A graph-hypergraph collaborative and low-cost adaptive agent optimization method for classroom teaching sentiment analysis
By combining attention with hypergraph temporal model (HAHT) and multiscale surrogate network, the problems of conflict between pairwise relationships and high-order group modeling and high computational cost in multimodal sentiment analysis are solved, and efficient multimodal sentiment analysis is achieved.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- TAIYUAN UNIVERSITY OF SCIENCE AND TECHNOLOGY
- Filing Date
- 2026-05-19
- Publication Date
- 2026-07-14
Smart Images
Figure CN122388694A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of deep learning and multimodal learning, specifically a classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization. Background Technology
[0002] With the widespread application of multimedia sensing technology and intelligent interactive devices in educational settings, a large amount of multimodal temporal data has emerged in classroom teaching, consisting of multi-source information such as teachers' voice tone, facial expressions, body movements, and teaching texts. Multimodal sentiment analysis (MSA) has become an important research topic in the fields of smart classrooms and teacher evaluation. In classroom settings, teachers' teaching emotions exhibit significant characteristics: the local temporal expressions of the same emotional state show obvious differences (e.g., the serious tone and gestures used when explaining key points and difficulties both point to "seriousness and rigor"). However, capturing these local emotional transitions requires the model to be sensitive to fine-grained interactions at temporal nodes, while identifying the overall teaching emotional tone requires smooth induction of group information across teaching segments. There is a conflict between the induction biases of the two—oversensitivity can amplify noise, while oversmoothing can drown out emotional change clues. How to resolve this conflict and adaptively adjust the balance within a unified framework is the core issue of current classroom teaching sentiment analysis. This invention focuses on a lightweight dedicated model for classroom scenarios. Targeting the characteristics of teachers' teaching emotions, it enhances temporal modeling capabilities through explicit design of induction biases to adapt to real-time analysis needs. From a research perspective, MSA time-series modeling methods have continuously evolved, but there are still significant gaps in modeling the two types of key dependencies. Sequence models (such as MulT and MISA) capture local associations between consecutive time steps using cross-modal attention mechanisms, and have strong discriminative power for fine-grained pairwise dependencies such as word modification and intonation transitions. However, attention naturally favors neighboring nodes, making it difficult to effectively model higher-order group dependencies such as multi-node co-evolution. The acquisition of global sentiment context often relies on deep stacking for indirect implementation. Graph neural networks flexibly represent pairwise relationships by constructing explicit topologies, such as LNLN, but can only characterize binary interactions and lacks the ability to express the multi-node group associations that are common in sentiment data. Traditional graph modeling is difficult to cover multiple complex relationships between nodes, while multi-path relationship perception can effectively mine implicit interaction information in the structure. Hypergraph neural networks, which associate multiple nodes simultaneously with hyperedges, are more suitable for this purpose. The need for high-order group modeling has been addressed by works such as MDH and MHN, but the discriminative power of fine-grained pairwise interactions is easily diluted during group aggregation—which is precisely the key to capturing emotional abrupt changes such as irony and transitions. In addition, the above hypergraph methods still rely heavily on predefined time segments for construction, and data-driven structural learning has not yet been fully realized. In summary, existing methods either focus on the local sensitivity of pairwise relationships and are difficult to achieve smooth global induction, or emphasize the high-order aggregation of hypergraphs and weaken the fine-grained discriminative power. Effective collaboration between the two within the same framework is still a gap. This leads to three core difficulties: (1) Pairwise relationship modeling and high-order group modeling conflict in induction bias and lack a unified collaborative framework; (2) Predefined time segments limit the model's adaptability to non-uniform time series; (3) View selection and hyperparameter tuning rely on repeated real training, resulting in high computational costs. In addition to the evolution of model structure, at the feature representation level, existing research has explored using multi-view representations such as spectrograms and time-frequency transformations to supplement the original time-series view. However, most of these methods adopt fixed, full-view combination strategies, which are difficult to adaptively select the optimal view subset based on data characteristics. This can easily introduce redundant noise and increase computational overhead. At the model tuning level, existing MSA methods still generally rely on human experience or grid search for hyperparameter selection, and repeated real training leads to high time costs. Although surrogate model-assisted optimization has made some progress in the field of evolutionary computation, there is still a lack of systematic research on lightweight, multi-scale surrogate optimization frameworks for deep learning models in MSA tasks. To address the aforementioned challenges, this invention proposes a Hybrid Attention and Hypergraph Temporal Model (HAHT) and constructs the MS²-Net (Multi-Scale Surrogate Network) multi-scale agent optimization framework. The core design principle of HAHT is to maintain both local sensitivity and global smoothness inductive biases in parallel under multimodal scenarios, and to achieve their dynamic synergy through a data-driven approach. The main contributions are as follows: (1) Multimodal pairwise-high-order collaborative modeling framework: The core contradiction of MSA time series modeling is condensed into the inductive bias conflict between "sensitive capture" and "smooth induction", and an adjustable collaborative modeling framework is designed. The model captures fine-grained pairwise dependencies within a modality with graph attention, describes high-order group associations between modalities with hypergraph diffusion, and achieves an adaptive balance between the two through learnable mixing coefficients, providing an effective way for the unified representation of the two types of relationships in multimodal scenarios. (2) Implicit temporal structure learning: A data-driven temporal structure learning mechanism without predefined time segments was designed. This mechanism forms a dynamic temporal topology through attention weight propagation and achieves cross-modal semantic grouping with hypergraph soft membership. The two are jointly optimized in end-to-end training, so that the discovery of temporal structure shifts from manual pre-setting to data-driven, and enhances the model's adaptability to non-uniform, multi-scale and multi-modal temporal expressions. (3) Agent optimization and efficiency assurance: A multi-scale agent network MS²-Net for joint search of multimodal view combination and hyperparameters was constructed. This network replaces a large amount of real training with lightweight agent evaluation and performs joint pre-screening of view combination and high-dimensional hyperparameters. While ensuring optimization accuracy, it effectively reduces computational overhead and provides efficiency support for high-precision multimodal sentiment analysis. The three tasks mentioned above support each other: problem condensation provides a theoretical anchor for multimodal temporal modeling, implicit structure learning enables cross-modal representation capabilities to be fully utilized without prior knowledge, and surrogate optimization adapts the optimal configuration for the former two at a low cost; the synergy of the three enables HAHT to take into account both local sensitivity and global inductive ability within a unified framework, providing a feasible technical route for balancing accuracy and efficiency in multimodal sentiment analysis. Summary of the Invention
