Agent high-order relationship modeling method based on edge attention weight

Abstract

This invention relates to the field of high-order relation modeling technology for intelligent agents, and particularly to a method for high-order relation modeling of intelligent agents based on edge attention weights. The method includes the following steps: acquiring multimodal interaction data of the intelligent agent, constructing an initial relation graph of the intelligent agent, and extracting edge attribute features; assigning initial attention weights based on the edge attribute features, generating an edge weight-enhanced graph, and performing multi-hop high-order relation propagation to obtain node representation enhancement data; constructing a task decision path based on the enhanced node representations and propagation information, and generating intelligent agent reflection fusion data through a three-level cognitive reflection mechanism; constructing an edge state reinforcement learning model based on the reflection fusion data, and iteratively optimizing the high-order propagation path; this invention improves the structural cognitive ability and policy adaptability of intelligent agents to complex tasks through high-order relation modeling.

Description

Technical Field

[0001] This invention relates to the field of high-order agent relation modeling technology based on edge attention weights, and particularly to a method for high-order agent relation modeling based on edge attention weights. Background Technology

[0002] Intelligent agents play an increasingly important role in multimodal perception, complex task execution, and collaborative decision-making. Especially in tasks such as social network understanding, intelligent recommendation systems, and cognitive reasoning systems, higher demands are placed on the accuracy and depth of modeling relationships between agents. Traditional graph neural network (GNN) methods largely rely on modeling strategies based on node ontology attributes, neglecting the dynamics and multi-layered semantic structures of interactions between agents. This results in a lack of interpretability and expressiveness when handling high-order relationships (such as multi-hop dependencies, structural coupling, and behavioral collaboration). Furthermore, existing technologies generally employ static graph structures and average aggregation strategies, making it difficult to effectively model the semantic features and interaction weights of edges in heterogeneous multimodal contexts, limiting the controllability and adaptability of information flow between agents. Some studies have introduced attention mechanisms to optimize the fusion of neighbor node features, but these often focus on the nodes rather than the edges themselves, ignoring the crucial attributes of edges as relationship carriers, such as temporal sequence, frequency, and contextual structure, thus failing to achieve deep modeling and dynamic adjustment at the semantic level of edges. However, traditional high-order relation modeling of agents with edge attention weights suffers from inaccurate construction of agent task decision paths and inaccurate processing of agent node representation enhancement. Summary of the Invention

[0003] Therefore, it is necessary to provide a method for modeling high-order relationships of agents based on edge attention weights to solve at least one of the above-mentioned technical problems.

[0004] To achieve the above objectives, a method for modeling high-order agent relationships based on edge attention weights includes the following steps:

[0005] Step S1: Obtain multimodal interaction data of the agent; construct an initial relational graph of the agent based on the multimodal interaction data of the agent; determine the edge attribute features of the agent based on the initial relational graph of the agent;

[0006] Step S2: Perform initial attention weight allocation based on the agent's edge attribute features to obtain initial edge attention weight data; construct an edge weight enhancement graph based on the edge attention weight data; perform multi-hop higher-order relation propagation based on the edge weight enhancement graph data to obtain multi-hop higher-order relation propagation data; perform agent node representation enhancement based on the multi-hop higher-order relation propagation data to obtain agent node representation enhancement data.

[0007] Step S3: Construct a task decision path based on agent node representation enhancement data and multi-hop high-order relationship propagation data to obtain the agent task decision path; perform a three-level cognitive reflection operation on the agent task decision path to obtain three-level cognitive reflection data; and fuse the three-level cognitive reflection data to obtain agent reflection fusion data.

[0008] Step S4: Construct a side-state reinforcement learning model based on the agent's reflection and fusion data; perform iterative optimization of the higher-order propagation path based on the side-state reinforcement learning model to obtain the higher-order propagation path iterative optimization data.

