Attackers inject the designed adversarial sample into the target recommendation system to achieve illegal goals，seriously affecting the security and reliability of the recommendation system. It is difficult for attackers to obtain detailed knowledge of the target model in actual scenarios， so using gradient optimization to generate adversarial samples in the local surrogate model has become an effective black⁃box attack strategy. However，these methods suffer from gradients falling into local minima，limiting the transferability of the adversarial samples. This reduces the attack's effectiveness and often ignores the imperceptibility of the generated adversarial samples. To address these challenges，we propose a novel attack algorithm called PGMRS⁃KL that combines pre⁃gradient⁃guided momentum gradient optimization strategy and fake user generation constrained by Kullback⁃Leibler divergence. Specifically，the algorithm combines the accumulated gradient direction with the previous step's gradient direction to iteratively update the adversarial samples. It uses KL loss to minimize the distribution distance between fake and real user data，achieving high transferability and imperceptibility of the adversarial samples. Experimental results demonstrate the superiority of our approach over state⁃of⁃the⁃art gradient⁃based attack algorithms in terms of attack transferability and the generation of imperceptible fake user data.

Using adversarial examples for training enhances the robustness of deep neural networks. Therefore，improving the success rate of adversarial attacks is significant in the field of adversarial example research. Diluting original samples can bring them closer to the decision boundary of the model，thereby increasing the success rate of adversarial attacks. However，existing dilution algorithms suffer from issues such as reliance on manually generated dilution pools and single dilution targets. This paper proposes a method to enhance the capability of text adversarial attacks based on automatic dilution，called the Automatic Multi⁃positional Dilution Preprocessing (AMDP) algorithm. The AMDP algorithm eliminates the reliance on manual assistance in the dilution process and generates different dilution pools for different datasets and target models. Additionally，AMDP extends the targeted words for dilution，broadening the search space of dilution operations. As an input transformation method，AMDP can be combined with other adversarial attack algorithms to further enhance attack performance. Experimental results demonstrate that AMDP increases the success rate by approximately 10% on average on BERT，WordCNN，and WordLSTM classification models，while reducing the average modification rate of original samples and the average number of accesses to the target model.

A deep learning⁃based semi⁃supervised prediction method for the P53 mutation status of glioma is proposed to address the current problems of poor utilisation of glioma image data and incomplete feature extraction. Firstly，NUGMB (Non⁃Uniform Granularity Multi⁃Batch) grey level partitioning algorithm is proposed to optimize the preprocessing methods of glioma MR image. Secondly，the K⁃means clustering algorithm of MCC (Multi Center Collaboration) is proposed for pseudo⁃labeling of glioma image data. Finally，a novel attention mechanism，LWAM (Local Longer and Wider Attention Modules)，is proposed to construct an improved MaxViT model based on LWAM for the preoperative non⁃invasive prediction of the P53 mutation status of glioma. The NML⁃MaxViT model based on NUGMB，MCC and LWAM algorithms predicts the P53 mutation status of glioma with an accuracy of 96.23%，which achieves non⁃invasive predictions to assist physicians in clinical diagnosis and treatment.

Single⁃cell RNA sequencing (scRNA⁃seq) technology enables researchers to measure gene expression across the transcriptome at single⁃cell resolution，progressively transforming our understanding of cell biology and human diseases. However，the high variability，sparsity，and dimensionality of single⁃cell sequencing data have significantly impeded downstream analysis，making dimensionality reduction crucial for the visualization and the subsequent analysis of high⁃dimensional scRNA⁃seq data. Yet，existing single⁃cell dimensionality reduction algorithms have not adequately considered relationships intercellular，nor have jointly optimized the tasks of dimensionality reduction and clustering. To overcome these limitations，this study focuses on scRNA⁃seq data and employs machine learning techniques to investigate a dimensionality reduction algorithm based on autoencoders. In light of the fact that most existing dimensionality reduction algorithms do not consider the use of pseudo⁃labels to supervise the training process of the encoder，leading to the loss of intercellular signals during the dimensionality reduction of data，this paper proposes a cell dimensionality reduction algorithm based on the classified autoencoder. The algorithm combines the classified autoencoder with deep embedded clustering to generate a low⁃dimensional representation of the gene expression matrix. Experimental results demonstrate that compared to six other benchmark testing algorithms，this algorithm exhibits competitive performance in a range of downstream scRNA⁃seq analysis tasks.