[0003] To address the shortcomings of existing technologies, the present invention aims to provide a classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization. This invention is achieved using the following technical solution: A classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization is characterized by: using a hybrid attention and hypergraph temporal model as the core architecture, and multi-scale agent optimization as an efficiency improvement strategy; constructing a dual-branch system of graph attention and hypergraph diffusion to capture fine-grained pairwise dependencies within classroom text, audio, and video modalities and high-order group associations between modalities respectively; relying on a data-driven mechanism to complete implicit temporal structure learning, breaking free from the limitations of predefined time segments; and employing a multi-scale agent network to complete multi-view representation selection and hyperparameter adaptive optimization, reducing model tuning overhead. First, a hybrid attention and hypergraph temporal model is constructed, dividing the modeling and training process of multimodal sentiment analysis into two stages: a multi-view representation extraction stage and a pairwise, higher-order relation co-training stage. In the multi-view representation extraction stage, multi-view feature mapping is performed on text, audio, and video multimodal temporal data, considering temporal domain, frequency domain, structural correlation, and temporal state. Adaptive fusion of multi-view features is achieved through a gated residual mechanism to obtain a unified multi-granular representation. In the co-training stage, a graph attention (GAT) branch and a hypergraph diffusion (HGNN) branch are constructed in parallel. The GAT branch is responsible for modeling fine-grained pairwise dependencies between multimodal nodes, maintaining local semantic sensitivity. The HGNN branch achieves cross-modal higher-order grouping through learnable soft membership. The smooth aggregation of relationships provides global sentiment constraints. Learnable mixing coefficients are used during training to dynamically adjust the contribution weights of GAT and HGNN based on the temporal characteristics of the input, replacing a fixed fusion strategy. Simultaneously, a multi-scale agent network is activated before model training, replacing extensive real-world training with lightweight agent evaluation to jointly pre-screen multi-view combinations and hyperparameters, reducing search overhead. The training process employs a joint loss function including hyperedge orthogonal regularization and membership entropy regularization to ensure a reasonable hypergraph structure and stable information propagation. Ultimately, the network achieves both local sensitivity capture and global smooth induction within a unified framework, improving the accuracy and training efficiency of non-uniform temporal and multimodal sentiment analysis, suitable for text, audio, and video joint sentiment recognition scenarios. To verify the effectiveness of this invention, the graph-hypergraph collaboration and low-cost adaptive agent optimization classroom teaching sentiment analysis method designed in this invention was compared with other multimodal learning methods on the public datasets MOSI, MOSEI, and CH-SIMS. This invention is rationally designed and applicable to the real-time perception and quantitative analysis of teachers' teaching emotions, students' classroom emotions, concentration status and willingness to interact in classroom teaching scenarios. It provides technical support for intelligent teaching evaluation and classroom quality improvement, and helps the digital transformation of education. Attached Figure Description
[0004] Figure 1 This is a flowchart illustrating the overall process of this invention. Figure 2This diagram illustrates the overall architecture of the hybrid attention and hypergraph temporal model. Figure 3 A schematic diagram illustrating the modeling of pairwise relations, higher-order relations, and collaborative optimization relations; Figure 4 This diagram illustrates the optimization framework for a multi-scale surrogate model. Figure 5 This is a flowchart illustrating a multi-view selection method based on a proxy model. Detailed Implementation
[0005] Many specific details are set forth in the following description in order to provide a full understanding of the invention, but the invention may also be practiced in other ways different from those described herein; obviously, the embodiments in the specification are only some embodiments of the invention, and not all embodiments. The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings; A classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization is characterized by: using a hybrid attention and hypergraph temporal model as the core architecture, and multi-scale agent optimization as an efficiency improvement strategy; constructing a dual-branch system of graph attention and hypergraph diffusion to capture fine-grained pairwise dependencies within classroom text, speech, and video modalities and high-order group associations between modalities respectively; relying on a data-driven mechanism to complete implicit temporal structure learning, thus overcoming the limitations of predefined time segmentation; and employing a multi-scale agent network to complete multi-view representation filtering and hyperparameter adaptive optimization, reducing model tuning overhead; and using the detection performance of the classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization on public datasets MOSI, MOSEI, and CH-SIMS as the evaluation standard in this embodiment. When constructing a hybrid attention and hypergraph temporal model, to further capture fine-grained pairwise dependencies within multiple modalities and high-order group associations between modalities, this invention constructs a dual-branch architecture of graph attention and hypergraph diffusion. It relies on a data-driven mechanism to complete implicit temporal structure learning, and finally employs a multi-scale surrogate network to complete multi-view representation filtering and hyperparameter adaptive optimization, reducing model tuning overhead. Its overall architecture diagram is shown below. Figure 2 As shown; The specific construction process of the hybrid attention and hypergraph temporal model is as follows: First, a hybrid attention and hypergraph temporal model is constructed, dividing the modeling and training process of multimodal sentiment analysis into two stages: a multi-view representation extraction stage and a pairwise, higher-order relation co-training stage. In the multi-view representation extraction stage, multi-view feature mapping of text, audio, and video multimodal temporal data is performed in the time domain, frequency domain, structural correlation, and temporal state. Adaptive fusion of multi-views is achieved through a gated residual mechanism to obtain a unified multi-granular representation. After entering the co-training stage, the graph attention GAT branch and the hypergraph diffusion HGNN branch are constructed in parallel. The GAT branch is responsible for modeling fine-grained pairwise dependencies between multimodal nodes, maintaining local semantic sensitivity discrimination ability. The HGNN branch uses learnable soft membership... This system achieves smooth aggregation of high-order group relationships across modalities, providing global sentiment constraints. Learnable mixing coefficients are used during training to dynamically adjust the contribution weights of GAT and HGNN based on input temporal characteristics, replacing fixed fusion strategies. Simultaneously, a multi-scale agent network is activated before model training, replacing extensive real-world training with lightweight agent evaluation to jointly pre-screen multi-view combinations and hyperparameters, reducing search overhead. The training process employs a joint loss function including hyperedge orthogonal regularization and membership entropy regularization to ensure a reasonable hypergraph structure and stable information propagation. Ultimately, the network achieves both local sensitivity capture and global smooth induction within a unified framework, improving the accuracy and training efficiency of non-uniform temporal and multimodal sentiment analysis, making it suitable for joint sentiment recognition scenarios involving text, audio, and video. The overall framework for the adaptive collaborative modeling of pairwise and higher-order relations described above is as