Classroom behavior analysis system based on adaptive algorithm

Abstract

The application relates to the technical field of classroom behavior analysis, and discloses a classroom behavior analysis system based on an adaptive algorithm, which comprises a detection module, a parameter extraction module, a feature integration module and a decision generation module. The detection module collects classroom multi-modal behavior data and identifies active student terminals; the parameter extraction module obtains real-time posture vectors of target students, speech spectrum features, interactive response time delays and teaching content semantic labels; the feature integration module cross-modally fuses data to generate posture, speech, interactive fusion features and semantic correlation features; and the decision generation module generates behavior matching degrees and determines target analysis objects through hierarchical correlation modeling and dynamic priority sorting based on a multi-head attention mechanism. The system realizes multi-modal data fusion by means of algorithms such as space-time decomposition and dynamic density clustering, improves the accuracy and real-time performance of classroom behavior analysis, and provides support for teaching optimization.

Description

Technical Field

[0001] This invention relates to the field of classroom behavior analysis technology, specifically a classroom behavior analysis system based on adaptive algorithms. Background Technology

[0002] In traditional classroom teaching scenarios, teachers struggle to grasp the learning status and behavioral performance of all students in a real-time and comprehensive manner, resulting in insufficient targeting and effectiveness of teaching interactions. With the development of educational informatization, using intelligent technology to analyze classroom behavior has become an important direction for improving teaching quality. However, existing technologies have significant shortcomings in terms of data processing accuracy, the effectiveness of multimodal information fusion, and dynamic adaptation to changes in teaching scenarios.

[0003] From a data acquisition perspective, traditional systems typically rely on single-modal data (such as visual images or audio), failing to comprehensively reflect students' real behavior. For example, capturing student posture solely through a camera makes it difficult to combine it with their speech engagement and interaction response speed for comprehensive analysis, leading to biased behavioral assessment results. While the introduction of multimodal data can enhance the information dimension, different modal data (such as posture vectors, speech spectrum features, and interaction response latency) exhibit heterogeneity in spatiotemporal dimensions. Achieving effective fusion of cross-modal data has become a key technical bottleneck. Existing fusion methods often employ linear superposition with fixed weights, lacking the ability to adaptively adjust to dynamic data characteristics, and are ill-equipped to handle complex scenarios in the classroom environment where teaching content and student states change over time.

[0004] In terms of feature analysis and modeling, existing systems do not utilize the semantics of teaching content in sufficient depth. Semantic tags of teaching content (such as knowledge point difficulty and learning objectives) are a crucial bridge connecting student behavior and learning outcomes, but traditional methods often treat them as independent variables, failing to establish a hierarchical relationship model between behavioral features and semantic tags. For example, when students learn knowledge points of varying difficulty, the importance weights of their posture, speech, and interaction behaviors should be dynamically adjusted. However, existing technologies lack an effective mechanism for this dynamic prioritization, resulting in the inability to accurately identify student behavioral patterns most relevant to the current teaching content.

[0005] Furthermore, existing classroom behavior analysis systems lack sufficient adaptability. The classroom environment is highly dynamic; students' attention distribution and participation levels change in real time with the progress of teaching and the forms of interaction (such as lecturing, questioning, and group discussions). Traditional algorithms typically employ fixed-parameter models, making it difficult to automatically optimize the calculation process and weight allocation based on real-time data, resulting in poor robustness in complex scenarios. For example, in group discussions, the importance of speech features should be significantly increased, but fixed-parameter models cannot adjust the weights of each modality feature in a timely manner, potentially leading to delayed or inaccurate analysis results.

[0006] In terms of computational efficiency and real-time performance, the high dimensionality and complex processing of multimodal data place higher demands on the system's computational resources. Existing technologies often employ traditional algorithms with high computational complexity in feature fusion and modeling, resulting in significant processing delays and making it difficult to meet the needs of real-time classroom analysis. For example, traditional density clustering algorithms lack adaptive optimization mechanisms for the clustering process when processing dynamic semantic tags, leading to low computational efficiency and an inability to extract semantic core clusters in a timely manner, thus affecting the timeliness of behavioral analysis. Summary of the Invention

[0007] The purpose of this invention is to provide a classroom behavior analysis system based on adaptive algorithms to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides the following technical solution: a classroom behavior analysis system based on an adaptive algorithm, the system comprising:

[0009] The detection module is used to collect multimodal behavioral data in the classroom environment and identify student terminal devices that are currently active.

[0010] The parameter extraction module is used to obtain the real-time posture vector, speech spectrum features, interaction response latency, and semantic tags of the teaching content of the target student.

[0011] The feature integration module is used to perform cross-modal fusion on the real-time pose vector, speech spectrum features, interaction response delay and semantic label respectively, to generate pose fusion features, speech fusion features, interaction fusion features and semantic association features;

[0012] The decision generation module is used to perform hierarchical association modeling and dynamic priority ranking of the posture fusion features, speech fusion features, and interaction fusion features with the semantic association features, generate the behavior matching degree of the target student, and take the data stream with the highest matching degree as the target analysis object.

[0013] Preferably, cross-modal fusion is performed on the real-time pose vector, speech spectrum features, interaction response delay, and semantic tags, including: spatiotemporal decomposition of the real-time pose vector to extract static pose components and dynamic trajectory components, and linear superposition to generate the pose fusion features; moving mean difference of the speech spectrum features to separate the fundamental frequency envelope and harmonic distribution, and standardized concatenation to generate the speech fusion features; piecewise exponential smoothing of the interaction response delay to generate a delay change gradient as the interaction fusion features; and dynamic density clustering of the semantic tags to extract semantic core clusters as the semantic association features.

[0014] Preferably, dynamic density clustering is performed on the semantic tags to extract semantic core clusters, including: performing a three-layer density reachability analysis on the semantic tags, calculating the contour coefficient of each cluster class, and filtering and generating core semantic vectors as the semantic association features according to a density threshold.