[0009] This invention achieves high-fidelity mapping of multimodal information to structural graphs by constructing an initial relational graph of intelligent agents and extracting edge attribute features, effectively improving the accuracy of relational modeling. It employs edge semantic vectors and a context fusion mechanism to allocate attention weights, making the edge weights more closely resemble real interaction intensity and enhancing the dynamic plasticity of the structural graph. Based on this, multi-hop higher-order relation propagation is performed, taking into account both semantic connections and behavioral similarities between remote nodes, achieving cross-level semantic connectivity. Node representation enhancement operations improve the node's ability to receive upstream signals and the accuracy of downstream path response, optimizing information transmission. The system employs a chain-like approach; task decision-making path construction combines high-order structure and behavioral characteristics to ensure the accuracy and efficiency of path selection; a three-level cognitive reflection mechanism performs backtracking analysis of the task path from multiple levels—results, processes, and strategies—enhancing the system's adaptability and task stability; the fused reflection data serves as the foundation for reinforcement learning modeling, enabling edge state perception to have real-time adjustment and optimization capabilities, achieving a dynamic self-feedback closed loop in path evolution; iterative optimization of high-order propagation paths gives the entire system high robustness and interpretability in complex structures, significantly improving the accuracy, stability, and generalization ability of agent high-order relationship modeling. Therefore, this invention optimizes traditional agent high-order relationship modeling based on edge attention weights, solving the problems of inaccurate agent task decision-making path construction and inaccurate agent node representation enhancement processing in traditional edge attention weight-based agent high-order relationship modeling, thus improving the accuracy of agent task decision-making path construction and agent node representation enhancement processing. Attached Figure Description

[0010] Figure 1 A flowchart illustrating the steps of a high-order agent relationship modeling method based on edge attention weights;

[0011] Figure 2 for Figure 1 A detailed flowchart illustrating the implementation steps of step S3.

[0012] Figure 3 This is a schematic diagram illustrating the construction of the initial relationship graph for the intelligent agent in this invention;

[0013] The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0014] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0015] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0016] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0017] To achieve the above objectives, please refer to Figures 1 to 3 A method for modeling high-order relationships between agents based on edge attention weights includes the following steps:

[0018] Step S1: Obtain multimodal interaction data of the agent; construct an initial relational graph of the agent based on the multimodal interaction data of the agent; determine the edge attribute features of the agent based on the initial relational graph of the agent;

[0019] In this embodiment of the invention, five types of heterogeneous modal data, including voice commands, gesture trajectories, environmental images, location information, and behavior logs, are extracted from a pre-defined intelligent agent interaction database to form the original multimodal interaction dataset of the intelligent agent. This dataset undergoes unified timestamp alignment, signal-noise filtering, image resolution resampling, and speech-to-text processing by a data preprocessing module to obtain a structured multimodal interaction data table. Subsequently, a graph structure generation engine is used to divide the above structured data into nodes and edges according to the interaction object and interaction type, constructing an initial relational graph for the intelligent agent. Nodes include three types of entities: the intelligent agent itself, external environment objects, and interaction targets. Edge types include three types of interaction types: observation, command, and feedback. Each edge records the specific interaction content, time label, and directional information. Next, the attribute vector of each edge is calculated using the metadata of the edges in the structured graph (such as interaction frequency, interaction interval, and data modality source), extracting edge attribute feature data. This edge attribute feature data includes fields such as interaction density, interaction latency, and semantic redundancy, and serves as the basis for subsequent edge attention initialization.

[0020] Step S2: Perform initial attention weight allocation based on the agent's edge attribute features to obtain initial edge attention weight data; construct an edge weight enhancement graph based on the edge attention weight data; perform multi-hop higher-order relation propagation based on the edge weight enhancement graph data to obtain multi-hop higher-order relation propagation data; perform agent node representation enhancement based on the multi-hop higher-order relation propagation data to obtain agent node representation enhancement data.

