As AI large models increasingly find applications in protein science and bioinformatics,their complexity often obscures how neural networks extract and interpret key features from intricate biological data. To understand why such computational models excel in inferring the structure,function,and interactions of biomacromolecules,we extended previous research on predicting therapeutic antibody specificity. We proposed a Residual Convolution Channel Attention Neural Network (RCANN) . This network effectively predicted the probability of antibody⁃specific binding across diverse amino acid sequences,achieving a cross⁃validation AUC (Area under Curve) of 0.943,a significant improvement over traditional methods. Additionally,leveraging non⁃linear transformations and integrated gradients,we identified residue contribution patterns crucial for binding affinity,thereby inferring residue distributions in antibody sequences. This novel approach uncovers latent information behind amino acid sequences and notably reduces the unknown mutation space in predicting specific antibodies. Our study underscores that the new network structure not only enhances performance but also elucidates the underlying logic of complex neural networks.
The technology of Cloth⁃Changing Person Re⁃Identification (CC Re⁃ID) aims to recognize the same individual in video or image sequences over long periods. Existing methods mainly utilize multimodal information to model body shape to mitigate the impact of clothing,but they suffer from poor generalization ability and require extensive additional work. On the other hand,the methods relying solely on RGB images fail to effectively extract information unrelated to clothing. To address these issues,this paper proposes a CC Re⁃ID method based on multiple attention mechanisms and a spatial transformer network. The proposed method integrates the Convolutional Block Attention Module (CBAM) and Spatial Transformer Network (STN) into the backbone network to enhance the network's ability to perceive the importance of different channels and spatial locations,as well as its adaptability to images from various angles. To further improve the extraction of fine⁃grained pedestrian features,a triple attention mechanism is introduced to focus on information from different dimensions,along with an adaptive feature extraction module to learn the importance of different regions in the features. Additionally,the model employs multiple loss functions,such as clothing classification loss and clothing adversarial loss,to guide the model in learning clothing⁃independent information. Extensive experiments are conducted on four CC Re⁃ID datasets (LTCC,PRCC,VC⁃Clothes,and DeepChange),and the results demonstrate that the proposed method outperforms some state⁃of⁃the⁃art CC Re⁃ID methods in terms of Rank⁃1 and mAP metrics.
In UAV (Unmanned Aerial Vehicle) aerial images,there are numerous dense and small instances,often resulting in suboptimal detection performance. To address this issue,this paper replaces the original C3 (CSP bottleneck with 3 convolutions) module in the backbone of YOLOv5 with a C3TR module incorporating a Transformer structure to enhance the backbone’s feature extraction capability. Then,a CA (Coordinate Attention) module is added after the SPPF layer to improve the model's focus on small⁃object regions. Additionally,the ConvNeXtBlock module is employed in the neck network to replace the C3 module,where the deep convolutions in ConvNeXtBlock further enhance the recognition of small⁃object details,thereby improving the detection accuracy. Finally,the ECIoU loss function is introduced in place of the CIoU loss function to further improve the model's convergence speed and accuracy. Experimental results indicate that the modified model achieves increases of 9.5% and 6.3% in
Multimodal aspect level sentiment analysis models may overly rely on text modalities during feature extraction,while ignoring the potential semantic associations between text and image content. And due to the heterogeneous encoding properties and differences in information quality between modalities,effective cross modal interaction cannot be performed. To address this issue,this paper proposes a Multiscale Semantic Perception and Attention Fusion Model (MSPAF). Firstly,this model fully explores multiscale image semantic information and conducts cross modal semantic association modeling,to promote effective interaction between text images in a unified feature space; a dynamic gated cross attention mechanism is proposed for visual feature extraction under aspect guidance. Secondly,combining graph convolutional neural networks deeply reproduces semantic dependency relationships between words,to obtain syntactic and semantic enhanced contextual representations. Finally,at the multimodal feature fusion stage,multi⁃layer attention pooling is used to learn the correlation between different modal features and reduce the dimensionality of the fused features. On a publicly available sentiment analysis dataset,the proposed model is evaluated,and the experimental results show that compared to a series of baseline models,this model has better sentiment classification performance.
The human brain is essentially an extremely complex network,and studying brain networks can provide new perspectives for understanding brain mechanisms and the pathogenesis of brain diseases. Currently,deep learning methods based on Graph Convolution Network (GCN) have been widely applied in brain network analysis due to their powerful learning capabilities for graph⁃structured data,achieving significant progress in brain disease diagnosis. However,existing GCN methods still have the following two issues: (1) They only focus on local features in the Euclidean space of nodes,making it difficult to effectively characterize global features such as scale⁃free and small⁃world properties in brain networks; (2)They use manually constructed adjacency matrices,which cannot capture the spontaneous fluctuations in the brain,thus ignoring potential topological structure information. To address the above issues,we propose a brain network classification model based on adaptive hyperbolic graph convolution network (AH⁃GCN). Specifically,AH⁃GCN first uses a Bidirectional Gated Recurrent Unit (BGRU) to adaptively generate the adjacency matrix,then embeds node features into hyperbolic space and performs graph representation learning in the hyperbolic space. Finally,AH⁃GCN feeds the entire graph representation into a classifier for disease prediction. Experimental results on the Autism Brain Imaging Data Exchange (ABIDE) and Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset demonstrate that AH⁃GCN outperforms the existing state⁃of⁃the⁃art methods in brain network analysis.
