At present， in skeleton-based action recognition task， there still are some shortcomings， such as unreasonable data preprocessing， too many model parameters and low recognition accuracy. In order to solve the above problems， a skeleton-based action recognition method based on feature interaction and adaptive fusion， namely AFFGCN（Adaptively Feature Fusion Graph Convolutional Neural Network）， was proposed. Firstly， an adaptive pooling method for data preprocessing to solve the problems of uneven data frame distribution and poor data frame representation was proposed. Secondly， a multi-information feature interaction method was introduced to mine deeper features， so as to improve performance of the model. Finally， an Adaptive Feature Fusion （AFF） module was proposed to fuse graph convolutional features， thereby further improving the model performance. Experimental results show that the proposed method increases 1.2 percentage points compared with baseline method Lightweight Multi-Information Graph Convolutional Neural Network （LMI-GCN） on NTU-RGB+D 60 dataset in both Cross-Subject （CS） and Cross-View （CV） evaluation settings. At the same time， the CS and Cross-Setup （SS） evaluation settings of the proposed method on NTU-RGB+D 120 dataset are increased by 1.5 and 1.4 percentage points respectively compared with those of baseline method LMI-GCN. And the experimental results on single-stream and multi-stream networks show that compared with current mainstream skeleton-based action recognition methods such as Semantics-Guided Neural network （SGN）， the proposed method has less parameters and higher accuracy of the model， showing obvious advantages of the model， and that the model is more suitable for mobile device deployment.

Word embedding technology plays an important role in text sentiment analysis, but the traditional word embedding technologies such as Word2Vec and GloVe (Global Vectors for word representation) will lead to the problem of single semantics. Aiming at the above problem, a text sentiment analysis model named Sentiment Lexicon Parallel-Embedding from Language Model (SLP-ELMo) based on sentiment lexicon and context language model named Embedding from Language Model (ELMo) was proposed. Firstly, the sentiment lexicon was used to filter the words in the sentence. Secondly, the filtered words were input into the character-level Convolutional Neural Network (char-CNN) to generate the character vector of each word. Then, the character vectors were input into ELMo model for training. In addition, the attention mechanism was added to the last layer of ELMo vector to train the word vectors better. Finally, the word vectors and ELMo vector were combined in parallel and input into the classifier for text sentiment classification. Compared with the existing models, the proposed model achieves higher accuracy on IMDB and SST-2 datasets, which validates the effectiveness of the model.

Existing research works on the influence of social networks mainly focus on the propagation of single-source information, and rarely consider the reverse form of propagation. Aiming at the problem of reverse influence maximization, the heat diffusion model was extended to the multi-source heat diffusion model, and a Pre-Selected Greedy Approximation (PSGA) algorithm was designed. In order to verify the validity of the algorithm, seven representative seed mining methods were selected, and the experiments were carried out on different kinds of social network datasets with the propagation revenue of reverse influence maximization, the running time of the algorithm and the degree of seed enrichment degree as evaluation indexes. The results show that the seeds selected by PSGA algorithm have stronger propagation ability, low intensity, and high stability performance, and have advantage in the early stage of propagation. It can be thought that PSGA algorithm can solve the problem of reverse influence maximization.

Aiming at the tracking problem of multiple extended targets under the circumstances with unknown measurement noise covariance, an extension of standard Extended Target Probability Hypothesis Density (ET-PHD) filter and the way to realize its analysis were proposed by using ET-PHD filter and Variational Bayesian (VB) approximation theory. Firstly, on the basis of the target state equations and measurement equations of the standard ET-PHD filter, the augmented state variables of target state and measurement noise covariance as well as the joint transition function of the above variables were defined. Then, the prediction and update equations of the extended ET-PHD filter were established based on the standard ET-PHD filter. And finally, under the condition of linear Gaussian assumptions, the joint posterior intensity function was expressed as the Gaussian and Inverse-Gamma (IG) mixture distribution, and the analysis of the extended ET-PHD filter was realized. Simulation results demonstrate that the proposed algorithm can obtain reliable tracking results, and can effectively track multiple extended targets in the circumstances with unknown measurement noise covariance.