[0015] Preferably, the hierarchical association modeling and dynamic priority ranking of the posture fusion features, speech fusion features, and interaction fusion features with the semantic association features to generate the behavior matching degree of the target student includes: performing spatiotemporal association calculation based on multi-head attention mechanism on the posture fusion features and the semantic association features to generate posture-semantic association coding; performing the spatiotemporal association calculation based on multi-head attention mechanism on the speech fusion features and the semantic association features to generate speech-semantic association coding; performing the spatiotemporal association calculation based on multi-head attention mechanism on the interaction fusion features and the semantic association features to generate interaction-semantic association coding; calculating posture weight coefficients, speech weight coefficients, and interaction weight coefficients according to the posture-semantic association coding, speech-semantic association coding, and interaction-semantic association coding respectively; and performing sliding weighted fusion based on the posture weight coefficients, speech weight coefficients, and interaction weight coefficients to generate the behavior matching degree.

[0016] Preferably, performing spatiotemporal association computation based on a multi-head attention mechanism on the pose fusion features and the semantic association features to generate pose-semantic association codes includes: constructing a block association tensor between the pose fusion features and the semantic association features, enhancing local features with deformable convolutional kernels, and generating the pose-semantic association codes through a temporal residual network.

[0017] Preferably, constructing the block association tensor between the pose fusion feature and the semantic association feature includes: performing an overlapping frame splitting operation on the pose fusion feature to generate multi-channel sub-vectors, and stacking each sub-vector with the semantic association feature after cosine similarity matching to form the block association tensor.

[0018] Preferably, the local features are enhanced using deformable convolution kernels, including: adaptively adjusting the offset of the convolution kernel according to the energy distribution of the block-related tensor, performing dilated convolution on the tensor, and then superimposing channel attention to generate enhanced feature maps.

[0019] Preferably, generating the pose-semantic association code through a temporal residual network includes: inputting the enhanced feature map into a depthwise separable convolutional layer along the time dimension, extracting the spatiotemporal dependencies, and then generating the pose-semantic association code by dimensionality reduction through a pooling compression layer.

[0020] Preferably, calculating the pose weight coefficient, speech weight coefficient, and interaction weight coefficient based on the pose-semantic association coding, speech-semantic association coding, and interaction-semantic association coding respectively includes: inputting the pose-semantic association coding into a probability distribution calibrator and outputting the pose weight coefficient according to a preset distribution template; inputting the speech-semantic association coding into a threshold adaptive network and generating the speech weight coefficient through linear combination after threshold segmentation; and inputting the interaction-semantic association coding into a sparse encoder and generating the interaction weight coefficient through normalization after feature selection.

[0021] Preferably, the step of generating the interaction weight coefficients by normalization after feature selection includes: inputting the interaction-semantic association code into a gated linear unit for feature filtering, using a sliding quantile to dynamically scale the output value, and finally generating the interaction weight coefficients in the interval of 0 to 1 through a piecewise saturation function mapping.

[0022] Preferably, after performing a three-layer density reachability analysis on the semantic tags, the contour coefficient of each cluster is calculated, including: generating candidate clusters based on dynamic neighborhood radius, calculating the contour coefficient through intra-cluster dispersion and cross-cluster overlap, and using gradient descent to iteratively optimize the density threshold to screen core semantic vectors that meet the confidence interval.

[0023] Preferably, the offset of the convolution kernel is adaptively adjusted according to the energy distribution of the block-related tensor, including: calculating the energy entropy value of each channel of the tensor, generating an offset adjustment matrix according to the entropy value distribution, and updating the sampling position of the convolution kernel through bilinear interpolation to achieve local feature focusing.

[0024] Preferably, the output value is dynamically scaled using a sliding quantile, which includes: calculating a dynamic quantile threshold based on the historical interaction weight sequence within the time window, performing a piecewise linear transformation on the current output value, and adjusting the scaling magnitude through an adaptive gain factor.

[0025] Compared with the prior art, the beneficial effects of the present invention are:

[0026] In terms of data acquisition and processing, the detection module achieves comprehensive perception of the classroom scene by collecting multimodal behavioral data and identifying active student terminals. The parameter extraction module acquires multi-dimensional data such as posture vectors, speech spectrum features, interaction response delays, and semantic tags in real time, providing a rich information foundation for subsequent analysis. The feature integration module adopts targeted cross-modal fusion algorithms, such as performing spatiotemporal decomposition on real-time posture vectors to extract static posture components and dynamic trajectory components, and linearly superimposing them to generate posture fusion features. This processing method fully explores the spatiotemporal features within single-modal data, avoiding the one-sidedness of traditional single-modal analysis. Moving mean difference is applied to speech spectrum features to separate the fundamental frequency envelope and harmonic distribution, and standardized concatenation is used to generate speech fusion features, effectively improving the resolution and distinguishability of speech features. Piecewise exponential smoothing is applied to interaction response delays to generate delay change gradients as interaction fusion features, which can sensitively capture the dynamic changing trends of student interaction behavior. Dynamic density clustering of semantic tags is used to extract semantic core clusters as semantic association features. Through three-layer density reachability analysis and silhouette coefficient calculation, combined with gradient descent method to iteratively optimize the density threshold, the accuracy and timeliness of semantic feature extraction are ensured, laying a solid foundation for the association analysis of behavioral features and teaching content.

[0027] At the feature association and modeling level, the decision generation module constructs a hierarchical association model between pose, speech, and interaction fusion features and semantic association features through spatiotemporal association calculation based on a multi-head attention mechanism. For example, it constructs a block association tensor for pose fusion features and semantic association features, adaptively adjusts the convolution kernel offset through deformable convolution kernels to focus on local features, and uses a temporal residual network to extract spatiotemporal dependencies to generate pose-semantic association codes. This approach not only enhances the spatial correlation between features but also captures dynamic changes in the temporal dimension, enabling the model to more accurately understand the spatiotemporal correlation between student behavior and teaching content. Similarly, the association calculation between speech and interaction fusion features and semantic association features generates corresponding weight coefficients through a threshold adaptive network and a sparse encoder, respectively, and combines sliding weighted fusion to generate behavior matching degrees, realizing dynamic priority ranking of multimodal features. This adaptive algorithm-based weight allocation mechanism can automatically adjust the importance of each modality feature according to real-time changes in teaching content and student status, significantly improving the accuracy and adaptability of behavior matching degree calculation.