[0021] In this embodiment of the invention, based on the edge attribute feature data obtained in step S1, an edge semantic encoder is used to perform embedding vector encoding processing on each edge. This encoder uses independent attribute embedding channels to perform numerical normalization, structural embedding, and feature fusion on interaction density, latency, and semantic redundancy respectively, obtaining edge semantic encoded data. Based on this, the historical interaction frequency intensity of each edge is statistically analyzed, and interaction temporal feature information is extracted to generate a time-series vector representation. Further, the semantic vector and the frequency-temporal joint vector are fused to form a unified edge semantic vector representation data. Combined with graph structure information, an aggregation operator is used to extract the adjacency structure distribution pattern of each node, forming edge adjacency structure feature data. Then, the semantic vector representation and adjacency structure features are fused and modeled in context to generate edge relationship context representation data. A weighted normalization mechanism combined with an attention scoring function is used to assign initial attention weights to all edges, obtaining initial edge attention weight data. This attention allocation result is used to perform edge weight enhancement processing on the graph, generating an edge weight enhanced graph. Subsequently, multi-order adjacency edge data of all nodes are extracted from the enhanced graph, a jump distance marking algorithm is executed, and the edge structure corresponding to each jump distance is labeled. Next, a candidate set of multi-hop paths is extracted based on the hop distance structure, and the behavioral similarity factor between nodes on the path is calculated. Multi-hop propagation aggregation is performed based on the similarity factor, merging the path propagation results into a unified representation to construct multi-hop propagation aggregation data. Then, by coupling the upstream and downstream edge weights, hop distance similarity, and path intersection degree between propagation paths, cross-node relationship coupling data is calculated. Based on this data, higher-order semantic extension data is calculated, and combined with the multi-hop higher-order relationship propagation processing completed by the propagation aggregation data, multi-hop higher-order relationship propagation data is obtained. Subsequently, the upstream and downstream interaction path information of nodes in this data is extracted, the interaction frequency distribution of each node is statistically analyzed, and high-frequency path reduction processing is performed based on the frequency distribution to obtain high-frequency path structure reduction data. The semantic and structural features of nodes are fused by combining the reduced data and the upstream and downstream path information to obtain node feature fusion data. This fusion data is used for node stability assessment calculation, forming node state stability data, and is used to identify high-fluctuation abnormal node feature dimensions, generating node abnormal feature data. Based on this, semantic correction is performed to obtain higher-order semantic correction node data. This data is then fused with the above data to update the original node representation, resulting in enhanced agent node representation data.

[0022] Step S3: Construct a task decision path based on agent node representation enhancement data and multi-hop high-order relationship propagation data to obtain the agent task decision path; perform a three-level cognitive reflection operation on the agent task decision path to obtain three-level cognitive reflection data; and fuse the three-level cognitive reflection data to obtain agent reflection fusion data.

[0023] In this embodiment of the invention, node representation enhancement data and multi-hop high-order relationship propagation data are used for path node screening and boundary identification processing to extract task-aware features of each node. The connectivity between paths with different hop distances is calculated using path node boundary information, and a graph traversal algorithm is used to determine the set of connected paths, constructing path connectivity data. Subsequently, a hop-distance priority and edge weight sorting algorithm is used for task path optimization processing, generating task path optimization sorting data. The sorting results are mapped to a node-edge sequence, outputting the agent's task decision path. After path generation, the path node load statistics function is called to obtain the number of processing requests, average response time, and volatility index for each node. Based on these indicators, a node response stability vector is constructed, and then the collaborative efficiency vector in task execution is calculated as the basis for path feasibility assessment, outputting agent path feasibility data. Combining path structure, task labels, and node performance, a three-level cognitive reflection module is designed to construct three types of data: result reflection (task completion rate and resource utilization rate), process reflection (path hop count and execution time), and strategy reflection (node ​​selection and path planning rules), generating three-level cognitive reflection data. By using tensor fusion, the three types of reflection datasets are transformed into a unified structure of agent reflection fusion data, providing data input for training reinforcement learning models.

[0024] Step S4: Construct a side-state reinforcement learning model based on the agent's reflection and fusion data; perform iterative optimization of the higher-order propagation path based on the side-state reinforcement learning model to obtain the higher-order propagation path iterative optimization data.

