In order to optimize the performance of generative adversarial networks on automatic advertisement image generation, this paper combines the variational self-encoder with generative adversarial networks, which consists of four parts: encoder network, decoder network, target-to-be-attacked network, and discriminator network to form a new adversarial sample generation method based on GANs, i.e., AdvAE-GAN model. To make the generated samples more clear and natural, the adversarial learning mechanism and similarity metric (PCE) are added to the AdvAE-GAN model. To obtain the performance of the model in diverse image coloring, multiple methods are elicited for subjective and objective qualitative evaluation and model complexity analysis, respectively. Combining the four standard datasets of AWA, CUB, SUN and FLO, zero-sample image recognition, generalized zero-sample learning experiments are carried out sequentially to derive the loss value curve of the model. The visual effects of animated advertisements generated by AdvAE-GAN model are rated using questionnaire research. For the product effect of animated advertisements generated by AdvAE-GAN model, the category diversity, design diversity, animation contour completeness, and image clarity indexes with scores above 7 account for 70.47%, 85.82%, 76.73%, and 84.02%, respectively. The animated advertisement generation model based on improved generative adversarial network is recognized by the market as well as the society and can be deepened.

Financial fraud, as a global problem in the financial industry, brings huge economic losses to financial institutions and customers. In this paper, a multi-task financial fraud detection model is constructed based on heterogeneous graph neural network with deep reinforcement learning, combined with variational self-encoder. In this model, the variational self-encoder is combined with graph convolutional network to construct the node input representation coding module, as a way to enhance the multi-task financial fraud data and better mine the structured features of different nodes. The attention mechanism is then introduced to build the relation-aware attention, which deeply mines the input node features, further acquires the neighbor-generated features of different nodes in the network, and combines the mutual information to measure the nonlinear correlation between different random nodes. Then the financial fraud node representation is mapped into the highdimensional space by the multilayer perceptron, and then the financial fraud prediction confidence of the model is obtained, and different types of loss functions are set to ensure the detection efficiency of the model. The results show that the F1-macro and AUC values of the financial fraud detection model on the self-constructed FFD dataset are 0.749 and 0.925, respectively. Relying on the heterogeneous graphical neural network and the variational autocoder, a multi-task financial fraud detection model can be constructed, which provides a new idea for solving the suspected fraud and money laundering cases that may exist in the field of finance and economy.

Under the current development trend of global economic integration, countries around the world are interconnected and influenced by each other in international trade, and the connection of world trade forms a complex network. This paper constructs a global trade network based on global trade theory and social network analysis theory, and selects indicators such as the number of network nodes and network diameter to characterize the topological structure of the global trade network. The Transformer model is designed based on the gating mechanism unit and dynamic attention mechanism to analyze the multimodal, high-dimensional and heterogeneous global trade time series data. The empirical analysis finds that the characteristics of the global trade network structure change over time, the trade network between countries and regions becomes more and more close, and there is an impulse effect of the country’s GDP and other influencing variables on the structure of the global trade network. This paper reveals the multi-path influence effect of global trade network through empirical analysis, and improves the related research on the structural change and positive evolution of global trade network, with a view to providing useful reference and guidance for the formulation of national trade countermeasures.

The article proposes a novel cross-modal adversarial learning framework for analyzing the emotional dynamics of non-English learners during classroom engagement and predicting their individualized behaviors. The framework combines multilevel feature extraction and Transformer CNN-LSTM integrated model to handle multimodal data more efficiently and capture the complex relationship between emotions and behaviors. Low-level and high-level multilevel features are then extracted from the raw multimodal data. Meanwhile, Transformer is utilized to mine long-distance dependencies between multimodal data, CNN extracts local features, and LSTM is used to model dynamic changes in time series. In addition, the framework introduces adversarial training to learn shared features across modalities. Before 50 rounds of training, the CL-Transformer model loss function, emotion recognition accuracy, and behavior prediction accuracy converge, showing the fastest training speed and training results. The algorithm in this paper has more than 90% precision, recall, and F1 scores for emotion recognition and behavior prediction, and the recognition accuracy for different emotions is up to 0.96. In the fifth stage of the case study, the classroom emotion conversion rate and arousal is up to 0.66, and the model predicts that the probability of cell phone playing behavior is the highest for learners who are in angry moods, which is 64.7%. The learners’ classroom emotional acceptance as well as behavioral integration have an impact on their classroom engagement.

The study combines hierarchical Bayesian model and adversarial neural network according to the model architecture of neural machine translation, and introduces the domain generalization method based on cross-domain gating to solve the domain generalization problem, and constructs the neural machine translation system based on hierarchical Bayesian model. Translation performance experiments are conducted on this translation system to test the cross-domain generalization performance of the neural machine translation system based on hierarchical Bayesian model in this paper. The translation method of this paper significantly outperforms the baseline system of statistical machine translation in the direction of translation for all the inter translated languages and medial languages of the European Parliament corpus. The statistical machine translation model and the standard neural machine translation model have maintained a stable performance during the growth of the interpolation coefficients, while the performance of this paper’s hierarchical Bayesian-based neural machine translation system grows rapidly to the maximum when the interpolation coefficients grow to 0.3 or 0.4, and its overall average BLEU value always outperforms that of the statistical machine translation model and the standard neural machine translation model. The BLEU values of the hierarchical Bayesian-based neural machine translation system are 35.26% and 34.28% for bidirectional Chinese-English translation, and 26.42% and 25.96% for bi-directional Chinese-Western translation, which are better than those of the neural machine translation based on the attentional mechanism and variational scoring. And the hierarchical Bayesian-based neural machine translation system has strong stability on the translation of low-resource languages.