Recently，the prediction of Multivariate Time Series (MTS) has gradually come into focus，especially with many Transformer⁃based models showing tremendous potential. However，existing Transformer⁃based models mainly focus on modeling cross⁃temporal dependencies，often overlooking the dependencies among different variables，which are crucial for MTS prediction. Therefore，this paper proposes a novel Multivariate Time Series prediction model，namely APDFinformer，designed to address the complex and dynamic nature of financial markets. The model integrates the Adaptive Multi⁃Scale Identifier (AMSI)，which extracts information from time series at different scales，helping reduce the impact of noise on time series and capture the interactions across different scales. Additionally，for the processed Multivariate Time Series data，the model utilizes the Decomposition method to divide it into trend and seasonal components. The trend component undergoes a simple linear processing，while the seasonal component，following the PatchTST approach，is sliced to shorten the sequence length，representing local features. This is advantageous for retaining local semantic information，facilitating the model's analysis of the correlations between time steps. Experimental results demonstrate that compared to traditional methods and various models similar to the Transformer model，APDFinformer more accurately captures the complex dynamics of financial markets and exhibits higher prediction accuracy. Specifically，compared to the Transformer model，APDFinformer reduces the MSE (Mean Squared Error) by 54%，24%，and 60% on the three selected cryptocurrency datasets，along with a reduction in MAE (Mean Absolute Error) by 39%，22%，and 44%. This study suggests that APDFinformer is a more reliable prediction tool for MTS in the financial domain and provides valuable insights for other application domains based on the Transformer model，meeting the evolving demands of financial markets.

With the development of Artificial Intelligence applications，the optimal automatic stock trading strategy to help investors achieve considerable returns in the volatile financial market has become a research hotspot at present. This paper proposes a stock trading decision⁃making algorithm LSTM⁃DDPG (Long Short⁃Term Memory Network⁃Deep Deterministic Policy Gradient). This algorithm combines the LSTM network that is better at capturing time series characteristics with the DDPG algorithm that is good at processing high⁃dimensional spatial data,and adds Dropout operation to reduce overfitting. In order to better grasp the dynamic changes of the market，six classic technical indicators in the stock market are introduced to expand the state space dimension of LSTM⁃DDPG. At the same time，two reward functions，cumulative return and Sharpe ratio，are used on LSTM⁃DDPG to provide investors with a variety of investment options. To verify its effectiveness,the proposed algorithm is applied to two kinds of trading tasks：single stock and stock portfolio. The datasets for the investment tasks include the data from both the US market and the Chinese market. The experimental results on multiple evaluation metrics such as cumulative return，Sharpe ratio，and Calmar ratio show that the proposed algorithm performs well in both domestic and foreign markets for the two kinds of investment tasks.

The multi⁃layer network community discovery algorithm aims to reveal the community structure of complex networks and has received widespread attention in recent years. However，existing algorithms only focus on the low⁃order structural information between nodes when measuring node similarity，ignoring the utilization of high⁃order structural information. Moreover，when fusing information from different layers of the network，there is a lack of consideration for the differences between different layers. To address these issues，the paper proposes a multi⁃layer network community discovery algorithm incorporating motif information. Specifically，firstly，each layer calculates a high⁃order adjacency matrix based on the motif information，fuses it with a adjacency matrix to obtain a reconstruction matrix，and then enhances the reconstruction matrix based on the importance of node neighbors to obtain a similarity matrix between nodes. Secondly，based on the reconstruction matrix，the importance of each layer of the network is calculated，and weighted fusion is used to obtain a unified similarity matrix. Finally，based on the obtained similarity matrix，the node influence is calculated，and the vector representation of the nodes is iteratively updated through the node embedding representation method to obtain the final embedding representation. Comparative experiments were conducted with existing multi⁃layer network community discovery algorithms on artificial multi⁃layer networks and real multi⁃layer network data. The results indicate that the proposed algorithm outperforms existing algorithms in terms of multi⁃layer modularity and normalized mutual information.

In the context of the rapidly evolving landscape of social media，where the number of users is substantial and the domains covered by social rumors shift swiftly，the challenge of developing an automated rumor detection system that quickly adapts to emerging domains while maintaining its detection capability remains significant. To address this challenge，this paper proposes an iterative rumor detection framework，LaReF，based on large language models and authoritative news sources. This framework leverages the strengths of large language models in natural language understanding and the credibility of authoritative news through an active learning approach to continuously optimize the rumor detection model. Specifically，LaReF comprises several key modules，an authoritative news retrieval module that enhances the model's detection capability using a dataset of authoritative news，a large language model feature extraction module and a cognitive pattern learning module，which utilizes prompt templates and attention mechanisms to extract features and learn the cognitive patterns of large language models,a feature validity prediction module，which automatically evaluates the importance of each feature and adjusts the weights accordingly，and a multi⁃feature fusion prediction module that integrates large language model features，sample semantic information，and authoritative news information for rumor detection. Experimental results demonstrate that LaReF exhibits strong performance in rumor detection tasks，effectively identifying the dissemination of rumors in emerging domains on social media. This provides a viable solution for constructing an information security ecosystem in cyberspace.