[0028] In terms of system adaptability, a series of adaptive algorithms, such as dynamic density clustering, adaptive adjustment of deformable convolutional kernel offsets, and dynamic scaling of sliding quantiles, enable the system to adjust computational parameters and processing flow in real time based on the energy distribution and historical sequence characteristics of the data. For example, an offset adjustment matrix is ​​generated based on the energy entropy value of the block association tensor to dynamically update the sampling position of the convolutional kernel, thereby focusing on key feature regions; a dynamic quantile threshold is calculated based on the historical interaction weight sequence within the time window, and piecewise linear transformation and adaptive gain adjustment are performed on the output value to ensure dynamic optimization of the weight coefficients. These adaptive mechanisms enable the system to efficiently cope with dynamic changes in the classroom environment, improving the robustness and generalization ability of the model.

[0029] In terms of real-time performance and computational efficiency, lightweight network structures such as depthwise separable convolutional layers and pooling compression layers, along with efficient algorithms like gradient descent and sparse coding, reduce computational complexity while maintaining analytical accuracy. For example, the temporal residual network extracts spatiotemporal dependencies through depthwise separable convolutions and then uses pooling compression layers for dimensionality reduction, effectively reducing computational load. The sliding weighted fusion mechanism, based on a pre-defined distribution template and dynamic weight coefficients, avoids complex global optimization calculations and ensures real-time generation of behavior matching scores. These techniques enable the system to meet the stringent requirements of real-time analysis in classroom scenarios, providing strong support for teachers to adjust their teaching strategies promptly. Attached Figure Description

[0030] Figure 1 This is a schematic diagram illustrating the working principle of the classroom behavior analysis system based on adaptive algorithms described in this invention.

[0031] Figure 2 A flowchart for dynamic density clustering of semantic tags and extraction of core semantic vectors;

[0032] Figure 3 A flowchart for hierarchical association modeling and dynamic priority ranking of behavioral matching degree;

[0033] Figure 4 A flowchart for generating pose-semantic association encoding based on a multi-head attention mechanism;

[0034] Figure 5 A flowchart illustrating how deformable convolutional kernels adaptively enhance local features;

[0035] Figure 6 A flowchart for generating pose-semantic association codes for temporal residual networks. Detailed Implementation

[0036] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0037] Please see Figures 1-6 This invention relates to a classroom behavior analysis system based on an adaptive algorithm. The system includes: a detection module, a parameter extraction module, a feature integration module, and a decision generation module. The specific implementation steps are as follows:

[0038] The detection module is used to collect multimodal behavioral data in the classroom environment and identify the student terminal devices that are currently active. Specifically, the detection module uses sensors such as cameras and microphones deployed in the classroom to collect multimodal data such as video images and voice signals from students in real time. At the same time, it uses methods such as network signal monitoring to determine the student terminal devices that are currently active in the classroom environment, such as tablets and laptops used by students.

[0039] The parameter extraction module is used to obtain the real-time posture vector of the target student, speech spectrum features, interaction response latency, and semantic tags of the teaching content. The real-time posture vector is obtained through analysis and processing of the student's video images, reflecting the student's posture information; the speech spectrum features are obtained through spectrum analysis of the student's speech signal; the interaction response latency refers to the response time when the student interacts with the teaching content through the terminal device; and the semantic tags of the teaching content are tags extracted after semantic analysis of the text, images, and other teaching content involved in the teaching process.

[0040] The feature integration module is used to perform cross-modal fusion on real-time pose vectors, speech spectrum features, interaction response delays and semantic labels respectively, to generate pose fusion features, speech fusion features, interaction fusion features and semantic association features.

[0041] The decision generation module is used to perform hierarchical association modeling and dynamic priority ranking of posture fusion features, speech fusion features, interaction fusion features and semantic association features to generate the behavior matching degree of the target student, and take the data stream with the highest matching degree as the target analysis object.

[0042] The present invention will be further described below with reference to Examples 1 to 5: Example 1

[0043] In this embodiment, the specific process by which the feature integration module performs cross-modal fusion on real-time pose vectors, speech spectrum features, interaction response latency, and semantic tags is as follows:

[0044] When processing real-time pose vectors, they are first decomposed using a spatiotemporal decomposition algorithm. This algorithm analyzes the characteristics of the pose vector in both spatial and temporal dimensions, separating it into static pose components and dynamic trajectory components. Static pose components represent the student's fixed posture attributes at a given moment, such as whether their sitting posture is upright and the position of their arms; dynamic trajectory components record the sequence information of posture changes over time, such as head rotation trajectories and hand movement paths. During decomposition, the spatial dimension analysis is based on the image coordinate system, using a skeletal keypoint detection algorithm to identify the coordinate positions of various parts of the student's body and construct a spatial pose vector. The temporal dimension analysis uses differential calculations across consecutive frames to track the rate of change of keypoint coordinates and trajectory curves, generating a temporal series feature vector. After decomposition, the static pose components and dynamic trajectory components are linearly superimposed, combined according to preset weighting coefficients (e.g., 0.6 for static components and 0.4 for dynamic components) to form a pose fusion feature that incorporates both spatiotemporal characteristics. This fusion feature retains the spatial attributes of the student's current posture while also reflecting the temporal dynamics of posture changes, providing more comprehensive basic data for subsequent behavior analysis.

[0045] For processing speech spectral features, a moving average difference algorithm is used to separate the fundamental frequency envelope and harmonic distribution. First, the speech signal's spectral data is divided into multiple overlapping time windows (e.g., each window is 50 milliseconds long with a 30% overlap), and the moving average of the spectral data is calculated within each window. The moving average smooths high-frequency noise and highlights the low-frequency trend components of the spectrum, i.e., the fundamental frequency envelope, which reflects the basic pitch changes of speech. Then, a difference operation is performed between the original spectral data and the moving average result to obtain the high-frequency fluctuation components, i.e., the harmonic distribution, which includes the timbre and tone quality features of speech. After separating the fundamental frequency envelope and harmonic distribution, both are standardized, mapping the data range uniformly to the [-1,1] interval to eliminate dimensional differences between different feature dimensions. After standardization, the fundamental frequency envelope and harmonic distribution are concatenated in chronological order to form a one-dimensional speech fusion feature vector. This vector integrates the pitch and timbre features of speech, facilitating subsequent association analysis with semantic tags.