follows: Figure 3 As shown, the specific construction process is as follows: (1) Pairwise temporal dependency modeling based on GAT In the MSA task, the semantic modification relationships between words within a modality and the fine-grained alignment relationships between modalities are essentially pairwise temporal dependencies between nodes. Graph Attention Networks (GATs) assign differentiated weight coefficients to different nodes through an attention mechanism, which can effectively capture such pairwise dependencies and is the core means of characterizing local temporal interactions. For the temporal modeling requirements of multimodal and multi-view features, the traditional GAT is adapted to handle the feature representations of multimodal temporal nodes and retain the attention calculation paths of query, key, and value, ensuring stable gradient propagation and effective scoring of pairwise relationships. Let the set of multimodal temporal features after time-frequency fusion and view fusion be . Temporal characteristics of any mode First, feature mapping of Query, Key, and Value is completed through linear projection: ; in, The projection matrix is learnable; Features of the original multimodal input sequence; To capture heterogeneous temporal relationships from multiple subspaces in parallel, a multi-head attention mechanism is employed to characterize pairwise dependencies between nodes from multiple perspectives. Both the projection matrix and features are divided into H attention heads, i.e. Each attention head independently calculates the attention score between nodes; the pairwise attention score for node i to node j in the h-th attention head is calculated as follows: ; in, Let h be the learnable weight vector of the h-th attention head; Indicates feature concatenation operation; A scaling factor to prevent gradient explosion, used to stabilize gradients; LeakyReLU is a non-linear activation function; Through attention masking Filter invalid nodes (if nodes i and j have no valid connection) ,otherwise The attention scores are normalized using the Softmax function to obtain the pairwise attention weights between nodes. ; Based on the normalized attention weights, the Value features are weighted and aggregated to obtain the pairwise relation feature representation of a single attention head. Then, the results of H attention heads are concatenated to complete the fusion of multi-head attention, ultimately obtaining the pairwise temporal dependency features based on GAT. : ; ; in, The output projection matrix of multi-head attention is used to realize the restoration and fusion of feature dimensions; In end-to-end training, the model autonomously learns the pairwise connection strength between nodes, captures local semantic transitions and intermodal alignment relationships in sentiment sequences, and provides a refined local feature foundation for subsequent high-order relationship modeling. (2) High-order temporal dependency modeling based on HGNN While pairwise relationship modeling can characterize binary interactions between nodes, higher-order group dependencies beyond pairwise relationships are common in MSA (Multi-person Dialogue): emotional resonance in multi-person dialogues, semantic aggregation of multimodal features at the global level, etc. These phenomena involve the collaborative effects of multiple nodes and cannot be effectively modeled by simple pairwise connections. Hypergraph neural networks, with hyperedges as the core unit, can simultaneously associate any number of nodes, making them naturally suitable for such higher-order group interaction needs. Therefore, a hypergraph structure adapted to temporal modeling is constructed to realize the diffusion and expression of higher-order group relationships in temporal data. The core of a hypergraph is the association between nodes and hyperedges. Unlike traditional hypergraphs that rely on predefined rules to construct hard association matrices, this method uses a data-driven soft hypergraph association matrix construction method. It calculates the soft membership degree of a node to a hyperedge by using the similarity between node features and hyperedge prototypes, enabling the hypergraph structure to be learned end-to-end. First, the set of hyperedge prototypes Where E is the number of hyperedges, and is a hyperparameter determined through experimental optimization. Each hyperedge prototype represents the semantic features of a high-order group; for the node features of the h-th attention head... Calculate the similarity score between the node and the prototype of the hyperedge, and use it as the initial membership degree of the node to the hyperedge: ; in, Let be the node of the h-th attention head, i.e., the hyperedge similarity matrix; i is the node index; e is the hyperedge index; The similarity matrix is normalized using the Softmax function to obtain the soft hypergraph correlation matrix. Its elements Let represent the probability that node i belongs to hyperedge e, thus implementing soft assignment from node to hyperedge: ; Compared to hard assignment, soft membership allows nodes to participate in multiple hyperedges in a probabilistic manner, more flexibly characterizing the multi-role attributes of nodes in time series data and avoiding the prior bias introduced by predefined hyperedge structures. Higher-order relationships in a hypergraph are achieved through information propagation between nodes via a hypergraph diffusion matrix, which captures the strength of higher-order associations formed by nodes through shared hyperedges; based on the soft hypergraph association matrix... First, calculate the degree matrix of the hyperedge. degree matrix of nodes ,in Let e represent the number of nodes connected by hyperedge e and the number of hyperedges to which node i belongs, respectively. To achieve smooth information diffusion on the hypergraph, the hypergraph correlation matrix is normalized to obtain a normalized soft hypergraph correlation matrix. Subsequently, the hypergraph diffusion matrix is constructed. Describe the higher-order relationships between nodes achieved through hyperedges: ; in, This represents the strength of the higher-order connection between node i and node j through a common hyperedge e. The larger the value, the higher the probability that the two nodes belong to the same higher-order group. Based on the hypergraph diffusion matrix, high-order information aggregation is performed on the Value features to obtain high-order relational feature representations for a single attention head; similar to GAT, multi-head fusion is used to obtain the final high-order temporal dependency features based on HGNN. ; ; Through hypergraph diffusion, the model extends information interaction from single-node pairs to multi-node group levels, capturing the overall emotional trend of the sentiment sequence and the global collaborative patterns between modalities, complementing pairwise modeling. In multimodal sentiment analysis tasks, the view composition and hyperparameter configuration of HAHT directly affect sentiment recognition performance. However, manual parameter tuning relies on experience and involves significant overhead from repeated real training. Therefore, a multi-scale agent network, MS²-Net, is constructed to replace extensive real training with lightweight agent evaluation, jointly selecting view compositions and key hyperparameters. The overall process is as follows: Figure 4 As shown, the framework mainly includes five modules: initializing the population, selecting reference points, multi-scale network modeling, data partitioning and training, and surrogate-assisted selection. This framework uses a multi-scale surrogate network to jointly pre-screen view combinations and hyperparameters, achieving efficient configuration search with a small number of real training iterations. The specific process of designing its multi-scale agent network is as follows: In HAHT tuning, discrete choices such as view combination and the number of hyperedges determine the model's topology, while continuous hyperparameters such as learning rate and Dropout control the training dynamics on this structure. These two are not independent: increasing the number of hyperedges enriches the associations of higher-order groups but also amplifies gradient propagation noise, requiring a smaller learning rate. View subsets determine the dimensionality and signal-to-noise ratio of input features, directly affecting the optimal range of Dropout values. Agent networks with a single receptive field struggle to simultaneously capture these cross-granular couplings—narrow