Aiming at the problem of rapidly growing and frequently changing requirement for risk map generation of immovable cultural heritage， and existing programs and tools cannot meet the needs of actual applications， a method for constructing semantic model was proposed. Based on the semantic model， a Domain-Specific Language （DSL） close to natural language was designed for experts in the field of immovable cultural heritage. Firstly， a business model was extracted by conducting in-depth research on various indicators of immovable cultural heritage， as well as methods and processes for generating risk maps. Secondly， the meta-calculation units of the risk value calculation rules were abstracted， and a semantic model was constructed by analyzing the business model. On this basis， a DSL that can express all semantics in the semantic model was designed. The language script can be programmed by the field experts themselves and used to quickly and efficiently generate risk maps. It is easy to expand and can meet the needs of frequently changing requirements. Compared with the mainstream method of generating risk maps by using Geographic Information System （GIS）， the use of DSL to generate risk maps can reduce work hours by more than 66.7%.

In the cross-domain sentiment analysis， the labeled samples in the target domain are seriously insufficient， the distributions of features in different domains are very different， and the emotional polarities expressed by features in one domain differ a lot from the emotional polarities in another domain， all of these problems lead to low classification accuracy. To deal with the above problems， an aspect-level cross-domain sentiment analysis method based on capsule network was proposed. Firstly， the feature representations of text were obtained by BERT （Bidirectional Encoder Representation from Transformers） pre-training model. Secondly， for the fine-grained aspect-level sentiment features， Recurrent Neural Network （RNN） was used to fuse the context features and aspect features. Thirdly， capsule network and dynamic routing were used to distinguish overlapping features， and the sentiment classification model was constructed on the basis of capsule network. Finally， a small amount of data in the target domain was used to fine-tune the model to realize cross-domain transfer learning. The optimal F1 score of the proposed method is 95.7% on Chinese dataset and 91.8% on English dataset， which effectively solves the low accuracy problem of insufficient training samples.

Considering the issues of unimodal feature representation and cross-modal feature fusion in sequential multimodal sentiment analysis, a multi-task learning based sentiment analysis model was proposed by combining with multi-head attention mechanism. Firstly, Convolution Neural Network (CNN), Bidirectional Gated Recurrent Unit (BiGRU) and Multi-Head Self-Attention (MHSA) were used to realize the sequential unimodal feature representation. Secondly, the bidirectional cross-modal information was fused by multi-head attention. Finally, based on multi-task learning, the sentiment polarity classification and sentiment intensity regression were added as auxiliary tasks to improve the comprehensive performance of the main task of sentiment score regression. Experimental results demonstrate that the proposed model improves the accuracy of binary classification by 7.8 percentage points and 3.1 percentage points respectively on CMU Multimodal Opinion Sentiment and Emotion Intensity (CMU-MOSEI) and CMU Multimodal Opinion level Sentiment Intensity (CMU-MOSI) datasets compared with multimodal factorization model. Therefore, the proposed model is applicable for the sentiment analysis problems under multimodal scenarios, and can provide the decision supports for product recommendation, stock market forecasting, public opinion monitoring and other relevant applications.

Concerning the problem of low accuracy in sentiment classification caused by short text, an end-to-end short text sentiment classifier was proposed based on adversarial learning and variational inference. First, the spectrum normalization technology was employed to alleviate the vibration of discriminator in training process. Second, an additional classifier was utilized to guide the updating of the inference model. Third, the Adversarial Variational Bayes (AVB) was used to extract the topic features of the short text. Finally, topic features and pre-trained word vector features were fused by three times of attention mechanism in order to realize the classification. Experimental results on one product review and two micro-blog datasets show that the proposed model improves the accuracy by 2.9, 2.2 and 8.4 percentage points respectively compared to the Bidirectional Long Short-Term Memory network based on Self-Attention (BiLSTM-SA). It can be seen that the proposed model can be applied to mine sentiments and opinions in social short texts, which is significant for public opinion discovery, user feedback, quality supervision and other related fields.