[0046] When handling interactive response latency, a piecewise exponential smoothing algorithm is used to generate latency change gradients. First, based on the type of interactive operation (e.g., click, swipe, input) and the time interval, the interactive response latency sequence is divided into multiple time periods, each corresponding to a different teaching segment or interactive scenario (e.g., classroom questioning, slide turning). Within each time period, the exponential smoothing algorithm is applied to process the latency data. By assigning higher weights to recent data, the exponential smoothing algorithm effectively captures short-term latency trends. Specifically, the initial smoothing value is set as the initial latency value for that time period, and then the smoothing is applied according to the recursive formula... Calculate the smoothed value, where This is a smoothing coefficient (range 0.1-0.3). This represents the delay value at the current moment. The value is the smoothed value from the previous moment. By calculating the difference between adjacent smoothed values, the time delay gradient is obtained, which reflects the rate of change of the interactive response speed over time. The time delay gradients of each time period are concatenated to form an interactive fusion feature, which can intuitively reflect the changing pattern of students' interactive response speed in different teaching stages.

[0047] For semantic tag processing, a dynamic density clustering algorithm is used to extract semantic core clusters. First, a three-layer density reachability analysis is performed on the semantic tag set: the first layer, based on a global density threshold, filters out core points whose density exceeds the initial threshold, forming preliminary candidate clusters; the second layer, based on the preliminary clusters, adjusts the density threshold (e.g., reducing it by 10%-20%), analyzes the density reachability relationships between core points, and merges overlapping or adjacent clusters; the third layer introduces density reachability analysis of boundary points to determine whether a boundary point belongs to a core cluster, further optimizing the clustering structure. After completing the three-layer analysis, the silhouette coefficient of each cluster is calculated using the formula... Calculation, where Let i be the average distance from sample i to other samples in the same cluster. The silhouette coefficient is the average distance from sample i to the nearest sample in a different cluster. The silhouette coefficient ranges from -1 to 1; a larger value indicates a denser cluster and higher inter-cluster discrimination. Clusters with silhouette coefficients greater than a preset threshold (e.g., 0.5) are selected based on their silhouette coefficients. The centroids of these clusters are used as core semantic vectors to form semantic association features. This feature can extract the core semantics from the teaching content, providing a semantic benchmark for subsequent behavior matching calculations.

[0048] During the entire feature integration process, the processing of each modality data follows the design concept of cross-modal fusion. Through targeted algorithms, feature extraction and fusion are performed on different modality data to ensure that the generated pose fusion features, speech fusion features, interaction fusion features, and semantic association features can accurately reflect the classroom behavior state of students and the semantic association of teaching content, providing reliable input data for the hierarchical modeling and dynamic priority ranking of the decision-making generation module. The algorithm implementation of each step is based on existing mature signal processing and machine learning technologies. Through parameter tuning and process design, the effectiveness and stability of feature integration are guaranteed. Embodiment 2

[0049] The process of the decision-making generation module performing hierarchical association modeling and dynamic priority ranking on the pose fusion features, speech fusion features, interaction fusion features, and semantic association features to generate the matching degree of the target student's behavior is as follows:

[0050] First, spatio-temporal association calculation is performed on the pose fusion features and semantic association features. The pose fusion features are input into the overlapping frame division module. The sliding window technique is used to segment the feature vectors. The window length is set to T time steps, and the step size is s (s < T), generating multiple overlapping sub-vector sequences. Each sub-vector contains the pose features at T consecutive time points, and there is an overlap of T - s time points between adjacent sub-vectors. This overlapping frame division method can retain the temporal continuity of pose changes and avoid information loss. For each sub-vector, its cosine similarity with the semantic association features is calculated. The semantic association features are pre-mapped to a low-dimensional vector space through a word vector model and are in the same feature space as the pose sub-vectors. The cosine similarity calculation is implemented using vectorized operations, and the similarity scores of all sub-vectors with the semantic association features are calculated in parallel through matrix multiplication to form a similarity matrix. The similarity matrix is stacked row by row to construct a block association tensor, where each block corresponds to the similarity distribution between a pose sub-vector and the semantic association features.

[0051] To enhance the local features of the block association tensor, deformable convolution kernels are used for feature extraction. First, the energy entropy value of each channel of the tensor is calculated. The energy entropy reflects the degree of chaos of the feature distribution, and the calculation formula is , where is the normalized energy value of each element within the channel. An offset adjustment matrix is generated according to the entropy value distribution. Regions with higher entropy values correspond to larger offsets, indicating that the features in these regions change complexly and require more attention. The offset adjustment matrix is mapped to the sampling positions of the convolution kernel through bilinear interpolation to achieve adaptive deformation of the standard convolution grid. When calculating the feature response, the deformed convolution kernel can dynamically adjust the sampling point positions according to the distribution of the input features, focusing on key local features. The deformable convolution operation is applied to the block association tensor to generate an enhanced feature map, highlighting the important association regions between the pose and semantics.

[0052] Enhanced feature maps are input into a temporal residual network for spatiotemporal dependency modeling. The network employs a depthwise separable convolutional structure, decomposing standard convolution into depthwise convolutions and pointwise convolutions, reducing the number of parameters while maintaining feature expressiveness. Depthwise convolutions apply a kernel to each input channel individually, capturing local spatial features; pointwise convolutions integrate inter-channel information through 1×1 convolutions, achieving feature dimension transformation. The network contains multiple residual blocks, each consisting of two depthwise separable convolutional layers and a skip connection. The skip connection directly adds input features to the output, mitigating the vanishing gradient problem. Through residual connections, the network learns temporal variation patterns of pose-semantic associations. In the final layer of the network, global average pooling is applied to compress the temporal features, mapping the feature vector at each time step to a fixed-length feature representation, generating pose-semantic association encodings.

[0053] A similar approach is used to process speech fusion features and interaction fusion features. For both speech fusion features and semantic association features, spatiotemporal correlation is calculated using a multi-head attention mechanism. Speech fusion features are segmented into temporal windows, and the feature vectors within each window capture internal dependencies using a self-attention mechanism. Then, mutual attention with semantic association features is calculated to generate a speech-semantic association tensor. Deformable convolution and a temporal residual network are applied to this tensor to generate speech-semantic association codes. The processing flow for interaction fusion features is the same as for semantic association features, involving frame segmentation, attention calculation, deformable convolution, and temporal modeling to generate interaction-semantic association codes.