[0025] In this embodiment of the invention, cognitive distillation is performed using agent reflection fusion data as input. This process employs a reflective semantic compression module to compress high-dimensional, multi-source cognitive data into lightweight knowledge units, outputting these lightweight knowledge unit data. This data is then used to construct an edge-state reinforcement learning model. In this model, edges serve as action units, edge state features serve as input state vectors, and reflection fusion weights serve as the basis for constructing reward signals. Iterative training is conducted using a Q-learning strategy. The reinforcement learning process employs a joint mechanism of policy iteration and edge state update. After each training round, a higher-order propagation path simulation is performed based on the current policy, and its path cost function and reflection fit are evaluated. The optimized higher-order propagation path iteration data is then output and fed back to the graph construction system to complete the dynamic update loop of the path structure.

[0026] Preferably, step S1 includes the following steps:

[0027] Step S11: Acquire multimodal interaction data of the intelligent agent;

[0028] In this embodiment of the invention, during the execution of collaborative tasks by the intelligent agent, multi-source perception data generated during its interaction with the external environment and other intelligent agents is collected. Specifically, this includes image data, voice data, text information data, and action execution sequence data. Image data is collected through a high-definition wide-angle visual perception module deployed at the front end, with a sampling frequency set to 30 frames / second and an image resolution of 1920×1080. Voice data is captured by a far-field microphone array at a sampling rate of 48kHz. Text information data is recorded as an interaction command log by a task control terminal and segmented according to the granularity of natural language sentences. Action execution sequence data is obtained from three-dimensional acceleration and angular velocity information by a six-axis inertial measurement unit (IMU), with a data sampling rate of 200Hz. All the above multimodal data are synchronized according to the timestamp alignment principle, with a time window width of 5 seconds and a sliding step size of 1 second, forming a multimodal interaction data sample group within a five-dimensional time period. Once synchronization is complete, the original agent multimodal interaction data is formed with the structure of: {image frame sequence, speech spectrum frame sequence, text semantic fragment, action vector sequence, timestamp sequence}, providing basic data support for subsequent feature extraction and relationship modeling.

[0029] Step S12: Perform structural feature fusion processing based on the agent's multimodal interaction data to obtain agent structural feature fusion data;

[0030] In this embodiment of the invention, the multimodal interaction data obtained in step S11 is input to a unified feature processing module, which consists of three sub-processing channels: an image and speech processing channel, a text semantic parsing channel, and an action information fusion channel. The image and speech processing channel uses edge operators to extract the contour features of moving objects within the visual region and extracts Mel-frequency cepstral coefficients (MFCC) from the speech spectrogram, forming a joint perceptual temporal matrix. The text semantic parsing channel uses a dependency parsing algorithm to extract verb-noun-object triples from the instruction text, constructing an action semantic chain. The action information fusion channel identifies stable interaction patterns by calculating the acceleration rate of change and angular velocity extremes of continuous action sequences. The structural features obtained from the three channels are aligned in chronological order, and the image frame contours, speech feature vectors, text semantic triples, and action trend data are mapped to a unified feature vector space. These are then fused and integrated using local feature matching to obtain a structural fusion vector corresponding to each interaction time window, defined as the agent structural feature fusion data. The fused data structure is a fixed-length sequence of feature vectors, each containing 12-dimensional image edge intensity features, 20-dimensional speech MFCC features, 10-dimensional action trend encoding, and 8-dimensional text semantic mapping values.

[0031] Step S13: Construct an initial relational graph of the agents based on the fusion data of the agents' multimodal interaction data and the agents' structural features;

[0032] In this embodiment of the invention, the multimodal interaction data of the agents in step S11 and the structural feature fusion data obtained in step S12 are jointly analyzed to construct an initial relationship graph between agents. Specifically, the initiating node and target node of the interaction behavior are determined based on the subject and object information in the text semantic triples. Secondly, the specific time period of the interaction and the corresponding action feature vector are identified using the time alignment results of image frames and action sequences. When constructing the node set, each agent's behavioral unit within a certain time window is considered an independent node, and each node contains a triple semantic label, a visual feature embedding vector, and an action trend encoding. When constructing the edge set, the interaction behavior between nodes is judged based on the semantic relationship subject-object connection and action synchronization features. If there is an instruction-triggered response relationship or overlapping action times and mutual semantic pointing between two nodes, an edge is established. This edge is marked as a "semantic-action coupling edge" and associated with the Euclidean distance between the timestamp of the interaction and the fused features. In this way, multimodal interaction events are transformed into structural connections in a graph, thus constructing the initial relational graph of the agent. The number of nodes in the graph is proportional to the number of samples in the time window, and the number of edge sets depends on the density of interaction events.