receptive fields emphasize local dependencies (such as the synergy between learning rate and Dropout), while wide receptive fields emphasize global structure (such as the impact of view combination on overall performance), but in HAHT, they are mutually constraining. To this end, this invention designs a proxy network with multi-scale parallel branches, extracting fine-grained local dependencies, medium-grained associations, and coarse-grained global structures through fully connected branches at three scales: 1, 3, and 5. The 1-scale branch focuses on univariate sensitivity (such as the impact of a single change in the learning rate on performance), the 3-scale branch captures pairwise couplings (such as the joint effect of the learning rate and Dropout), and the 5-scale branch models higher-order interactions (such as the synergistic constraints of view composition, the number of hyperedges, and the learning rate). These three aspects are then fused across scales through feature concatenation, enabling the proxy network to simultaneously complete local sensitivity evaluation and global structure perception in a single forward propagation, explicitly modeling the cross-granular coupling relationship between continuous hyperparticipants and discrete choices in HAHT. The number of neurons in each branch scales dynamically with the dimension D of the decision variable. , , No manual structural adjustments are needed for different problems; each branch consists of a fully connected layer, batch normalization, and ReLU activation, which are then concatenated and output as predicted values via a fusion head; this multi-scale parallel design enables the network to adapt to nonlinear mappings in high-dimensional hybrid search spaces with relatively low computational cost. By combining the aforementioned proxy network with evolutionary optimization, a multi-scale proxy-assisted evolutionary optimization framework is formed. In each iteration: a reference solution is selected based on a screening strategy, the population is stratified and sampled according to dominance relationship, and the proxy network is trained. A prediction error-driven adaptive strategy is used to screen high-value candidate solutions and perform real evaluation. The proxy network is built once and reused throughout the iteration, avoiding the extra overhead of repeated modeling in each generation of traditional methods. This framework improves optimization efficiency and performance through multi-stage collaborative design. The stratified sampling data partitioning method maintains a balanced class distribution between the training and test sets, preventing model bias towards the majority class due to sample imbalance and further ensuring the prediction accuracy of the surrogate model. Simultaneously, the prediction error-driven adaptive sample selection strategy dynamically adjusts sampling rules based on the model's prediction errors on dominant and non-dominant samples, preferentially selecting high-value candidate solutions in areas with reliable model predictions and performing random exploration in unreliable areas, achieving a dynamic balance between global exploration and local development. Furthermore, the framework adopts a multi-scale network model that is built once and reused throughout the iteration process, avoiding the high computational overhead of repeated modeling in each generation of traditional surrogate models and effectively improving overall optimization efficiency. The dependence on views varies across different emotional scenarios: in monologues, the time-frequency patterns of acoustic features are relatively stable, and frequency views often provide effective supplementation; however, in film and television dialogues with frequent scene changes, changes in temporal structure are more critical, and structural views such as GASF / GADF may contribute more. This difference is difficult to exhaustively enumerate through human experience, and using all views consistently would introduce redundant noise; therefore, view selection is modeled as a discrete binary search space {0, 1}. K (K is the number of candidate views), the proxy model quickly evaluates each candidate combination on a lightweight task, and adaptively selects the optimal subset that is sensitive to the current task. The data is fed into a real model for training, and redundant views are removed while retaining key information; the overall process is as follows: Figure 5 As shown; The view selection problem can be modeled as a discrete binary search space; let the total number of candidate views be K, and the view selection vector be defined as: ; in, This indicates that the k-th candidate view is enabled. This indicates that the proxy model is not enabled; the proxy model quickly evaluates each candidate combination on a lightweight proxy task and selects the combination with the best validation performance. This final configuration is then fed into the network for formal training. The proxy model determines whether to use a particular view. It rapidly evaluates candidate combinations on lightweight proxy tasks and selects the combination with the best validation performance. As the final configuration, it is fed into the network for formal training, thereby reducing computational redundancy while maintaining the model's expressive power; To verify the effectiveness of HAHT, it was compared with mainstream methods such as MulT, MISA, BERT-MAG, Self-MM, UniMF, MPLMM, and MDHMHN on CMU-MOSI and CMU-MOSEI. The results are shown in Table 1. Table 1 Performance Comparison of Multimodal Sentiment Analysis Methods ; To verify the contribution of each core module, ablation experiments were conducted on CMU-MOSI, and the results are shown in Table 2. Table 2 Ablation Experiment Results under Different Core Module Combinations ; To explore the contribution of different modalities to emotion recognition and the cross-modal fusion capability of HAHT, the performance of single-modal, dual-modal and trimodal combinations was tested on CMU-MOSI. The results are shown in Table 3. Table 3 Performance Comparison under Different Modes and Mode Combinations ; To test the Pareto front approximation capability of MS²-Net in this type of multi-objective scenario, benchmark experiments were conducted on the DTLZ series test set; the results of the three-objective experiment are shown in Table 4. Table 4. Comparison of the mean (standard deviation) of IGD for each algorithm on the DTLZ test set with 20 and 50-dimensional decision variables for objective 3. ; Note: The significance test uses the Wilcoxon rank-sum test (α=0.05); "+" "-" "=" indicates that the corresponding algorithm is significantly better than, significantly worse than, and has no significant difference from MS²-Net, respectively; the value after "+ / - / =" in the last row of the table is the cumulative number of the three types of results for each algorithm in 14 cases; all cases were run independently 20 times; This invention is based on a 13th Gen Intel(R) Core(TM) i5-13400F CPU and an NVIDIA GeForce RTX4070Ti SUPER GPU, using PyTorch 2.4.1 and CUDA 12.4, and Python 3.8.10. All experiments use binary classification accuracy (Acc) and weighted F1 score as evaluation metrics, and the calculation method follows the official evaluation protocol of each dataset. The model supports binary, quadruple, and seven-class classification for multi-category determination: binary classification divides into two categories of emotional labels: negative and positive; quadruple classification includes four basic emotional labels: joy, sadness, anger, and calmness; and seven-class classification covers seven refined emotional labels: neutral, joy, sadness, anger, surprise, fear, and disgust. After the model is loaded, it performs inference and prediction on the test samples. The terminal console can print out key information such as the predicted class label, confidence probability of each class, and model prediction score for single or batch samples in a structured manner, intuitively presenting the model's class determination results and confidence level for the samples, as shown below: ========== Sample Prediction Result ========== Input Modalities: Text(L) + Audio(A) + Video(V) Predicted Class: Positive Confidence Probability: 0.9427 Model Prediction Score: 0.8765 ============================================= The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the present invention. Although detailed descriptions have been made with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments, and they should all be covered within the protection scope of the claims.