[0054] The weight coefficients of the three types of associative encoding are calculated. The pose-semantic associative encoding is input into a probability distribution calibrator, which is based on a Gaussian mixture model. First, the probability density function parameters of the encoded data are estimated using the EM algorithm to determine the mean, variance, and weights of multiple Gaussian components. Then, based on a preset reference distribution template (such as a uniform or normal distribution), the Kullback-Leibler divergence between the current distribution and the reference distribution is calculated. By minimizing the divergence, the parameters of the Gaussian mixture model are adjusted to make the output distribution approximate the reference distribution. Finally, the probability density values ​​are extracted from the calibrated distribution as pose weight coefficients, which reflect the degree of matching between pose features and semantic content.

[0055] Speech weight coefficients are generated using a threshold adaptive network. The speech-semantic association code is input into a neural network consisting of multiple fully connected layers. The last layer of the network applies a sigmoid activation function to generate initial weight values. Simultaneously, the network includes an adaptive threshold generation module that dynamically calculates a threshold based on the statistical characteristics of the input code (such as mean and variance). The initial weight values ​​are compared with the threshold; values ​​greater than the threshold are retained, while values ​​less than the threshold are adjusted using linear interpolation. Finally, the adjusted weight values ​​are normalized to generate speech weight coefficients ranging from 0 to 1, reflecting the contribution of speech features to semantic understanding.

[0056] The calculation of interaction weight coefficients involves a more complex processing flow. First, the interaction-semantic association encoding is input into a gated linear unit (GLU). The GLU generates a gating value using a sigmoid function, selectively filtering the input features and suppressing irrelevant information. The filtered features are then dynamically scaled using a sliding quantile method. A sliding window is defined on the time series, and the quantiles of the feature values ​​within the window (e.g., the 0.25 and 0.75 quantiles) are calculated. The scaling factor is determined based on these quantiles. Outliers outside the quantile range are compressed using a piecewise linear function to avoid impacting the overall weights. The scaled features are then adjusted for amplitude using an adaptive gain factor, which is automatically calculated based on the dynamic range of the features. Finally, the processed features are mapped to the 0-1 interval using a piecewise saturation function to generate interaction weight coefficients, which reflect the strength of the association between the interaction behavior and the semantic content.

[0057] In generating the final behavior matching score, the posture-semantic association encoding, speech-semantic association encoding, and interaction-semantic association encoding are weighted and fused according to their respective weight coefficients. Specifically, the three encoding vectors are multiplied element-wise by their respective weight coefficients, and then vector addition is performed to obtain the fused behavior feature vector. This vector is input into a softmax classifier to calculate its similarity score with multiple preset behavior pattern templates. The behavior type corresponding to the template with the highest score is the behavior matching result for the target student. The entire process, through hierarchical association modeling and dynamic priority ranking, comprehensively considers the correlation between the behavioral features of posture, speech, and interaction dimensions and the teaching semantics, ensuring that the generated behavior matching score accurately reflects the student's actual behavioral state in the classroom.

[0058] In implementation, all algorithms employ numerically stable computation methods to avoid gradient explosion or vanishing gradient problems. Key steps such as offset calculation for deformable convolution, parameter estimation for probability distribution calibration, and adaptive threshold generation all utilize iterative optimization methods, ensuring algorithm stability through the setting of reasonable convergence conditions. Batch normalization and dropout techniques are applied during network training to improve the model's generalization ability. Data transfer between modules uses tensor format to ensure data consistency and traceability during data flow. Through this systematic design and implementation, the decision generation module can efficiently and accurately complete the behavior matching degree calculation task, providing a reliable decision-making basis for the classroom behavior analysis system. Example 3

[0059] When processing semantic tags, the system employs a dynamic density clustering algorithm to extract semantic core clusters. First, all semantic tags are mapped to a high-dimensional vector space, with each tag represented as a point in the space. Based on the density characteristics of the spatial point distribution, the algorithm dynamically adjusts the neighborhood radius parameter to cluster points that are density-connected. In the initial stage, the algorithm calculates the distance distribution between all point pairs and determines the minimum and maximum neighborhood radii based on the distribution characteristics. For each point, it starts by trying the minimum neighborhood radius, counting the number of points within the neighborhood. If the number of points exceeds a preset minimum point threshold, the point is considered a core point, and the neighborhood radius is gradually expanded until the number of points within the neighborhood no longer increases significantly or the maximum neighborhood radius limit is reached.

[0060] During cluster growth, the algorithm employs a layer-by-layer expansion strategy. First, all core points are identified, each serving as a seed for an initial cluster. For each seed point, based on its currently determined neighborhood radius, all points within its neighborhood are collected to form a preliminary cluster. Then, it is checked whether any new core points exist among these points. If so, the neighborhood is expanded around the new core point, and points within that neighborhood are added to the cluster. This process iterates until no new core points can be found or the termination condition is met.

[0061] To evaluate clustering quality, the algorithm calculates the cohesion and separation of each cluster. Cohesion is measured by the average distance between points within a cluster, while separation is measured by the minimum distance between points in different clusters. By comparing clustering results under different neighborhood radii, the algorithm selects a clustering scheme with high cohesion and high separation. This dynamic neighborhood radius mechanism allows the algorithm to adapt to the clustering needs of regions with different densities, using smaller neighborhood radii in high-density regions and larger neighborhood radii in low-density regions.

[0062] After generating candidate clusters, the system further calculates intra-cluster dispersion and cross-cluster overlap. Intra-cluster dispersion reflects the density of points within a cluster and is evaluated by calculating the average distance from each point to the cluster center. Cross-cluster overlap measures the similarity between different clusters and is determined by calculating the proportion of shared points between clusters and the degree of overlap in the distance distribution of point pairs. The silhouette coefficient takes both metrics into account; a higher value indicates denser intra-cluster distribution and higher inter-cluster separation.

[0063] The system employs gradient descent to iteratively optimize the density threshold. Initially, a high density threshold is set, retaining only points in high-density regions as core points. The threshold is then gradually decreased, generating new candidate clusters with each iteration. In each iteration, the average silhouette coefficient of all clusters at the current threshold is calculated and used as the optimization objective function. During gradient descent, the threshold descent step size is adjusted according to the rate of change of the objective function. Iteration stops when the change in the objective function is less than a preset convergence threshold.