Frequent Pattern Mining (FPM) is crucial in graph data analysis and mining,aiming to discover patterns in large⁃scale graph data with support above a specified threshold. Traditional FPM algorithms rely on the support threshold for search space pruning,often resulting in excessive redundantpatterns. While algorithms for the top⁃k patterns mining only return k frequent patterns,their results often fail to fully reflect users' interests and preferences as these patterns are picked based on objectiveindicators such as support. To address these issues,we propose a Pattern Interestingness Evaluation with Active Learning (PIEAL) method. PIEAL employs an active learning strategy to select representative patterns from frequent patterns that are mined from a sample graph and leverages limited human⁃computer interaction to gather users’ preferences on representative patterns,therefore enabling the prediction of pattern interestingness and the discovery of patterns of interest to users. During the interaction phase,PIEAL adopts a pairwise comparison strategy to collect users' feedback on the representative patterns,effectively reducing the difficulty of subjective evaluation. Experimental results on real life graphs show that with few interactions PIEAL can discover patterns of interest to users,achieving an accuracy of up to 95% on the test set.
Data sparsity and user selection bias negatively impact the recommendation performance of collaborative filtering algorithms. However,existing methods merely rely on user rating information for weighted matrix factorization to identify uninteresting items and use a uniform low⁃value filling for these items to mitigate these issues,overlooking the variations in user rating habits and item quality differences. To address this problem,this paper proposes a method based on user rating habits that combines the advantages of explicit and implicit feedback and employs multiple low⁃value filling strategies. The proposed method consists of two stages: identifying uninteresting items and filling uninteresting items. At the stage of identifying uninteresting items,user rating habits are mined using explicit feedback data,along with the inference of pre⁃usage preferences from implicit feedback data. At the stage of filling uninteresting items,the concept of item quality is introduced,and uninteresting items are divided into low⁃quality and high⁃quality categories based on explicit feedback data,and filled with different low values accordingly. Experimental results on two public datasets demonstrate that the proposed method outperforms the existing methods in both identifying and filling uninteresting items,significantly improving the performance of collaborative filtering algorithms in top⁃N recommendations.
As a reliable means of network security defense,intrusion detection technology is of great significance in ensuring network security. DBN (Deep Belief Network) combined with SVM (Support Vector Machine) is a machine learning method with good generalization ability and classification performance,which is widely used in the field of intrusion detection. However,this method is prone to the problem of "dimension disaster" when dealing with high⁃dimensional data,and the parameter selection has a great impact on the classification performance. In view of the above shortcomings,an intrusion detection method based on GWO (Gray Wolf Algorithm) to optimize DBN⁃SVM is proposed. In the GWO algorithm,the adaptive hunting weight coefficient and the improved head wolf position update formula are introduced to accelerate the convergence speed and expand the search range of the wolf group,and the optimal grey wolf individual adaptive disturbance strategy is added to avoid falling into local optimum. The improved GWO algorithm is further used to optimize DBN⁃SVM and it is applied to intrusion detection. The experimental results show that the accuracy of the proposed method on NSL⁃KDD and UNSW⁃NB15 datasets is 6.5% and 5.7% higher than that of the unimproved DBN⁃SVM,respectively,which can meet the application requirements of intrusion detection.
The problem of privacy disclosure arising from location services in social network has received increasing attention in recent years,and a measurement method of privacy disclosure for location k⁃anonymity is proposed. Firstly,the circulation process of user privacy in the location service is analyzed. Secondly,a correlation rule algorithm based on Analytic Hierarchy Process (AHP) is proposed to represent user background knowledge as a two⁃dimensional table of user attributes. The AHP method is used to determine the relative importance between attributes,and key background knowledge is filtered out based on the user's sensitivity to quantify the attacker's background knowledge. On this basis,a measurement method of privacy disclosure based on mutual information is proposed. The proposed model and background knowledge quantification method can provide a feasible measurement basis for knowledge quantification and privacy disclosure risk analysis and evaluation under attack background. Finally,the results of experimental verification and analysis indicate that the measurement model is feasible and effective.