Existing generation models have difficulty in directly generating high-resolution images from complex semantic labels. Thus, a Generative Adversarial Network based on Semantic Labels and Noise Prior (SLNP-GAN) was proposed. Firstly, the semantic labels (including information of shape, position and category) were directly used as input, the global generator was used to encode them, the coarse-grained global attributes were learned by combining the noise prior, and the low-resolution images were generated. Then, with the attention mechanism, the local refined generator was used to query the high-resolution sub-labels corresponding to the sub-regions of the low-resolution images, and the fine-grained information was obtained, the complex images with clear textures were thus generated. Finally, the improved Adam with Momentum (AMM) algorithm was introduced to optimize the adversarial training. The experimental results show that, compared with the existing method text2img, the proposed method has the Pixel Accuracy (PA) increased by 23.73% and 11.09% respectively on COCO_Stuff and the ADE20K datasets; in comparison with the Adam algorithm, the AMM algorithm doubles the convergence speed with much smaller loss amplitude. It proves that SLNP-GAN can efficiently obtain global features as well as local textures and generate fine-grained high-quality images.

In big data environment, the mass of nodes in graph and the complexity of analysis bring forward higher requirement for the speed and accuracy of maximum clique problems. In order to solve the problems, a Parallel Multi-layer Graph Partitioning method for Solving Maximum Clique (PMGP_SMC) was proposed. Firstly, a new Multi-layer Graph Partitioning method (MGP) was proposed, the large-scale graph partitioning was executed to generate subgraphs while the clique structure of the original graph was maintained and not destroyed. Large-scale subgraphs were divided into multi-level graphs to further reduce the size of subgraphs. The graph computing framework of GraphX was used to achieve MGP to form a Parallel Multi-layer Graph Partitioning (PMGP) method. Then, according to the size of partitioned subgraph, the iteration number of Local Search algorithm Based on Penalty value (PBLS) was reduced, and the PBLS based on Speed optimization (SPBLS) was proposed to solve the maximum clique of each subgraph. Finally, PMGP method and SPBLS were combined to form PMGP_SMC. The large-scale dataset of Stanford was used for running test. The experimental results show that, the proposed PMGP reduces the maximum subgraph size by more than 100 times and the average subgraph size by 2 times compared with Parallel Single Graph Partitioning method (PSGP). Compared with PSGP for Solving Maximum Clique (PSGP_SMC), the proposed PMGP_SMC reduces the overall time by about 100 times, and its accuracy is consistent with that of Parallel Multi-layer Graph Partitioning for solving maximum clique based on Maximal Clique Enumeration (PMGP_MCE) in solving the maximum clique. The proposed PMGP_SMC can solve the maximum clique of large-scale graph quickly and accurately.

In order to further improve the execution efficiency of Frequent Pattern-Growth (FP-Growth) algorithm on Spark platform, a new parallel FP-Growth algorithm based on Spark, named BFPG (Better Frequent Pattern-Growth), was presented. Firstly, the grouping strategy F-List was improved in the size of the Frequent Pattern-Tree (FP-Tree) and the amount of partition calculation to ensure that the load sum of each partition was approximately equal. Secondly, the data set partitioning strategy was optimized by creating a list P-List, and then the time complexity was reduced by reducing the traversal times. The experimental results show that the BFPG algorithm improves the mining efficiency of the parallel FP-Growth algorithm, and the algorithm has good scalability.

Considering the problem that traditional active learning algorithms can only handle small and medium size data sets, a big data active learning algorithm based on MapReduce was proposed. Firstly, a classifier was trained by Extreme Learning Machine (ELM) algorithm on an initial training set, and the outputs of the classifier were transformed into a posterior probability distribution by softmax function. Secondly, the big data set without labels was partitioned into l subsets, which were deployed to a cloud computing platform with l nodes. On each node, the information entropies of instances of each subset were calculated by the trained classifier, and q instances with maximum information entropies were selected for labeling, then the l× q labeled instances were added into the training set. Repeat the above steps until the predefined termination criterion was satisfied. Contrast test with ELM-based active learning algorithm were conducted on 4 data sets including Artificial, Skin, Statlog and Poker. Experimental results show that the proposed algorithm can complete active instance selection on 4 data sets, while the active learning algorithm based on ELM can only complete active instance selection on the smallest data set, indicating that the proposed algorithm outperforms the active learning algorithm based on ELM.