[0064] When selecting core semantic vectors, the system first sorts all clusters based on their silhouette coefficients, selecting clusters with silhouette coefficients higher than a set threshold. For each qualified cluster, its centroid vector is calculated as the representative semantic vector for that cluster. Then, a confidence interval test is applied to calculate the distribution confidence of points in the neighborhood of each semantic vector. Only semantic vectors with confidence scores higher than a preset level are retained as core semantic vectors. This selection mechanism ensures that the final extracted semantic core clusters accurately represent the key themes and concepts of the teaching content.

[0065] In its implementation, the system employs efficient data structures to accelerate density calculations. By constructing a KD-tree to partition the high-dimensional space, potential nearest neighbors can be quickly located when calculating the neighborhood of a point, reducing unnecessary distance calculations. For large-scale semantic label data, the system adopts a batch processing strategy, dividing the data into multiple small batches for clustering, and then merging the results, effectively reducing memory consumption.

[0066] The entire clustering process is completed in the feature integration module, and the generated core semantic vectors are output as semantic association features. These features retain the core semantic information of the teaching content, remove redundancy and noise, and provide high-quality input for hierarchical association modeling in the subsequent decision generation module. By dynamically adjusting clustering parameters and iteratively optimizing, the system ensures the accuracy and stability of semantic core cluster extraction, and can adapt to the semantic analysis needs of different disciplines and teaching scenarios. Example 4

[0067] When constructing the block association tensor between pose fusion features and semantic association features, the system first preprocesses the pose fusion features. The pose fusion features are multi-dimensional vector sequences linearly superimposed from static pose components and dynamic trajectory components, containing information about students' pose changes in the classroom. The preprocessing stage employs overlapping framing technology to divide the continuous pose fusion feature sequence into multiple fixed-length sub-sequences. The framing process uses a sliding window mechanism, with the window length set to cover the time span of typical pose change cycles, for example, the number of feature frames corresponding to 5 seconds. The window sliding step is smaller than the window length, forming overlapping regions. The overlap rate is typically set to 50% to ensure sufficient contextual information is retained between adjacent sub-vectors. This overlapping framing method captures both local pose change features and preserves the temporal continuity of pose changes.

[0068] After each sub-vector is generated, the system standardizes it to eliminate dimensional differences between features of different dimensions. Standardization uses the Z-score method to calculate the mean and standard deviation of each feature dimension, mapping the feature values ​​to a distribution with a mean of 0 and a standard deviation of 1. The standardized sub-vector is treated as a multi-channel feature vector, with each channel corresponding to different dimensions of pose information, such as joint angles and displacement velocities.

[0069] Next, the system performs similarity matching between each standardized sub-vector and the semantic association feature. The semantic association feature is a core semantic vector extracted from the teaching content through dynamic density clustering, representing the theme and key concepts of the current teaching segment. The matching process uses cosine similarity calculation, treating the sub-vector and the semantic association feature as vectors in a high-dimensional space, and measuring the similarity by calculating the cosine value of the angle between them. The cosine similarity value ranges from [-1, 1], with a value closer to 1 indicating that the two vectors are more similar in direction, meaning that the pose feature is more relevant to the semantic content.

[0070] To improve matching efficiency, the system employs a vectorized computation method, forming matrices from all sub-vectors and expanding semantic association features into matrices of the same dimension. Matrix multiplication is then used to calculate the similarity scores between all sub-vectors and the semantic association features in a single operation. This batch computation method fully leverages the parallel processing capabilities of modern computing devices, significantly improving computation speed.

[0071] After matching, the system stacks the obtained similarity scores in chronological order of the sub-vectors. Due to the use of overlapping frames, there is temporal continuity between adjacent sub-vectors. Therefore, the stacked similarity scores form a two-dimensional matrix, where each row corresponds to the similarity distribution between a sub-vector and the semantic association feature, and each column corresponds to the similarity change at different time points. This two-dimensional matrix is ​​regarded as the basic unit of the block association tensor.

[0072] To construct a complete chunked association tensor, the system further expands the two-dimensional similarity matrix along the channel dimension. Each channel corresponds to a different dimension of pose features, such as head pose and hand movements. For each channel, the above-mentioned frame segmentation, normalization, similarity calculation, and stacking process is repeated to obtain multiple two-dimensional similarity matrices. These matrices are then concatenated along the third dimension to form a three-dimensional chunked association tensor. Each slice of this tensor represents the change in the association strength between a specific pose dimension and semantic content over time.

[0073] During the construction process, the system also considers the ambiguity of semantic association features. A concept in the teaching content may correspond to multiple postures, so the system performs multimodal expansion on each semantic association feature, generating multiple semantic variant vectors. Each variant vector represents a different semantic emphasis of the same concept. For example, the concept of "listening attentively" may be expanded into multiple variant vectors such as "eye focus," "leaning forward," and "taking notes." When calculating similarity, each sub-vector is matched with the original semantic association feature and all its variant vectors, resulting in a set of similarity scores. The system performs weighted fusion of these scores, with the weights pre-set according to the importance of the variant vectors, ultimately obtaining a comprehensive similarity score used to construct the block association tensor.

[0074] To enhance the expressive power of the block association tensor, the system also performs feature enhancement processing. First, a Gaussian smoothing filter is applied to denoise the tensor, reducing the impact of random noise on association analysis. Then, histogram equalization is used to enhance the contrast of the tensor, highlighting strongly correlated regions. These preprocessing operations enable the block association tensor to more clearly reflect the relationship between pose features and semantic content.

[0075] The constructed block association tensor, serving as the foundational representation of pose-semantic association, is passed to subsequent deformable convolutional modules for local feature enhancement. Each element of the block association tensor contains information on the association strength between a specific time point, a specific pose dimension, and semantic content, providing a rich feature base for subsequent behavior analysis. Through this method of constructing block association tensors, the system can effectively capture the temporal and semantic associations between pose changes and teaching content, laying a solid foundation for accurately analyzing students' classroom behavior.

[0076] In practical applications, the system dynamically adjusts the construction parameters of the block association tensor based on changes in the teaching content. For example, when explaining abstract concepts, it may pay more attention to changes in students' facial expressions and eye contact; in practical operation sessions, it may focus more on the correlation between hand movements and operation steps. This dynamic adjustment mechanism enables the system to adapt to the needs of different teaching scenarios, improving the accuracy and relevance of behavior analysis.