[0033] Step S14: Determine the edge attribute features of the agent based on the initial relation graph of the agent.

[0034] In this embodiment of the invention, based on the constructed initial relationship graph, edge attribute feature extraction is performed on each edge in the graph. Specifically, the following processing steps are included: First, the interaction duration is calculated based on the time window identifiers of the two nodes connected by the edge, defined as the time span feature of the interaction edge; second, the structural fusion feature vectors in the nodes connected by the edge are extracted, and their Euclidean distance is calculated as a semantic action difference measure, defined as the structural difference feature; third, the frequency of recurrence of the edge within five consecutive time windows is counted, serving as the interaction frequency feature; fourth, after word embedding processing of the text semantic triples in the nodes connected by the edge, their semantic similarity scores are calculated, serving as the semantic fit feature of the edge. These four types of features (interaction time span, structural difference, interaction frequency, and semantic fit) are concatenated to form the edge attribute feature vector. The edge attribute features serve as the basic input data for initializing attention weights in the higher-order relationship propagation stage.

[0035] Preferably, Figure 3 This is a schematic diagram of the process for constructing the initial relationship graph of the intelligent agent in this invention;

[0036] Please see Figure 3 This is a schematic diagram of the construction of the initial relationship graph of the intelligent agent in this invention;

[0037] Preferably, step S13 includes the following steps:

[0038] Step S131: Perform interaction semantic extraction processing based on the agent's multimodal interaction data to obtain interaction semantic feature data;

[0039] In this embodiment of the invention, the multimodal interaction data obtained synchronously in step S11 is used to divide the data into unit groups according to time windows. Text semantic data is selected as the main semantic source, and a rule extraction method based on syntactic dependency relations is used to decompose each instruction text into components, extracting the behavioral expression unit centered on verbs. In specific operations, part-of-speech tagging, dependency syntax tree construction, and semantic role tagging are performed on the text to identify the action, agent, and patient respectively. Based on this, a semantic feature fragment in the form of a triple (Agent, Action, Target) is constructed. Next, the image frame data and action sequence data are temporally aligned to identify the spatial consistency between the target object in the image and the action instruction. The position of the subject within the frame is located by color clustering and edge detection, and the existence of a spatial movement relationship between the action initiator and the target object is determined by optical flow change detection. The semantic features are further mapped to a 128-dimensional embedding space through word embedding and weighted combination of text and image semantic weights to generate interactive semantic feature vector data in each interaction window, forming interactive semantic feature data with the structure {triple, image matching confidence, action direction encoding, semantic vector}.

[0040] Step S132: Perform node functional attribute identification processing based on the fusion data of the agent's structural features to obtain node functional attribute data;

[0041] In this embodiment of the invention, based on the agent structure feature fusion data generated in step S12, a sub-vector representing the action pattern and image feature encoding in the fusion vector is selected as the functional behavior recognition input. The action trend encoding part is extracted, and the continuity, acceleration change range, and directional stability of the action pattern are identified. Action sequences with high response frequency and stable directionality are defined as functional behavior segments. Secondly, the density and change gradient information of image edge features are extracted to determine whether the visual focus area has undergone interactive scene switching. If a rapid focusing-discrete pattern appears, it indicates that the node has an "observation-response" attribute. The above two types of features are combined to form a functional attribute vector, which is classified by a regular function, and three types of functional labels are set: control type, response type, and mediator type. Each node is assigned a functional label and functional response parameters (such as reaction time, action duration, and operation frequency). The output structure of the node functional attribute data is: {node number, functional label, action feature vector, visual feature descriptor, response statistical parameters}.