Aiming at the problems of inaccurate feature extraction and high complexity of traditional ElectroCardioGram (ECG) detection algorithms based on morphological features, an improved Long Short-Term Memory (LSTM) neural network was proposed. Based on the advantage of traditional LSTM model in time series data processing, the proposed model added reverse and depth calculations which avoids extraction of waveform features artificially and strengthens learning ability of the network. And supervised learning was performed in the model according to the given heart beat sequences and category labels, realizing the arrhythmia detection of unknown heart beats. The experimental results on the arrhythmia datasets in MIT-BIH database show that the overall accuracy of the proposed method reaches 98.34%. Compared with support vector machine, the accuracy and F1 value of the model are both improved.

Concerning the problem that traditional collaborative filtering algorithm just utilizes users' ratings on items when generating recommendation, without considering users' labels or comments, which can not reflect users' real preference on different items and the prediction accuracy is not high and easily overfits, a Stacked Denoising AutoEncoder (SDAE)-based improved Hybrid Recommendation (SDHR) algorithm was proposed. Firstly, SDAE was used to extract items' explicit features from users' free-text labels. Then, Latent Factor Model (LFM) algorithm was improved, the LFM's abstract item features were replaced with extracted explicit ones to train matrix decomposition model. Finally, the user-item preference matrix was used to generate recommendations. Experimental tests on the dataset MovieLens showed that the accuracy of the proposed algorithm was improved by 38.4%, 16.1% and 45.2% respectively compared to the three recommendation models (including the model based on label-based weights with collaborative filtering, the model based on SDAE and extreme learning machine, and the model based on recurrent neural networks). The experimental results show that the proposed algorithm can make full use of items' free-text label information to improve recommendation performance.

Concerning the problem of insufficient accuracy of the traditional image segmentation algorithm in segmentation of blue-green alga images, a new network structure named Deep Generative Adversarial Net (DGAN) based on Deep Neural Network (DNN) and Generative Adversarial Net (GAN) was proposed. Firstly, based on Fully Convolutional neural Network (FCN), a 12-layer FCN was constructed as the Generater ( G), which was used to study the distribution of data and generate the segmentation result of blue-green alga images ( Fake). Secondly, a 5-layer Convolutional Neural Network (CNN) was constructed as the Discriminator ( D), which was used to distinguish the segmentation result generated by the generated network ( Fake) and the true segmentation result with manual annotation ( Label), G tried to generate Fake and deceive D, D tried to find out Fake and punish G. Finally, through the adversarial training of two networks, a better segmentation result was obtained because Fake generated by G could cheat D. The training and test results on image sets with 3075 blue-green alga images show that, the proposed DGAN is far ahead of the iterative threshold segmentation algorithm in precision, recall and F ₁ score, which are increased by more than 4 percentage points than other DNN algorithms such as FCNNet (SHELHAMER E, LONG J, DARRELL T. Fully convolutional networks for semantic segmentation. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2017, 39(4):640-651) and Deeplab (CHEN L C, PAPANDREOU G, KOKKINOS I, et al. Semantic image segmentation with deep convolutional nets and fully connected CRFs. Computer Science, 2014(4):357-361). The proposed DGAN has obtained more accurate segmentation results. In the aspect of segmentation speed, the DGAN needs 0.63 s per image, which is slightly slower than the traditional FCNNet with 0.46 s, but much faster than Deeplab with 1.31 s. The balanced segmentation accuracy and speed of DGAN can provide a feasible technical scheme for image-based semantic segmentation of blue-green algae.

Emerging non-volatile, byte-addressable memories like phase-change memory can make data persistent at main memory level instead of storage. Since the read/write latency of Non-Volatile Memory (NVM) is very low, the overhead of software in a NVM system has become the main factor in determining the performance of the entire persistent memory system. Pmfs is a file system specifically designed for NVM. However, it still has an undesirable characteristic:each directory operation (create, open or delete) of Pmfs requires a linear search of the entire directory files, resulting in a cost linearly increased with the number of files in the directory. The performance of Pmfs under various workloads was evaluated and the test showed that the overhead of the directory operations had become the bottleneck of the whole system in some circumstance of particular workloads. To solve this problem, a persistent directory entry index was implemented in Pmfs to speed up directory operations. The experimental results show that under a single directory with 100 000 files, the file creation speed is increased by 12 times, the bandwidth is improved by 27.3%.