This article proposes a distributed node backup strategy to address the network disconnection issue caused by node failures in the lunar surface human⁃machine joint exploration network. Firstly，considering the unique characteristics of lunar missions and their networking methods，we model the lunar surface human⁃machine joint exploration network by categorizing the behaviors in the network and imposing constraints on the node behaviors. To reduce the information overhead maintained by each node，critical points are determined by two⁃hop neighbor information，and the critical points select backup nodes based on principles such as node type，recovery cost，and impact on the network using one⁃hop neighbor information and Hello message information. When recovering from failures，the independence of node tasks is considered，and a fault propagation strategy is used to offset the responsibility of moving astronaut nodes. Experiments demonstrate that this node backup mechanism can generate a backup recovery path for each critical node in the network with redundant nodes，and can reconnect the network without moving astronaut nodes.

When a high⁃orbit satellite (Geostationary Orbit，GEO) is used to provide real⁃time relay service for a low⁃orbit remote sensing satellite (Low Earth Orbit，LEO)，the allocation of GEO relay resources directly affects the data transmission efficiency，and at the same time，the return of transmitting the same size but different level of the task is different，which affects the overall network revenue. To solve the above problems，using the Stackelberg game model，this paper proposes a task⁃aware real⁃time relay transmission scheme for remote sensing data. First，to reasonably describe the topological changes between layers and reduce the number of link interruptions in each time slot，we determine the identity and number of LEOs participating in relaying under each time slot using the non⁃uniform time slot division method; second，according to the spatio⁃temporal attributes of remote sensing data and user needs，we utilize the naming mechanism in Named Data Networking (NDN) to name the tasks，which can realize the network's perception of the level and size of the content; finally，when multiple LEOs offload data to the GEO，a task⁃aware GEO resource allocation model is constructed using the Stackelberg game model，and a priority⁃based GEO resource allocation algorithm is proposed，where the optimal allocation of the GEO rate can be achieved according to the task priority under each time slot to maximize the benefit of the GEO，and the Nash equilibrium is proved to exist. Experimental results confirm the effectiveness of the proposed scheme in terms of pricing，GEO resource allocation，transmission efficiency，and network revenue.

Aiming to address the challenges of the current UAV aerial image de⁃fogging method，including difficulties in considering de⁃fogging for different depths of field，excessive loss of edge details，and ineffective results，this paper proposes a de⁃fogging method based on the Pyramid⁃Kuwahara filter. Firstly，atmospheric light estimation and transmittance are solved using an improved dark channel prior method. Secondly，a multi⁃scale filter called Pyramid⁃Kuwahara is designed to optimize and extract atmospheric light details. Then，a method named MFRTV is proposed to enhance detail information in transmission based on the designed filter. Finally，the restored fog⁃free image is obtained by utilizing the atmospheric scattering model along with optimized transmittance and atmospheric light maps generated by the algorithm. Experimental results demonstrate that our proposed fog removal algorithm effectively restores image details in different depths of field while significantly reducing fog presence in experimental images. Moreover，it successfully removes fog even at further depths of field achieving in enhanced subjective visual effects and increased information richness compared to other control algorithms. The proposed algorithm exhibits significant improvements in parameters such as entropy，FADE (fog area density estimation)，structural similarity index (SSIM)，and average gradient.

The competence⁃based knowledge space theory provides a scientific and effective theoretical framework for educational assessment and learning guidance. In this theory，the purpose of educational assessment is to understand students' competence levels (competence state) in specific domains based on their answers (knowledge state)，and provide personalized learning guidance based on the assessment results. However，there is no one⁃to⁃one correspondence between competence state and knowledge state. Solely relying on knowledge state only determines a category of competence state，without accurately judging the students' competence state. Therefore，this paper proposes a general method for the direct calculation of the master fringe of the knowledge state and personalized learning guidance based on the ceiling or floor of the same category of competence. The first step is to obtain the master fringe of the current knowledge state,the second step is to choose the target knowledge state based on the master fringe,the third step is to attain the ceiling or floor of the competence state corresponding to the current knowledge state and the target knowledge state,and the fourth step is to push the skill set to be learned soon based on the ceiling or floor，to achieve the target knowledge state. This paper also provides two characterization theorems,using the skill function to describe the ceiling or floor of the competence state，and using the problem function to describe the master fringe of the knowledge state. Obtaining the ceiling or floor of the competence state，and the master fringe of the knowledge state without establishing the knowledge structure improves acquisition efficiency. Finally，this paper presents algorithms to obtain the master fringe based on definitions and characterization theorems，and comparative practice demonstrates that the latter has shorter time consumption and smaller memory usage.