[0042] Step S133: Perform agent node association analysis based on interaction semantic feature data and node functional attribute data to obtain agent node association relationship data;

[0043] In this embodiment of the invention, an association evaluation matrix between nodes is established by combining the interactive semantic feature data generated in step S131 and the node functional attribute data identified in step S132. The specific processing flow includes the following two steps: First, the agent and recipient nodes in each interaction triple are matched, and the corresponding node number is retrieved from the node functional attribute data and its functional label is read; if the triple is {A, transitive, B}, and the label A is control type and B is responsive type, then a basic association is established; Second, the association is confirmed by combining the cosine similarity of the semantic vector (threshold set to 0.65) and the consistency label of the action direction. If the semantic similarity exceeds the threshold and there are continuous movement trajectories between the image targets, then a strong association is confirmed. The above analysis results are constructed into a triple format of "node number pair + association strength + interaction direction", and the output is the intelligent agent node association relationship data. The structure of this dataset is: {node A, node B, association level (strong / medium / weak), interaction direction (A→B)}.

[0044] Step S134: Construct node relationship connection structure data based on agent node association data;

[0045] In this embodiment of the invention, the agent node association data obtained in step S133 is input to the relationship structure generation module to express the behavioral connection relationships between nodes in a directed edge manner. For each association triple, the start and end nodes of the edge are constructed according to the interaction direction, and the association level is assigned as the weight index of the edge; at the same time, the association strength is converted into numerical weight values: strong = 1.0, medium = 0.6, weak = 0.3, which serve as the weight attributes of the edges in the graph structure. All edges constitute an edge set, and each edge element has the structure {source_node_id, target_node_id, weight, direction}, together with the corresponding node set (including node number and functional attribute information), to form the node relationship connection structure data. This data can describe the behavioral paths, response chains, and interaction strengths between various agent nodes.

[0046] Step S135: Perform relation edge generation processing based on node relationship connection structure data and interaction semantic feature data to obtain structural relation edge data;

[0047] In this embodiment of the invention, based on the node relationship connection structure data, each edge is matched with the interaction semantic feature data, and semantic identifiers and semantic similarity scores are added. Specifically, for the starting and ending nodes, their corresponding triples are searched in the interaction semantic feature data, their embedding vectors are read, and their semantic consistency scores are calculated. If the score is lower than 0.4, the edge is discarded. For the retained edges, their corresponding semantic verb tags and behavioral context content are further extracted, serving as semantic annotations to output the structural relationship edge data. This includes: {source_id, target_id, edge_weight, action_verb, context_score, direction}, constructing a structural relationship edge set to describe the semantically enhanced connections in the graph.

[0048] Step S136: Perform node-edge information fusion processing based on node functional attribute data and structural relationship edge data to obtain relation graph structure data;

[0049] In this embodiment of the invention, node functional attribute data and structural relationship edge data are fused. The functional attribute labels of the starting and ending nodes connected by each edge are retrieved to establish edge context structure information. Specifically, the action semantic label of the edge is jointly constructed with the functional labels of the starting and ending nodes to create a "behavior pattern-function comparison structure." For example, if the edge's action is "scheduling," the starting point is "control type," and the ending point is "response type," then the edge is classified as a "dominant-execution link." The fused information is recorded as a six-dimensional tuple: {source_id, target_id, edge_type, function_pair, semantic_verb, weight}. This structure constitutes the relational graph structure data, providing three-dimensional connectivity support for the graph structure: entity-action-response.

[0050] Step S137: Construct the initial relation graph of the agent based on the structural relation edge data and the relation graph structure data.

[0051] In this embodiment of the invention, node functional attribute data and structural relationship edge data are fused. The functional attribute labels of the starting and ending nodes connected by each edge are retrieved to establish edge context structure information. Specifically, the action semantic label of the edge is jointly constructed with the functional labels of the starting and ending nodes to create a "behavior pattern-function comparison structure." For example, if the edge's action is "scheduling," the starting point is "control type," and the ending point is "response type," then the edge is classified as a "dominant-execution link." The fused information is recorded as a six-dimensional tuple: {source_id, target_id, edge_type, function_pair, semantic_verb, weight}. This structure constitutes the relational graph structure data, providing three-dimensional connectivity support for the graph structure: entity-action-response.