The single feature based vehicle license plate localization algorithms are hard to be adapted to the complex environment. In order to solve the problem, a multi-feature fusion algorithm was proposed, which made use of multi-features such as edge, color and texture. The localization process was divided into two phases: Hypothesis Generation (HG) and Hypothesis Verification (HV). In HG, feature point detection algorithm and mathematical morphology were used as the primary techniques, and the character texture and color information of vehicle license plate were extracted as the features to generate the candidates. In HV, gray projection technology and constant feature of vehicle license plate were used to verify the candidates from HG, then the correct license plate was located. The experimental results show that the proposed algorithm can achieve the localization success ratio of 96.6% and the precision of 95.4% in the testing image set in real environment. Moreover, the rationality and validity of the multi-feature fusion algorithm are verified.

In the smart grid, the development of electric power Demand Response (DR) brings great change to the traditional power utilization mode. Combined with real-time electricity price, consumers can adjust their power utilization mode by their energy demand. This makes load forecasting more complicated. The multi-input and two-output Least Squares Support Vector Machine (LS-SVM) was proposed to preliminarily predict the load and price at the same time. Considering the interaction between the real-time electricity price and load, the fuzzy recursive inference system based on data mining technology was adopted to simulate the game process of the forecasting of the price and load, and then the preliminary forecast results of multi-variable LS-SVM prediction algorithm were recursively corrected until the forecasting results were tending towards stability. Multi-variable LS-SVM can avoid running into local optima and has an excellent capacity of generalization, the improved association rules mining algorithm and loop predictive control algorithm have good completeness and robustness, and can correct the forecasting result approximately in every real situation. Simulation results of the actual power system show that the proposed method has better application effects.

In order to recommend trustworthy Web services, the differences between Web service recommendation and electronic commerce recommendation were analyzed, and then based on the collaborative filtering recommendation algorithm, a trustworthy Web service recommendation approach was proposed. At first, non-functional requirements of trustworthy software were evaluated. According to the evaluation results, similar users were filtered for the first time. Then, by using the rating information and basic information, the similar users were filtered for the second time. After finishing these two filtering procedures, the final recommendation users were determined. When using users' ratings information to calculate the similarity between the users, the similarity of the different services to the users was taken into consideration. When using users' basic information to calculate the similarity between the users, the Euclidean distance formula was introduced because of the nonlinear characteristics of the users. The problems of the dishonesty and insufficient number of users were also considered in the approach. At last, the experimental results show that the recommendation approach for trustworthy Web services is effective.

An automatic segmentation for pollen microscopic images was proposed in this paper, which was useful to develop a recognition system of airborne pollen. First, the image patch classifier was trained with normalized color component. Then, conditional random field was employed to model pollen images and Maximum A Posterior (MAP) was used to segment the pollen areas in microscopic images, with graph cut algorithm for optimization. In the experiments, the respective average values of mean distance error was 7.3 pixels and the true positive rate was 87% on 133 images. The experimental results show that image patch classifier and conditional random field model can be used to accomplish segmentation of the pollen microscopic images.

Aiming at several key issues in the fuzzy comprehensive evaluation, the structure of cooperating fuzzy comprehensive evaluation system(CFCES) was intorduced which used of CSCW technology. Analyzing the merits and demerits of both traditional concentrating and distributing control methods, gave formally a basic model of CFCES based on roles,A basic model of CFCES based on roles was given formally after analyzing the merits and demerits of both traditional concentrating and distributing control methods, and the implementing procedure of role-based cooperating components in CFCES was described.

More and more information, especially text information,has spread widely on Internet. To detect hot topics from plenty of Chinese text information,a term weight counting algorithm SDTF*PDF(Short Document Term Frequency * Proportional Document Frequency)was discussed. There were lots of channels in the system implementing this algorithm of detecting topics from short Chinese passages, and the passages in channels were usually short. Results worked out by it indicate that the system of detecting topic from short Chinese passages based on this algorithm can accurately extract the hot topics in a period of time, a day or a week, from enormous Chinese text information.