Claims
1. A classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization, characterized in that: Addressing the emotional perception scenarios of teachers in classroom teaching, this research adopts a hybrid attention and hypergraph temporal model adapted to the temporal characteristics of the classroom as its core architecture, and multi-scale proxy optimization as an efficiency improvement strategy. It constructs a dual-branch system of graph attention and hypergraph diffusion to capture fine-grained pairwise emotional dependencies of individual teachers within the text, audio, and video modalities of classroom teaching, as well as high-order emotional group associations between teachers across multiple modalities. It leverages a data-driven mechanism to learn implicit temporal structures, overcoming the limitations of predefined time segments. Furthermore, it employs a multi-scale proxy network to complete multi-view representation selection and adaptive hyperparameter optimization, reducing model tuning overhead. First, a hybrid attention and hypergraph temporal model is constructed, dividing the modeling and training process of multimodal teaching emotions of classroom teachers into two stages: the multi-view representation extraction stage and the pairwise and higher-order relation collaborative training stage. In the multi-view representation extraction stage, multi-view feature mapping of text, audio, and video multimodal temporal data is performed in the time domain, frequency domain, structural correlation, and temporal state. Multi-view adaptive fusion is completed through a gated residual mechanism to obtain a unified multi-granular representation. After entering the collaborative training phase, the graph attention GAT branch and the hypergraph diffusion HGNN branch are constructed in parallel. The GAT branch is responsible for modeling fine-grained pairwise dependencies between multimodal nodes and maintaining local semantic sensitivity discrimination ability. The HGNN branch achieves smooth aggregation of cross-modal high-order group relations through learnable soft membership and provides global sentiment constraints. Learnable mixing coefficients are used during training to dynamically adjust the contribution weights of GAT and HGNN based on the temporal characteristics of the input, replacing the fixed fusion strategy. Meanwhile, before model training, a multi-scale agent network is started to replace a large amount of real training with lightweight agent evaluation, and joint pre-screening of multi-view combination and hyperparameters is performed to reduce search overhead. The training process employs a joint loss function that includes hyperedge orthogonal regularization and membership entropy regularization to ensure a reasonable hypergraph structure and stable information propagation. Ultimately, the network achieves both local sensitivity capture and global smooth induction within a unified framework, improving the recognition accuracy and training efficiency of non-uniform classroom teaching sequences and teacher multimodal sentiment analysis. It is suitable for offline / online classroom text, audio, and video joint sentiment recognition scenarios.
2. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 1, characterized in that: The core architecture uses a hybrid attention and hypergraph temporal model. The specific construction process of the hybrid attention and hypergraph temporal model is as follows: First, a multimodal input layer is constructed to receive aligned teacher classroom teaching text (such as teaching language and classroom instructions), audio (such as teaching tone, speaking speed, and intonation), and video (such as facial expressions, body movements, and gestures) three-modal temporal features, and then enters the multi-view representation extraction process; The first stage performs multi-view temporal feature mapping, generating five complementary features: time domain view TD, frequency domain view FD, Gram angle and field GASF, Gram angle difference field GADF, and Markov transfer field MTF. Through a gated residual fusion mechanism, the effective information of each view is adaptively aggregated with the TD view as the backbone to complete the unified representation of multi-granularity and multi-modal features. The second stage establishes a dual-branch relationship modeling backbone, and implements pairwise dependency modeling and high-order group modeling in parallel. The graph attention GAT branch generates query Q, key K, and value V through linear projection, and uses multi-head attention to calculate fine-grained pairwise attention weights between nodes. It focuses on capturing the emotionally sensitive information of binary temporal dependencies between different modalities within the same modality of teachers, and realizes the accurate capture of subtle emotional changes in teachers' teaching process. The hypergraph diffusion HGNN branch calculates the soft membership degree of nodes through learnable hyperedge prototypes, and constructs dynamic soft hypergraph association matrix and diffusion matrix to achieve smooth aggregation of cross-modal multi-node high-order group relationships. The third stage introduces an adaptive hybrid fusion module, which sets learnable single parameters and obtains hybrid coefficients through Sigmoid mapping. Based on the input temporal characteristics, it dynamically weights and fuses GAT pairwise features and HGNN high-order features, and combines residual connections and layer normalization to stabilize training. The fourth stage involves cross-modal temporal interaction, using any single emotional modality feature of the teacher as the query and other modalities as key-value pairs to complete the joint guidance enhancement of cross-modal pairwise relationships and higher-order relationships, resulting in multimodal deep integration of the teacher's teaching emotional temporal features. The fifth stage incorporates hypergraph structure regularization constraints, setting hyperedge prototype orthogonal regularization terms and soft membership entropy regularization terms, which are jointly optimized with the sentiment classification task loss to ensure that the hypergraph structure does not degenerate and semantics do not overlap. The sixth stage integrates the multi-scale proxy optimization pre-module MS²-Net, which uses 1 / 3 / 5 multi-scale parallel branches to lightweight evaluate view combinations and hyperparameters such as learning rate, Dropout, number of hyperedges, and number of attention heads before formal training, completing optimal configuration pre-screening to reduce search overhead. Finally, the fused features are fed into the multilayer perceptron (MLP) classification head, outputting sentiment category prediction results and completing loss backpropagation, realizing end-to-end modeling of multimodal sentiment analysis of individual teacher classroom teaching.
3. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 2, characterized in that: Multi-view mapping and adaptive fusion method; In the analysis of teachers' emotional states in classroom teaching, the temporal characteristics on which different emotional states of teachers depend differ: the emotional intensity of teachers' tone of voice during explanation is reflected in the temporal envelope, the emotional tendency of text semantics is implied in the long-range context, and cues such as laughter and sighs have cross-modal periodicity; a single temporal view is difficult to capture these heterogeneous characteristics at the same time, and the introduction of a full view is prone to bringing redundant information that is not related to emotion. To this end, a complementary view is constructed from five perspectives: time domain, frequency domain, structural correlation, temporal difference, and state transition, to provide multi-granular representations of the same sequence that meet the needs of sentiment discrimination. Given the original multimodal input sequence The view extractors are defined as follows: The temporal view (TD) preserves the original temporal structure through linear projection, capturing fine-grained emotional cues (such as intonation shifts and speech rate changes); the frequency view (FD) extracts global periodic features through Fourier transform, revealing the fluctuation patterns of emotional intensity and cross-modal resonance patterns; different emotional states are not only reflected in the amplitude of temporal numerical changes, but also implicit in the structural patterns and state transition rules of the temporal sequence—for example, "happy → surprised" and "calm → sad" have fundamental differences in their state transition paths, which are difficult to distinguish using simple temporal amplitudes; therefore, structural correlation and temporal state views are introduced to characterize them; Gram angle and field... The Gaussian External Surface (GASF) and Gram Aspect Difference Field (GADF) map time series to polar coordinate space and capture the structural correlation between time series points by calculating the inner product of angle differences: GASF characterizes the static similarity of emotional semantic segments in polar coordinates, while GADF captures the dynamic turning direction of emotional polarity; Markov Transition Field (MTF) models the dynamic laws of time series from the perspective of state transition, transforming continuous time series into state sequences through quantile discretization, and then characterizing the evolution path of emotional states—such as the gradual or abrupt process from calm to excitement, from sadness to relief, etc.—by using the state transition probability matrix, which makes up for the shortcomings of the first two types of views in modeling the dynamics of time series states; The above five views provide complementary representations of the same time series from the four perspectives of amplitude, frequency, structure and state, and together constitute a candidate view set; Teachers' dependence on different views varies across different classroom teaching scenarios: rhythmic teaching emotions (such as excited guidance during class introduction and enthusiastic emphasis during explanation of key points and difficulties) strongly depend on the periodic patterns in the time-frequency domain, while semantically driven teaching emotions (such as tactful correction of students' mistakes and sincere encouragement of excellent performance) require the semantic structural associations captured by GASF / GADF. Fixed-weight fusion strategies cannot adapt to these scenario differences. Therefore, based on the temporal view TD, a gating mechanism is used to adaptively adjust the information inflow of each derived view. It should be noted that which of the above five views is ultimately enabled is not manually preset, but rather determined by the MS²-Net agent network based on validation performance before training starts, in order to retain complementary information while eliminating redundant views that are irrelevant to emotions. The optimal view combination was determined through a proxy model. Then, each enabled view extractor outputs features. The data needs to be integrated into a unified multi-view representation; therefore, a gated residual fusion mechanism is designed to utilize temporal view features. Using the main framework, the remaining views are adaptively aggregated using the gating weights it guides: ； After splitting the gated outputs by view, they are superimposed as residuals: ； in, The gating weight is the value corresponding to the k-th view. It represents element-wise product; the mechanism is dominated by the temporal view for gating decisions, and adaptively adjusts the contribution ratio of each derived view according to the sentiment characteristics of the current input, while retaining the original temporal discriminative power and flexibly absorbing the sentiment semantics of complementary views.
4. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 1, characterized in that: Pairwise temporal dependency modeling based on GAT; In classroom teaching sentiment analysis tasks, the core of identifying individual teachers' teaching sentiments is capturing the temporal correlations of teachers' multimodal sentiment features. The semantic modification relationships between words within a modality and the fine-grained alignment relationships between modalities are essentially pairwise temporal dependencies between nodes. Graph Attention Networks (GAT) assign differentiated weight coefficients to different nodes through an attention mechanism, which can effectively capture such pairwise dependencies and is a core means of characterizing local temporal interactions. To meet the temporal modeling needs of multimodal and multi-view features, the traditional GAT is adapted to handle the feature representations of multimodal temporal nodes and retain the attention calculation paths of query, key, and value, ensuring stable gradient propagation and effective scoring of pairwise relationships. Let the set of multimodal temporal features after time-frequency fusion and view fusion be . Temporal characteristics of any mode First, feature mapping of Query, Key, and Value is completed through linear projection: ； in, The projection matrix is learnable; Features of the original multimodal input sequence; To capture heterogeneous temporal relationships from multiple subspaces in parallel, a multi-head attention mechanism is employed to characterize pairwise dependencies between nodes from multiple perspectives. Both the projection matrix and features are divided into H attention heads, i.e. Each attention head independently calculates the attention score between nodes; the pairwise attention score for node i to node j in the h-th attention head is calculated as follows: ； in, Let h be the learnable weight vector of the h-th attention head; Indicates feature concatenation operation; A scaling factor to prevent gradient explosion, used to stabilize gradients; LeakyReLU is a non-linear activation function; Through attention masking Filter invalid nodes (if nodes i and j have no valid connection) ,otherwise The attention scores are normalized using the Softmax function to obtain the pairwise attention weights between nodes. ； Based on the normalized attention weights, the Value features are weighted and aggregated to obtain the pairwise relation feature representation of a single attention head. Then, the results of H attention heads are concatenated to complete the fusion of multi-head attention, ultimately obtaining the pairwise temporal dependency features based on GAT. : ； ； in, The output projection matrix of multi-head attention is used to realize the restoration and fusion of feature dimensions; In end-to-end training, the model autonomously learns the pairwise connection strength between teachers' emotional nodes, accurately capturing local semantic emotional transitions in teachers' teaching emotional sequences (such as the change in teachers' emotions from patience to seriousness when asking questions in class) and intermodal emotional alignment relationships (such as the emotional synergy between an enthusiastic tone in speech and facial expressions such as smiling and raising one's head). This provides a refined foundation of local emotional features for subsequent modeling of higher-order relationships in teachers' teaching emotions, further improving the accuracy of individual teachers' teaching emotion analysis.
5. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 1, characterized in that: Hybrid relationship adaptive fusion mechanism; For the identification of teaching emotions of individual teachers, the fine-grained cues and high-order patterns that the multi-scale attention module needs to capture are consistent with the characteristics of teachers' emotional expression in the teaching process: fine-grained cues such as local semantic transitions and instantaneous emotional alignment rely on pairwise relationships for capture, while high-order patterns such as overall emotional tone and cross-modal collaboration require group-level modeling. The contributions of the two vary depending on the individual teacher's teaching style, teaching process, and intensity of emotional expression. The local pairwise relationships characterized by GAT and the global high-order relationships characterized by HGNN are complementary, and their collaboration can more completely characterize the complex dependency features of multimodal time-series data. To address this, a learnable adaptive fusion mechanism is designed, using the mixing coefficients... This enables dynamic weighted fusion of pairwise relation features and higher-order relation features, adapting to the expressive pattern of teachers' teaching emotions as "local instantaneous changes and global unified tone"; Mixing coefficient Parameters can be learned through a single variable. Obtained through mapping using the Sigmoid function, the value range is: ,Right now: ； The design allows The initial values can be biased towards pairwise relationship modeling of GAT, preserving its ability to characterize local pairwise interactions, and then adaptively adjusted according to data characteristics during end-to-end training. Based on mixing coefficient Paired features of GAT With HGNN higher-order features Perform element-wise weighted fusion to obtain the fusion time-dependent features. : ； To further enhance the expressive power of fused features, residual connections and layer normalization operations are introduced after weighted fusion to alleviate gradient decay in deep networks and improve training stability. ； Where X represents the original multi-view fusion feature; Dropout is a random deactivation operation used to prevent model overfitting; LayerNorm is a layer normalization operation that standardizes the distribution of features. The adaptive fusion mechanism enables the model to flexibly adjust the contribution ratio of the two types of relationships according to different scenarios of individual teachers' teaching emotions, adapting to the actual needs of classroom teaching emotion analysis; for example, when there are significant local semantic changes in the emotion sequence (such as irony or transition), the model increases... To enhance GAT's pairwise modeling capabilities; when the overall emotional atmosphere is relatively consistent (such as a consistently warm narrative passage), the model reduces... To highlight the higher-order diffusion effect of HGNN; this design, driven by teachers' teaching emotion data, replaces the fixed weight allocation of pairwise and higher-order relations, effectively solving the identification problem caused by differences in the emotional expression of different teachers and the diverse emotional changes in different teaching stages, and improving the accuracy and adaptability of individual teachers' classroom teaching emotion analysis.
6. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 1, characterized in that: Implicit temporal structure learning mechanism; Teachers' emotional expression in teaching is multi-scale and non-uniform in rhythm—the pace of their speech fluctuates, and emotions may gradually accumulate or suddenly erupt due to classroom feedback or unexpected teaching situations. Traditional predefined time segmentation strategies are difficult to adapt to this dynamic and are prone to introducing prior bias. Based on the aforementioned attention mechanism, implicit temporal structure learning can be achieved through a combination of two paths without explicit time segmentation: (1) Topological guidance of attention propagation; The pairwise attention weights αij of GAT characterize the dynamic connection strength between nodes; During training, the weights are clustered towards semantically related nodes (such as adjacent emotional keywords, cross-modal aligned node pairs), forming an adaptive local temporal topology, which gets rid of the dependence on predefined time windows. (2) Semantic aggregation of soft grouping of hypergraph; Soft membership Hie assigns nodes to different hyperedges in the form of probability. The hyperedge prototype is continuously optimized during training so that nodes under the same hyperedge have consistent semantic features (such as nodes with the same emotional state), forming adaptive global semantic grouping. The end-to-end joint optimization of attention weights and hypergraph membership calculations with the teaching sentiment prediction task deeply couples the temporal structure learning with the ultimate goal of teacher teaching sentiment analysis. As a result, the model output feature Xout simultaneously encodes the fine-grained multimodal pairwise interactions, higher-order group associations, and the implicit temporal structure of teachers' teaching sentiments during the teaching process, thereby improving the accuracy of individual teacher teaching sentiment analysis.
7. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 1, characterized in that: Interaction modeling of temporal dependencies in multimodal heterogeneous views; The aforementioned pairwise-higher-order relation modeling mechanism is designed to characterize the temporal dependencies of a single modality; however, the core of multimodal sentiment analysis lies in the information complementarity and synergy between different modalities—text provides semantic content, audio conveys tone and emotion, and video captures facial expressions and postures. Only through mutual guidance can the emotional expression be fully understood. To this end, based on a cross-modal attention mechanism, the temporal dependency modeling of a single modality is extended to multimodal and multi-view scenarios, enabling cross-modal interaction of pairwise and higher-order relations between language, audio, and video modalities. For any two modes and modality Multi-view fusion features as queries, modal The multi-view fusion features are used as Key and Value, and the modality is calculated through the attention mechanism of the GAT-HGNN fusion described above. For modes Cross-modal pairwise attention and higher-order diffusion features enable temporally dependent interactions between modalities: ； in, Representing modes For modes Cross-modal temporal dependency features; For each modality, the cross-modal interaction features of the other two modalities are fused to obtain the temporal dependency features of multimodal multi-view fusion: ； in, These are the two modes other than m; It is a learnable fusion matrix that enables dimensionality reduction and information aggregation of cross-modal features; The core of multimodal and multi-view temporal dependency interaction modeling is to enable the temporal dependency features of different modalities in the classroom scenario to guide and enhance each other, accurately capture the fine-grained pairwise alignment relationships and global high-order collaborative patterns between modalities related to the teacher's teaching emotions. For example, the semantic information of language features guides the extraction of emotional features from audio and video features, and the temporal dynamic features of audio and video features assist the semantic understanding of language features, ultimately achieving deep integration of multimodal temporal dependencies. Through the aforementioned cross-modal interactions, the temporal dependence features of different modalities mutually guide and reinforce each other around the teacher's teaching emotions: the semantic information of language features provides contextual guidance for the emotional cues in audio and video, while the temporal dynamic features of audio and video assist language in dissolving ambiguity (such as a sudden change in the teacher's tone of voice revealing the emotional intention of "serious reminder" rather than ordinary statement); this cross-modal collaboration enables the model to capture both fine-grained alignment relationships between modalities and to uncover high-order cross-modal collaboration patterns at the global level, thereby accurately depicting the changes in the teacher's emotional state throughout the entire classroom teaching process.
8. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 2, characterized in that: Design of a joint loss function with hypergraph structure regularization; The aforementioned pairwise-higher-order collaborative modeling and implicit structure learning constitute a complete framework for temporal dependency modeling. For the task of analyzing sentiment in individual classroom teaching, due to the special nature of the dataset annotation format in the teaching scenario, the model training process needs to ensure the effectiveness of cross-modal feature fusion and the rationality of hypergraph structure learning. To this end, cross-entropy loss is adopted and a hypergraph structure regularization term is introduced to constrain the orthogonality of hyperedge prototypes and the rationality of the hypergraph association matrix, so as to achieve joint optimization of task objectives and structure learning. Hypergraph structure regularization loss: To ensure the effectiveness of hypergraph structure learning and avoid hyperedge prototype degradation and overly uniform distribution of the hypergraph association matrix, a hyperedge prototype orthogonality regularization term L_orth (constraining the orthogonality of different hyperedge prototypes) and a hypergraph association matrix entropy regularization term L_entropy (constraining the discriminability of soft membership distribution) are designed to constitute the hypergraph structure regularization loss. The formula is: ； in, , This is the regularization coefficient, used to balance the contribution of structural constraints and task losses; Hyperedge Prototype Orthogonal Regularization Term To constrain the orthogonality between different hyperedge prototypes, avoid semantic overlap of hyperedges, and ensure that each hyperedge corresponds to a unique high-order group feature, the formula is as follows: ； Where E is the number of superedges; Let be the prototype vector of the e-th hyperedge. The orthogonality constraint maximizes the expressive power of the hyperedge prototype space. Hypergraph Incidence Matrix Entropy Regularization Term : The entropy value of the soft membership degree distribution of a node to its hyperedge is constrained to avoid the node being evenly distributed across all hyperedges (resulting in excessively large membership degree entropy values), ensuring that the membership degree of a node to its hyperedges is discriminative. The formula is: ； in, The elements of the soft hypergraph association matrix represent the probability that the i-th node belongs to the e-th hyperedge. By minimizing this regularization term, nodes are guided to cluster towards semantically matching hyperedges, thus enhancing the discriminative power of membership. The overall loss function of the model is a weighted sum of the task loss and the hypergraph structure regularization loss, and the formula is: ； in, The loss function is used for classification tasks; it jointly optimizes the sentiment prediction objective and the hypergraph structure learning objective, ensuring classification accuracy while constraining the rationality of the hypergraph structure.
9. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 1, characterized in that: Multi-scale agent network design; In HAHT tuning, discrete choices such as view combination and the number of hyperedges determine the model's topology, while continuous hyperparameters such as learning rate and Dropout control the training dynamics on this structure. These two are not independent: increasing the number of hyperedges enriches the relationships between higher-order groups but also amplifies gradient propagation noise, requiring a smaller learning rate. View subsets determine the dimensionality and signal-to-noise ratio of input features, directly affecting the optimal range of Dropout values. Agent networks with a single receptive field struggle to capture these cross-granular couplings simultaneously—narrow receptive fields emphasize local dependencies (such as the synergy between learning rate and Dropout), while wide receptive fields emphasize global structure (such as the impact of view combination on overall performance), but in HAHT, the two are mutually constrained. To address this, a multi-scale parallel branching proxy network is designed, extracting fine-grained local dependencies, medium-grained associations, and coarse-grained global structures through fully connected branches at three scales: scale 1 focuses on univariate sensitivity (e.g., the impact of a single change in the learning rate on performance), scale 3 captures pairwise couplings (e.g., the joint effect of the learning rate and Dropout), and scale 5 models higher-order interactions (e.g., the synergistic constraints of view composition, the number of hyperedges, and the learning rate). These three aspects are then fused across scales through feature concatenation, enabling the proxy network to simultaneously complete local sensitivity evaluation and global structure perception in a single forward propagation, explicitly modeling the cross-granularity coupling relationships between continuous hyperparticipants and discrete choices in HAHT. The number of neurons in each branch scales dynamically with the dimension D of the decision variable. , , No manual structural adjustments are needed for different problems; each branch consists of a fully connected layer, batch normalization, and ReLU activation, which are then spliced together and output as predicted values via a fusion head; this multi-scale parallel design enables the network to adapt to nonlinear mappings in high-dimensional hybrid search spaces with a relatively low computational cost. The above-mentioned proxy network is combined with evolutionary optimization to form a multi-scale proxy-assisted evolutionary optimization framework. In each iteration: a reference solution is selected based on the screening strategy, the population is stratified and sampled according to the dominance relationship and then the proxy network is trained. A prediction error-driven adaptive strategy is used to screen high-value candidate solutions and perform real evaluation. The agent network is built once and reused throughout the entire iteration process, avoiding the extra overhead of repeated modeling in each generation of traditional methods; This framework improves optimization efficiency and performance through multi-stage collaborative design. The stratified sampling data partitioning method maintains a balanced class distribution between the training and test sets, preventing model bias towards the majority class due to sample imbalance and further ensuring the prediction accuracy of the surrogate model. Simultaneously, the prediction error-driven adaptive sample selection strategy dynamically adjusts sampling rules based on the model's prediction errors on dominant and non-dominant samples, selecting high-value candidate solutions in reliable regions and performing random exploration in unreliable regions, achieving a dynamic balance between global exploration and local development. Furthermore, the framework adopts a multi-scale network model that is built once and reused throughout the iteration process, avoiding the high computational overhead of repeated modeling in each generation of traditional surrogate models and effectively improving overall optimization efficiency.
10. The classroom teaching sentiment analysis method based on graph-hypergraph collaboration and low-cost adaptive agent optimization according to claim 9, characterized in that: Multi-view selection and hyperparameter search based on proxy model; HAHT includes a dual-branch structure of graph attention and hypergraph diffusion, and the time consumption of a single complete training is already higher than that of a lightweight model. At the same time, there are multiple ways to combine the five candidate views, and after crossing with continuous and discrete hyperparameters such as learning rate, Dropout, and number of hyperedges, the search space grows exponentially. If we rely on grid search or manual parameter tuning to train each configuration one by one, the computational cost will be unbearable. Multiple view selection; different emotional scenarios have different dependencies on views: in monologues, the time-frequency patterns of acoustic features are relatively stable, and frequency views can often provide effective supplementation; however, in film and television dialogues with frequent scene changes, changes in temporal structure are more critical, and structural views such as GASF / GADF may contribute more; this difference is difficult to exhaustively enumerate through human experience, and using all views in a fixed way will introduce redundant noise; therefore, view selection is modeled as a discrete binary search space {0, 1}. K (K is the number of candidate views), the proxy model quickly evaluates each candidate combination on a lightweight task, and adaptively selects the optimal subset that is sensitive to the current task. The real model is fed in for training, and redundant views are removed while retaining key information. The view selection problem can be modeled as a discrete binary search space; let the total number of candidate views be K, and the view selection vector be defined as: ； in, This indicates that the k-th candidate view is enabled. This indicates that the proxy model is not enabled; the proxy model quickly evaluates each candidate combination on a lightweight proxy task and selects the combination with the best validation performance. This final configuration is then fed into the network for formal training. The proxy model determines whether to use a particular view. It rapidly evaluates candidate combinations on lightweight proxy tasks and selects the combination with the best validation performance. As the final configuration, it is fed into the network for formal training, thereby reducing computational redundancy while maintaining the model's expressive power; Real model interaction and hyperparameter optimization; the collaborative mechanism between the surrogate model and the real model, through unified encoding of decision variables, parameter constraint mapping, cross-platform invocation and result feedback, aligns the surrogate space with the real target space; let the objective function of the real model be: ； The decision variables include continuous hyperparameters (such as learning rate, various Dropout methods, and initial values of fusion coefficients), discrete hyperparameters (such as batch size, number of network layers, projection dimension, number of attention heads, and number of hyperedges), and view combination indexes. The optimization objective is to simultaneously maximize Acc and F1 in this search space and minimize training time. The surrogate model aims to learn an approximate mapping. To reduce the number of evaluations of the true function; The PlatEMO optimization platform is used for scheduling and execution: the MATLAB side is responsible for parameter normalization mapping and population evolution, and the Python deep learning environment is called through the command line to execute training. The output file is parsed to obtain three indicators: Acc, F1 and training time, which are then sent back to the surrogate model. Infeasible configurations are avoided through parameter boundary constraints and running status verification, which provides support for the iterative evaluation of surrogate-assisted evolutionary optimization. The surrogate model shares the same decision space, constraints, and optimization objectives with the real model. Surrogate evaluation replaces a large amount of real training, reducing the computational overhead of hyperparameter search and multi-view selection while maintaining HAHT modeling capabilities. This provides a cross-platform collaborative solution for efficient and automated tuning of complex multimodal time series modeling.