[0077] The entire process of constructing the segmented association tensor is completed in the feature integration module, which is a key step in the classroom behavior analysis system. By effectively associating pose features with semantic content, the system can mine valuable behavioral patterns from multimodal data, providing teachers with more comprehensive and in-depth classroom behavior analysis results to assist in teaching decisions and improvements. Example 5

[0078] When performing local feature enhancement on the block-related tensor, the system employs deformable convolutional kernels. The core of deformable convolutional kernels lies in their ability to adaptively adjust the sampling position based on the distribution of input features, thereby capturing local features more accurately. The system first calculates the energy distribution of each channel of the block-related tensor, which reflects the spatial concentration of features. By analyzing the energy distribution, the system can determine which regions contain more important feature information and require greater attention.

[0079] To quantify energy distribution, the system performs statistical analysis on each channel of the block-related tensor. For each channel, the sum of squares of its pixel values ​​is calculated, and this sum is taken as the energy value of that channel. Then, the energy values ​​of all channels are normalized to a range between 0 and 1. The normalized energy values ​​form an energy distribution map, where regions with larger values ​​indicate a more concentrated concentration of feature energy and may contain more important semantic information.

[0080] Based on the energy distribution map, the system generates an offset adjustment matrix. Each element of this matrix corresponds to a sampling point of the deformable convolution kernel, and the element value represents the offset of that sampling point relative to the standard convolution grid position. The magnitude and direction of the offset are determined by the energy distribution map. Specifically, sampling points corresponding to regions with higher energy values ​​will receive larger offsets, allowing them to deviate from the standard grid position and sample more critical feature regions.

[0081] When calculating the offset, the system employs bilinear interpolation. Bilinear interpolation is a technique that performs continuous sampling on a two-dimensional grid, enabling the calculation of interpolation results at any position based on the values ​​of the four nearest surrounding points. Through bilinear interpolation, the system can obtain energy values ​​at non-integer coordinate positions on the energy distribution map, thereby generating a more refined offset adjustment matrix. This method ensures the continuity and smoothness of the offset calculation, avoiding feature extraction instability caused by sudden changes in the sampling point positions.

[0082] After generating the offset adjustment matrix, the system applies it to the sampling process of the deformable convolution kernel. In standard convolution, the kernel samples the input feature map according to a fixed grid pattern. In deformable convolution, the sampling positions are dynamically adjusted based on the offset adjustment matrix. Specifically, the actual position of each sampling point is equal to its standard grid position plus the corresponding offset. In this way, the deformable convolution kernel can adaptively focus on important local regions in the input features, enhancing the ability to extract key features.

[0083] When processing enhanced feature maps, the system inputs them into depthwise separable convolutional layers along the time dimension. Depthwise separable convolution is an efficient convolutional structure that decomposes standard convolution into two independent operations: depthwise convolution and pointwise convolution. Depthwise convolution applies a convolution kernel to each channel of the input features individually, capturing the spatial features within each channel; pointwise convolution uses a 1×1 convolution kernel to linearly combine the outputs of the depthwise convolution, integrating information between channels. This decomposition significantly reduces the number of model parameters and computational cost while maintaining good feature extraction capabilities.

[0084] In the depthwise separable convolutional layer, the system designs multiple convolutional blocks, each consisting of a deep convolutional layer, a batch normalization layer, and a ReLU activation function. The batch normalization layer accelerates model training and improves stability by normalizing the input data, ensuring that the input to each neural network layer maintains the same distribution. The ReLU activation function introduces non-linear characteristics into the model, enhancing its expressive power. By stacking multiple such convolutional blocks, the system can extract and enhance multi-level spatiotemporal dependencies in the feature maps.

[0085] After extracting the spatiotemporal dependencies, the system performs dimensionality reduction on the features using a pooling compression layer. This layer employs global average pooling, averaging the spatial dimensions of the feature map for each channel, compressing the two-dimensional feature map of each channel into a single scalar value. This dimensionality reduction method effectively preserves the key information of the features while significantly reducing the amount of data and lowering the computational complexity of subsequent processing.

[0086] Global average pooling has the advantage of translation invariance, meaning that the pooling result remains unchanged regardless of how the features are translated in space. This property makes the system insensitive to the location of pose changes, allowing it to focus more on the essential features of the pose. Through global average pooling, the system transforms the enhanced feature map into a fixed-length feature vector, which contains the core information of the pose-semantic association.

[0087] In the design of the temporal residual network, the system employs residual connection technology. Residual connection is a technique that adds the input directly to the output through skip connections. It effectively alleviates the vanishing gradient problem in deep neural networks, allowing the network to be trained to greater depths. In the temporal residual network, each residual block contains two depthwise separable convolutional layers and one residual connection. The input features are first transformed through the two convolutional layers and then added to the original input to obtain the output of the residual block. This structure allows the network to learn the residual information of the input features, rather than directly learning the target mapping, thus making it easier to train and optimize.

[0088] To further enhance the model's ability to capture temporal information, a gating mechanism is introduced into the temporal residual network. The gating mechanism selectively focuses on key parts of the sequence by learning importance weights for the input sequence. Specifically, the system uses a Gated Recurrent Unit (GRU) to process temporal information. A GRU is a recurrent neural network that controls the flow of information through update and reset gates. The update gate determines how much of the hidden state from the previous time step needs to be passed to the current time step, while the reset gate determines how much of the hidden state from the previous time step needs to be ignored. In this way, the GRU can effectively capture long-term dependencies in the sequence and model temporal patterns of pose changes.

[0089] In practical applications, the system dynamically adjusts the parameters of the deformable convolutional kernel and the temporal residual network based on the characteristics of the input data. For example, for rapidly changing poses and actions, the system increases the offset range of the deformable convolutional kernel, enabling it to adapt more flexibly to changes in features; for long-term teaching processes, the system adjusts the hidden layer size and time step of the GRU to better capture long-term temporal dependencies. This dynamic adjustment mechanism allows the system to adapt to different teaching scenarios and behavioral patterns, improving the accuracy and robustness of behavior analysis.