Word segmentation is fundamental in natural language processing. This paper addresses the problem of how dictionary size influences word segmentation performance,proposes two novel measures:square overlap ratio (SOR) and relaxed square overlap ratio (RSOR),and validates their effectiveness through experiments. The SOR measure is the product of the dictionary overlap ratio and corpus overlap ratio,while the RSOR measure is the relaxed version of the SOR measure under unsupervised learning. The two measures both indicate the suitable degree between the segmentation dictionary and the object corpus to be segmented. The experimental results of Vietnamese word segmentation show that the RSOR⁃based unsupervised optimal selection can detect the most suitable,neither smaller nor larger,Vietnamese dictionary without manual labeling to achieve the best word segmentation performance for dictionary⁃based word segmenters.
As quantum computing has entered the noisy intermediate⁃scale quantum era,the effective control of qubits and the realization of quantum operations have become critical research directions. Meanwhile,machine learning plays an important role in many fields. Therefore,we aim to combine machine learning with quantum computing,utilizing machine learning methods to address quantum computing challenges and seek high⁃quality quantum control schemes. Based on superconducting quantum systems,we utilize the Continuous⁃time Variational Quantum Algorithm (CTVQA) to perform variational optimization on the parameters of a two⁃level superconducting qubit system. This approach optimizes the parameter control schemes for three quantum operations: the Controlled⁃NOT (CNOT) gate,the Controlled⁃Phase (CZ) gate,and the Quantum Fourier Transform (QFT). Then,we perform robustness analysis on the three parameter control schemes by introducing noise. The final results indicate that CTVQA demonstrates good universality in achieving quantum control implementations and can provide parameter schemes for experiments. Additionally,the parameter schemes obtained using CTVQA show favorable performance in terms of fidelity and robustness.
Phonon lasers have emerged as a focal point of research due to their immense potential in precision measurement,with the quest for low⁃threshold,high⁃conversion efficiency. This paper introduces a novel approach to achieve ultra⁃low threshold phonon lasers within a magnon⁃based cavity magnonic system. By altering the dissipation within the cavity of the magnonic system,this study manipulates the system's Parity⁃Time symmetry,examining the dynamics of the system under Parity⁃Time symmetry and its broken state. Further investigation into the steady⁃state dynamics of the mechanical modes within the cavity magnonic system under microwave drive reveals a strong correlation between the gain in the system's mechanical modes and its Parity⁃Time symmetry. Notably,near the system's exceptional point,it is possible to achieve phonon laser action with near⁃zero threshold power. Additionally,this paper demonstrates the relationship between magnon⁃phonon interactions and the threshold power of the phonon laser,highlighting a significant reduction in threshold power near the system's exceptional point,which is directly linked to a marked increase in effective magnon⁃phonon coupling.
Utilizing the active calcium in steel slags for carbon sequestration and the production of high⁃purity calcium carbonate products has emerged as a prominent research focus in the field of CCUS (Carbon Capture,Utilization,and Storage). Ammonium chloride (NH₄Cl) is commonly used in steel slags carbonation technology,facilitating the cyclic leaching and mineralization of active calcium,but it results in residual Cl-. To efficiently remove the residual Cl- from carbonated steel slags,this study conducts adsorption experiments varying initial Cl- concentrations,adsorption times,and reaction temperatures,as well as washing experiments on carbonated steel slags under different liquid⁃to⁃solid ratios and washing procedures.The experimental results indicate that the Cl- adsorption kinetics of steel slags adhere to a pseudo⁃second⁃order kinetic model,while the thermodynamics of adsorption align more closely with the Langmuir model. The mechanism of Cl- adsorption by steel slags can be attributed to chemisorption,in which activated aluminum components in the slags interact with Cl- in solution,as well as physisorption,facilitated by the amorphous calcium silicate hydrate (CSH) gel formed during the partial hydration of the slags. With a liquid⁃to⁃solid ratio of 2 L·kg-1,three washes are required for meeting the chlorine content standard in washed steel slags. However,with the ratios of 1 L·kg-1 and 0.5 L·kg-1,four washes are necessary. Countercurrent washing can be applied to steel slags,which saves 65%~75% of water compared to conventional washing methods.
With the improvement of science and technology,polymer composites are widely used.However,in the preparation of polymer composites,due to the surface wetting degree,roughness degree and surface energy of different polymer materials affect the service life of the materials. Therefore,in order to obtain better interfacial adhesion performance,the surface of the polymer often needs to be modified first.In this review,different treatment methods are introduced,mainly including plasma treatment,flame treatment,ultraviolet light treatment and atomic layer deposition method in energetic surface treatments of polymer modification,as well as surface grafting modification and surface chemical coating treatments in chemical treatment of polymer surface modification. Moreover,the principle of different methods and the progress of existing research are briefly described,and the future development direction of polymer surface modification to enhance interface bonding is prospected.