As a conventional technique in lacquer painting, the abrasion painting technique is widely used in the creation of modern lacquer painting. In order to promote the digital innovation of the abrasion painting technique in the creation of lacquer paintings, a fusion scheme of the abrasion painting technique and color distribution in the creation of lacquer paintings is formulated. According to the relationship between color and gray scale, the color mapping of image coloring algorithm is proposed under the framework of energy optimization algorithm to realize algorithm-driven lacquer painting color generation. In addition, with the technical support of the renderer, the color distribution of lacquer paintings is integrated with the milling technique according to the principle of texture mapping. With the help of evaluation indexes and experimental platforms, we simulate and analyze the techniques and colors in lacquer painting. In the color generation of lacquer paintings, the indicators of this paper’s method are 34.09, 0.964, 0.025 and 4.28 in order, which verifies the application effect of this paper’s method in the color generation of lacquer paintings. In addition, the speed of this paper’s rendering method (42-86FPS), fully meets the requirements of real-time drawing, this method better promotes the fusion of grinding and painting techniques and color distribution in the creation of lacquer paintings, which is of great significance to the digital dissemination of traditional culture of non-heritage.

As an emerging form of cultural communication, microshort dramas have emerged in the audiovisual industry. In order to explore the optimization method of international communication of short microdramas, this paper takes the selected short micro-dramas of an international video platform as samples, selects the influencing factors of the international communication effect of short microdramas, constructs the optimization model of international communication of short micro-dramas by using Bayesian network, and adopts the Great Likelihood Estimation Algorithm as its parameter learning method. The performance of the Bayesian network model is explored through model comparison, node sensitivity analysis and scenario simulation. The results show that the Bayesian network model has good prediction performance, and its AUC value is greater than 0.8 in both training and testing results. The entropy reduction percentages of publisher’s fan number, video duration and localized creation are all greater than 0.07%, which have the most obvious influence on the effect of international dissemination of microshort dramas. Scenario simulation verifies the influence of each variable on the optimization of the international dissemination effect of micro-short dramas, and the probability value of the obtained optimal solution with a strong dissemination effect is 83.5%. It is recommended to actively guide the creation of high-quality products, carry out in-depth localized creation, accelerate the integration of art and technology, and strengthen the comprehensive governance of the industry, so as to promote the global dissemination of China’s online micro short dramas.

In order to solve the problem of vagueness and uncertainty, which is difficult to deal with in traditional education assessment, this paper introduces the theory of fuzzy matrix logic, and constructs a multilevel assessment model of education quality by means of the affiliation function and multilevel weight allocation. Through fuzzy reasoning and cognitive estimation techniques, combined with knowledge graph visualization, the cognitive level of learners is accurately estimated to achieve personalized learning resource recommendation. The quality assessment of physical education teaching in colleges and universities is taken as an example to verify the application value of the model. The constructed PE teaching quality evaluation index system contains 3 level 1 indicators, 11 level 2 indicators, and 38 level 3 indicators.The initial index scoring result of the PE classroom by 5 raters is an average score of 100.8, which is 89.2 points different from the full average score.The weights of the indicators within the 3 levels do not differ much. Students’ levels of knowledge of the 6 initial physical education concepts ranged from 0.53 to 0.86 points. The maximum inter-conceptual influence strength was 0.86 and the minimum was 0.18. After the interference of the resource recommendation, the cognitive level increased to between 0.67-0.98 points. The maximum inter-conceptual influence intensity reaches 1. The Sig value is greater than 0.05, and the results of the model calculations have reliability and can be used for education quality assessment and dynamic learning planning method improvement.

Due to the complexity of the ship product structure and process, long production cycle and other factors, ship enterprises are plagued by the problem of profitability. Strengthening cost prediction and budget control is a very important means for ship enterprises to improve their profit margins. By analyzing the cost structure of shipbuilding, this paper proposes a rolling forecast model of shipbuilding cost based on long and short-term memory neural network (LSTM) as the estimation method of shipbuilding cost. Meanwhile, the traditional earned value method and target cost method are combined to sort out the shipbuilding cost control process and prepare the cost control plan as the control strategy of shipbuilding cost. Then we take the manufacturing data of a shipyard as the experimental object, use this paper’s model for data mining, compare the data performance of this paper’s model with similar algorithms, and verify the feasibility of this paper’s model. Finally, the model of this paper is applied to real cases. In the comparison of the estimation results between this paper’s model and the commonly used algorithms, the average error of cost estimation of this paper’s model is ±4.95%, which is better than the average error of the commonly used algorithms. The superior accuracy of this paper’s model in shipbuilding cost estimation is verified.

In this paper, K-prototype algorithm is chosen to cluster and analyze the data of students’ behavior in the educational field. Further, a model of students’ employment interest is constructed based on the job rating data of different classes of students. The timeliness is introduced in the model to improve the recommendation accuracy. Synthesize the algorithm and model to build an employment support system. Apply the system to the clustering study of college students’ behavioral data to verify its career recommendation value. Set up comparison experiments to find the optimal similarity fitting parameters and number of neighbors to improve the system recommendation accuracy and judge the system recommendation effect. Preliminarily divide students into 3 categories by analyzing students’ online behavior and book borrowing behavior. Preliminarily categorize students into 4 categories based on their grades. Combined with the performance labels and grade categories of professional courses, the employment direction of students was finally clustered into four categories, namely “postgraduate entrance examination”, “civil servant application”, “company work” and “others”. The highest accuracy of the system job recommendation is achieved when the similarity fitting parameter λ = 0.5 and the number of neighbors N = 50.The RMSE value of the K-prototype algorithm ranges from 0.6011 to 0.731, and the recommendation effect is better than the comparison algorithm.