For SLP (Superword Level Parallelism) algorithm cannot effectively process the large-scale applications covered with few parallel codes, and the codes which can be vectorized may be adverse to vectorization. A new improved algorithm for SLP was proposed, namely NSLPO. First of all, the non-isomorphic statements which cannot be vectorized were transformed to isomorphic statements as far as possible, thus locating the opportunities of vectorization which SLP has lost. Secondly, the Max Common Subgraph (MCS) was built by adding redundant nodes, and the supplement diagram of SLP was got by using some optimization such as redundancy deleting, which can greatly increase the parallelism of program. At last, the codes which are harmful to vectorization were exclued out of vectorization by using cutting method and executed in serial, only the valuable codes for vectorization were vectorized to improve the efficiency of programs as far as possible. Experiments were conducted on widely used kernel test sets. The experimental results show that compared with the SLP algorithm, the proposed NSLPO algorithm has better performance and its running time was reduced by 9.1%.

According to the implicit characteristics of IPTV (Internet Protocol Television) user viewing behavior data, a novel implicit rating model was proposed. Firstly, the main features of IPTV user viewing behavior data were introduced, and a new mixed feature implicit scoring model was proposed, which combined with viewing ratio, user interest bias factor and video type influence factor. Secondly, the strategy of viewing behavior based on viewing time and viewing ratio was proposed. Finally, a distributed model architecture based on Hadoop was designed and implemented. The experimental results show that the proposed novel model effectively improves the quality of the recommended results in the IPTV system, improves the time efficiency, and has good scalability for large amounts of data.

To generate an effective visual dictionary for representing the scene of images, and further improve the accuracy of semantic annotation, a scene annotation model based on Formal Concept Analysis (FCA) was presented by means of an abstract from the training image set with the initial visual dictionary as a form context. The weight value of visual words was first marked with information entropy, and FCA structures were built for various types of scene. Then the arithmetic mean of each visual word's weight values was used to describe the contribution among different visual words in the intent to the semantic, and each type of visual vocabularies for the scene was extracted from the structure according to the visual vocabularies thresholds. Finally, the test image was assigned with the class label by using of the K-nearest method. The proposed approach is evaluated on the Fei-Fei Scene 13 natural scene data sets, and the experimental results show that in comparison with the methods of Fei-Fei and Bai, the proposed algorithm has better classification accuracy with β=0.05 and γ=15.

A new threshold function under non-Gaussian noise background was presented to overcome the limitations of wavelet threshold algorithm under the Gaussian noise background. The shortcomings of conventional function, such as discontinuity of hard threshold function and the invariable dispersion of soft threshold function, can be solved. The new function which employed high order power function was put forward based on Garrote threshold. First, the signal with a class of non-Gaussian noise was decomposed by wavelet. Secondly, each high frequency wavelet coefficient was quantified based on new threshold function. Thirdly, signal was reconstructed by the low frequency coefficients of wavelet decomposition and quantified high frequency coefficients. The simulation results under non-Gaussian noise background indicate that the new threshold function gets higher Signal-to-Noise Ratio (SNR) gains and lower minimum Mean Square Error (MSE) compared to the soft and hard threshold, two types of improved threshold and Garrote threshold.

A fast calculation approach for large-scale matrix operation, which can be accomplished by Graphic Processing Unit (GPU), was designed. For taking full advantage of the parallel architecture of GPU to enhance the calculation speed, special matrix partitioning and memory allocation mechanism according to the features of GPU were designed to decrease the frequency of data access. Meanwhile Kahan's summation formula was introduced to ensure the precision of the calculation. The result shows that the approach can achieve better effect and greatly enhance the speed and the precision of the large matrix multiplication.

The theoretical calculation of the participant payload was given, the relationship between the participant payload and the rerouting road's dependency was investigated，and the factors that decided the system payload performance were analyzed. The result shows that the determinants are the mathematical expectations of the road's length and the requested dependency. The system administrator can balance the payload and the dependencies to fit the different usages by configuring these two parameters.

An improved ant colony algorithm was provided in this paper for robot path planning in a static environment. In this algorithm the model of robot's workspace was established with grid method and foldback iterating was used to search the aims by simulating the foraging behavior of ant colony. A heuristic factor based on the most pheromone in a moving direction range and a goal guiding function were used during the searching process. Furthermore, according to the features of the pheromone strewing when solving the problem by ant colony algorithm, the strewing method and updating strategy of pheromone were reconstructed. The simulation results show that these improvements make searching of the best path rapid and efficient. With this method a best path can be found rapidly even if the obstacles are exceedingly complicated.