[0090] The entire feature enhancement and temporal modeling process is completed in the decision generation module, with the generated pose-semantic association code serving as a crucial input for subsequent behavior matching calculations. Through local feature enhancement using deformable convolutional kernels and temporal modeling using temporal residual networks, the system can extract more representative and discriminative features from multimodal data, providing a solid foundation for accurately analyzing student classroom behavior. This method, combining spatial feature enhancement and temporal modeling, fully leverages the complementarity of pose and semantic information, enabling a more comprehensive capture of students' behavioral performance and learning status in the classroom.

[0091] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0092] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A classroom behavior analysis system based on an adaptive algorithm, characterized in that, include: The detection module is used to collect multimodal behavioral data in the classroom environment and identify student terminal devices that are currently active. The parameter extraction module is used to obtain the real-time posture vector, speech spectrum features, interaction response latency, and semantic tags of the teaching content of the target student. The feature integration module is used to perform cross-modal fusion on the real-time pose vector, speech spectrum features, interaction response delay and semantic label respectively, to generate pose fusion features, speech fusion features, interaction fusion features and semantic association features; The decision generation module is used to perform hierarchical association modeling and dynamic priority sorting of the posture fusion features, speech fusion features, and interaction fusion features with the semantic association features to generate the behavior matching degree of the target student, and take the data stream with the highest matching degree as the target analysis object. The hierarchical association modeling and dynamic priority ranking of the posture fusion features, speech fusion features, and interaction fusion features with the semantic association features are used to generate the behavior matching degree of the target student. This includes: performing spatiotemporal association calculation based on a multi-head attention mechanism on the posture fusion features and the semantic association features to generate posture-semantic association codes; performing the spatiotemporal association calculation based on a multi-head attention mechanism on the speech fusion features and the semantic association features to generate speech-semantic association codes; performing the spatiotemporal association calculation based on a multi-head attention mechanism on the interaction fusion features and the semantic association features to generate interaction-semantic association codes; calculating posture weight coefficients, speech weight coefficients, and interaction weight coefficients based on the posture-semantic association codes, speech weight coefficients, and interaction weight coefficients, respectively; and performing sliding weighted fusion based on the posture weight coefficients, speech weight coefficients, and interaction weight coefficients to generate the behavior matching degree. Performing spatiotemporal association computation based on multi-head attention mechanism on the pose fusion features and the semantic association features to generate pose-semantic association codes includes: constructing a block association tensor between the pose fusion features and the semantic association features, and using deformable convolutional kernels to enhance local features before generating the pose-semantic association codes through a temporal residual network; Constructing a block association tensor between the pose fusion feature and the semantic association feature includes: performing an overlapping frame splitting operation on the pose fusion feature to generate multi-channel sub-vectors, and stacking each sub-vector with the semantic association feature after cosine similarity matching to form the block association tensor. The method of enhancing local features by using deformable convolution kernels includes: adaptively adjusting the offset of the convolution kernel according to the energy distribution of the block-related tensor, and performing dilated convolution operation on the tensor and then superimposing channel attention to generate enhanced feature maps. The pose-semantic association encoding is generated by a temporal residual network, including: inputting the enhanced feature map into a depthwise separable convolutional layer according to the time dimension, extracting the spatiotemporal dependency, and then reducing the dimensionality of the pose-semantic association encoding by a pooling compression layer.

2. The classroom behavior analysis system based on adaptive algorithm according to claim 1, characterized in that, Cross-modal fusion is performed on the real-time pose vector, speech spectrum features, interaction response delay, and semantic tags, including: spatiotemporal decomposition of the real-time pose vector to extract static pose components and dynamic trajectory components, and linear superposition to generate the pose fusion features; moving mean difference of the speech spectrum features to separate the fundamental frequency envelope and harmonic distribution, and standardized concatenation to generate the speech fusion features; piecewise exponential smoothing of the interaction response delay to generate a delay change gradient as the interaction fusion features; and dynamic density clustering of the semantic tags to extract semantic core clusters as the semantic association features.

3. The classroom behavior analysis system based on adaptive algorithm according to claim 2, characterized in that, Dynamic density clustering is performed on the semantic tags to extract semantic core clusters, including: performing a three-layer density reachability analysis on the semantic tags, calculating the contour coefficient of each cluster class, and filtering by density threshold to generate core semantic vectors as the semantic association features.

4. The classroom behavior analysis system based on adaptive algorithm according to claim 1, characterized in that, The pose weight coefficient, speech weight coefficient, and interaction weight coefficient are calculated based on the pose-semantic association coding, speech-semantic association coding, and interaction-semantic association coding, respectively. The calculation includes: inputting the pose-semantic association coding into a probability distribution calibrator and outputting the pose weight coefficient according to a preset distribution template; inputting the speech-semantic association coding into a threshold adaptive network and generating the speech weight coefficient by linear combination after threshold segmentation; and inputting the interaction-semantic association coding into a sparse encoder and generating the interaction weight coefficient by normalization after feature selection.

5. The classroom behavior analysis system based on adaptive algorithm according to claim 4, characterized in that, The step of generating the interaction weight coefficients by normalization after feature selection includes: inputting the interaction-semantic association code into a gated linear unit for feature filtering, using a sliding quantile to dynamically scale the output value, and finally generating the interaction weight coefficients in the interval of 0 to 1 through a piecewise saturation function mapping.

6. The classroom behavior analysis system based on adaptive algorithm according to claim 3, characterized in that, After performing a three-layer density reachability analysis on the semantic tags, the contour coefficient of each cluster is calculated, including: generating candidate clusters based on dynamic neighborhood radius, calculating the contour coefficient through intra-cluster dispersion and cross-cluster overlap, and using gradient descent to iteratively optimize the density threshold to screen core semantic vectors that meet the confidence interval.

7. The classroom behavior analysis system based on adaptive algorithm according to claim 1, characterized in that, The offset of the convolution kernel is adaptively adjusted according to the energy distribution of the block-related tensor, including: calculating the energy entropy value of each channel of the tensor, generating an offset adjustment matrix according to the entropy value distribution, and updating the sampling position of the convolution kernel through bilinear interpolation to achieve local feature focusing.

8. The classroom behavior analysis system based on adaptive algorithm according to claim 5, characterized in that, The output value is dynamically scaled using a sliding quantile, which includes: calculating a dynamic quantile threshold based on the historical interaction weight sequence within the time window, performing a piecewise linear transformation on the current output value, and adjusting the scaling magnitude through an adaptive gain factor.

Images