based on joint point features in the video fall behavior analysis.

Aiming at the problem that most algorithms of granular computing ignore the class-imbalance of data， a feature selection algorithm integrating pseudo-label strategy was proposed to deal with class-imbalanced data. Firstly， to investigate feature selection from class-imbalanced data conveniently， the sample consistency and dataset consistency were re-defined， and the corresponding greedy forward search algorithm for feature selection was designed. Then， the pseudo-label strategy was introduced to balance the class distribution of the data. By integrating the learned pseudo-label of a sample into consistency measure， the pseudo-label consistency was defined to estimate the features of the class-imbalanced dataset. Finally， an algorithm for Pseudo-Label Consistency based Feature Selection （PLCFS） for class-imbalanced data was developed based on the preservation of the pseudo-label consistency measure for the class-imbalanced dataset. Experimental results indicate that the proposed PLCFS has the performance only lower than max-Relevancy and Min-Redundancy （mRMR） algorithm， but outperforms Relief algorithm and algorithm for Consistency-based Feature Selection （CFS）.

With the rapid development of online education platforms represented by Massive Open Online Courses （MOOC）， how to evaluate the large-scale subjective question assignments submitted by platform learners is a big challenge. Peer grading is the mainstream scheme for the challenge， which has been widely concerned by both academia and industry in recent years. Therefore， peer grading technologies for online education were survyed and analyzed. Firstly， the general process of peer grading was summarized. Secondly， the main research results of important peer grading activities， such as grader allocation， comment analysis， abnormal peer grading information detection and processing， true grade estimation of subjective question assignments， were explained. Thirdly， the peer grading functions of representative online education platforms and published teaching systems were compared. Finally， the future development trends of peer grading was summed up and prospected， thereby providing reference for people who are engaged in or intend to engage in peer grading research.

Focusing on the malicious cheating behaviors of Third Party Auditor （TPA） in cloud audit， a trusted cloud auditing scheme without bilinear pairings was proposed to support the correct judgment of the behaviors of TPA. Firstly， the pseudo-random bit generator was used to generate random challenge information， which ensured the reliability of the challenge information generated by TPA. Secondly， the hash value was added in the process of evidence generation to protect the privacy of user data effectively. Thirdly， in the process of evidence verification， the interactive process between users and TPA results was added. The data integrity was checked and whether TPA had completed the audit request truthfully or not was judged according to the above results. Finally， the scheme was extended to realize batch audit of multiple data. Security analysis shows that the proposed scheme can resist substitution attack and forgery attack， and can protect data privacy. Compared with Merkle-Hash-Tree based Without Bilinear PAiring （MHT-WiBPA） audit scheme， the proposed scheme has close time for verifying evidence， and the time for generating labels reduced by about 49.96%. Efficiency analysis shows that the proposed scheme can achieve lower computational cost and communication cost on the premise of ensuring the credibility of audit results.

Aiming at the low embedding capacity of Reversible Data Hiding （RDH） in encrypted videos， a high-capacity RDH scheme in encrypted videos based on histogram shifting was proposed. Firstly， 4×4 luminance intra-prediction mode and the sign bits of Motion Vector Difference （MVD） were encrypted by stream cipher， and then a two-dimensional histogram of MVD was constructed， and （0，0） symmetric histogram shifting algorithm was designed. Finally， （0，0） symmetric histogram shifting algorithm was carried out in the encrypted MVD domain to realize separable RDH in encrypted videos. Experimental results show that the embedding capacity of the proposed scheme is increased by 263.3% on average compared with the comparison schemes， the average Peak Signal-to-Noise Ratio （PSNR） of encrypted video is less than 15.956 dB， and the average PSNR of decrypted video with secret can reach more than 30 dB. The proposed scheme effectively improves the embedding capacity and is suitable for more types of video sequences.