Tin dioxide (SnO₂) is an intrinsic semiconductor material with direct band gap of about 3.6 eV at room temperature. Because of the advantages of good chemical and thermal stability，high electron mobility and light transmittance，low cost，non⁃toxicity and environmental friendliness，SnO₂ has been widely used in gas sensors，catalytic materials，electrode materials，photoelectric materials and solar cells. Aimed at the contradiction between the application requirement of SnO₂ dispersible in both water and oil and commercial SnO₂ dispersible only in water，a solvothermal crystallization method was developed to successfully prepare the SnO₂ quantum dots (SnO₂ QDs) dispersible in both water and oil solvents. By adjusting the solvothermal crystallization time (0~24 h)，the crystal size，energy gap，and ratio of (110) plane of the obtained SnO₂ QDs were modified in the ranges of 2.12~4.16 nm，3.72~4.15 eV and 12.6%~43.3%，respectively. The crystalline SnO₂ QDs can be well dispersed in water，ethanol，ethyl acetate and other water or oily solvents as needed，to form a yellow transparent solution，which is long⁃term stable at room temperature.

As a macroscopic coherent state formed by the condensation of electron⁃hole bound pairs⁃ excitons，excitonic insulators still lack clear experimental evidence for their realization in actual materials since the commencement of theoretical research on them. Therefore,the theoretical exploration of more novel systems and the search for more candidate materials that can realize excitonic insulators are still hot topics in current research. In this paper，we systematically investigated the excitonic insulators generated from parent materials such as semiconductors，normal semimetals，and Dirac semimetals. Using impurities as probes，utilizing mean⁃field theory as well as the T ⁃matrix approximation，we found that nonmagnetic impurities have pair⁃breaking effect in excitonic insulators，inducing the generation of bound states. For excitonic insulators based on parent materials of normal semimetals and semiconductors，the intra⁃band and inter⁃band impurity scattering strength has a significant modulation effect on the numerical value and number of bound state energy levels. In contrast，the bound state energy levels in excitonic insulations with Dirac semimetals as the parent materials are hardly affected by impurity scattering. In addition，the impurity induced bound states show very different energy characteristics in the three different excitonic insulators. Lower energy inter⁃gap bound states can be produced in the excitonic insulators generated from normal semimetals than semiconductors，while the excitonic insulators generated from the Dirac semimetals have bound states tightly situated near the band⁃gap edge.

Human comfort is an important indicator of urban livability,and it is the response of human organs and skin and other receptors to environment changes through the nervous system. Human comfort is directly affected by meteorological factors such as temperature,humidity and wind speed. In this paper,the changes of human comfort indicators,including HI (Heat Index),Humidex,WBGT,WCI,in Nanjing area were analyzed by using meteorological observations from 2015 to 2021 at the SORPES station located on Xianlin Campus of Nanjing University. The interannual variation magnitude of each index is basically similar,ranging from 15 to 20. The annual and monthly trends of each index are highly consistent with the temperature trend,because in the definition and calculation of each index,temperature is the meteorological factor with the greatest weight,and on this basis,it is revised by humidity,wind speed and other factors. From the perspective of the probability density of each index interval,the probability density distribution of each index has a weak "three⁃peak type",which is caused by the superposition of the "bimodal type" in autumn and the "unimodal type" in other seasons. The "bimodal" distribution of each index in autumn is mainly caused by the large temperature difference between day and night. From the perspective of statistical characteristics,the human body perception in Nanjing is mainly comfortable,and the probability of hot environment risk is much greater than that of cold environment risk. The risk probability is slightly higher with Humidex than that is with HI. Using the training records of 200 m sprint of athletes from Nanjing University,the impact of human comfort on outdoor sports was analyzed,and it was found that there are great individual differences in the sensitivity response of different athletes to the thermal environment. That is 30% of the subjects are affected by the thermal environment comfort,and 10% of the subjects are significantly affected by the thermal environment comfort,and the sensitivity of female athletes is higher than that of male athletes. The results show that the impact of environmental human comfort on athletes in future sports training cannot be ignored,entailing more attention and research.