Fake news not only leads to misconceptions and damages people's right to know the truth， but also reduces the credibility of news websites. In view of the occurrence of fake news in news websites， a fake news detection method based on blockchain technology was proposed. Firstly， the smart contract was invoked to randomly assign reviewers for the news for determining the authenticity of the news. Then， the credibility of the review results was improved by adjusting the number of reviewers and ensuring the number of effective reviewers. At the same time， the incentive mechanism was designed with rewards distributed according to the reviewers' behaviors， and the reviewers' behaviors and rewards were analyzed by game theory. In order to gain the maximum benefit， the reviewers' behaviors should be honest. An auditing mechanism was designed to detect malicious reviewers to improve system security. Finally， a simple blockchain fake news detection system was implemented by using Ethereum smart contract and simulated for fake news detection， and the results show that the accuracy of news authenticity detection of the proposed method reaches 95%， indicating that the proposed method can effectively prevent the release of fake news.

The event that the user is interested in is extracted from the unstructured information， and then displayed to the user in a structured way， that is event extraction. Event extraction has a wide range of applications in information collection， information retrieval， document synthesis， and information questioning and answering. From the overall perspective， event extraction algorithms can be divided into four categories： pattern matching algorithms， trigger lexical methods， ontology-based algorithms， and cutting-edge joint model methods. In the research process， different evaluation methods and datasets can be used according to the related needs， and different event representation methods are also related to event extraction research. Distinguished by task type， meta-event extraction and subject event extraction are the two basic tasks of event extraction. Among them， meta-event extraction has three methods based on pattern matching， machine learning and neural network respectively， while there are two ways to extract subjective events： based on the event framework and based on ontology respectively. Event extraction research has achieved excellent results in single languages such as Chinese and English， but cross-language event extraction still faces many problems. Finally， the related works of event extraction were summarized and the future research directions were prospected in order to provide guidelines for subsequent research.

Focusing on the issue that the Low-Rank Representation （LRR） subspace clustering algorithm does not consider the local structure of the data and may cause the loss of local similar information during learning， a Low-Rank Representation subspace clustering algorithm based on Hessian regularization and Non-negative constraint （LRR-HN） was proposed to explore the global and local structure of the data. Firstly， the good speculative ability of Hessian regularization was used to maintain the local manifold structure of the data， so that the local topological structure of the data was more expressive. Secondly， considering that the obtained coefficient matrix often has positive and negative values， and the negative values often have no practical significance， non-negative constraints were introduced to ensure the effectiveness of the model solution and make it more meaningful in the description of the local structure of the data. Finally， the low-rank representation of the global structure of the data was sought by minimizing the nuclear norm， so as to cluster high-dimensional data better. In addition， an effective algorithm for solving LRR-HN was designed by using the linearized alternating direction method with adaptive penalty， and the proposed algorithm was evaluated by ACcuracy （AC） and Normalized Mutual Information （NMI） on some real datasets. In the experiments with clusters number 20 on ORL dataset， compared with LRR algorithm， LRR-HN has the AC and NMI increased by 11% and 9.74% respectively， and compared with Adaptive Low-Rank Representation （ALRR） algorithm， LRR-HN has the AC and NMI increased by 5% and 1.05% respectively. Experimental results show that the LRR-HN has great improvement in AC and NMI compared with some existing algorithms， and has the excellent clustering performance.

The traditional Graph Convolutional Network （GCN） and many of its variants achieve the best effect in the shallow layers， and do not make full use of higher-order neighbor information of nodes in the graph. The subsequent deep graph convolution models can solve the above problem， but inevitably generate the problem of over-smoothing， which makes the models impossible to effectively distinguish different types of nodes in the graph. To address this problem， an adaptive deep graph convolution model using initial residual and decoupling operations， named ID-AGCN （model using Initial residual and Decoupled Adaptive Graph Convolutional Network）， was proposed. Firstly， the node’s representation transformation as well as feature propagation was decoupled. Then， the initial residual was added to the node’s feature propagation process. Finally， the node representations obtained from different propagation layers were combined adaptively， appropriate local and global information was selected for each node to obtain node representations containing rich information， and a small number of labeled nodes were used for supervised training to generate the final node representations. Experimental result on three datasets Cora， CiteSeer and PubMed indicate that the classification accuracy of ID-AGCN is improved by about 3.4 percentage points， 2.3 percentage points and 1.9 percentage points respectively， compared with GCN. The proposed model has superiority in alleviating over-smoothing.