[0052] Preferably, in step S2, the initial attention weight allocation is performed based on the agent's edge attribute features to obtain the initial edge attention weight data, which includes:

[0053] Edge semantic encoding is performed based on the edge attribute features of the agent to obtain the agent edge semantic encoding data;

[0054] In this embodiment of the invention, based on the agent edge attribute feature data determined in step S14, multi-dimensional attribute features of the edges are extracted, including edge weights, semantic verb labels, interaction direction, node function labels, and interaction duration. A fixed-length multi-dimensional vector representation is constructed for the above attributes. Word vector encoding technology is used to encode the semantic labels of the edges, converting the verb labels into corresponding word vector representations. A pre-trained word embedding dictionary (such as GloVe or FastText) is used to map the verb labels to word vectors, resulting in a 300-dimensional semantic vector. Other numerical attributes (edge ​​weights, interaction duration, etc.) are normalized and then concatenated with the semantic vector to form an edge attribute vector. The edge attribute vector undergoes non-linear mapping processing using a multilayer perceptron (MLP), and the output dimension is uniformly adjusted to 256 dimensions, serving as the edge semantic encoding data. This edge semantic encoding data is a high-dimensional dense vector containing the semantic information and numerical features of the edges, serving as the basic input for subsequent frequency intensity and temporal feature extraction.

[0055] Extracting the frequency and intensity of agent-side interactions based on agent-side semantic encoding data;

[0056] In this embodiment of the invention, the interaction trigger counts for each edge in the semantically encoded edge data within a defined time period are calculated. Specifically, the activation frequency of the edge within the time window is calculated by scanning the timestamp sequence contained in the edge attributes, forming edge interaction frequency data. The edge interaction frequency data is normalized according to the edge number, with the range limited to 0 to 1, reflecting the strength of different edge interaction frequencies. The normalization method uses max-min normalization, that is, the maximum edge interaction frequency within the time period is set to 1, and the minimum is set to 0, linearly mapping the frequency values ​​of each edge. The edge interaction frequency strength data is appended to the corresponding edge semantically encoded data in vector form for subsequent temporal feature analysis and semantic vector generation.

[0057] Extracting temporal features of agent-side interactions based on agent-side semantic encoding data;

[0058] In this embodiment of the invention, temporal series features of edge interactions are extracted by combining edge interaction frequency and intensity data. Specifically, temporal statistical features are calculated using the activation timestamp sequence of edges, including average activation interval, activation interval variance, activation time distribution skewness, and kurtosis. A sliding window technique is employed, using a fixed-length time window (e.g., 10 seconds) to analyze the trend of activation interval changes and identify periodic and abrupt changes in frequency. After normalizing the above temporal statistical indicators, they are concatenated with the semantic encoding vector of the edge to form an edge interaction temporal feature vector. This temporal feature vector reflects the interaction rhythm and dynamic changes of the edge, providing temporal information dimensions for the subsequent construction of the edge semantic vector representation.

[0059] The semantic vector representation data of the agent's side interaction is determined based on the temporal characteristics of the agent's side interaction and the frequency and intensity of the agent's side interaction.

[0060] In this embodiment of the invention, a composite semantic vector is constructed by combining the normalized edge interaction frequency intensity with the edge interaction temporal feature vector. Specifically, the two are fused element-wise with weights, the weight allocation being fixed based on the contribution ratio of frequency and temporal sequence to agent behavior prediction in historical data, such as a frequency weight of 0.6 and a temporal sequence weight of 0.4. The fused vector is input to the feature fusion module, which uses a hierarchical feature fusion algorithm to extract the cross-relationships between features layer by layer, outputting an edge semantic vector representation with a unified dimension of 256. This semantic vector represents the comprehensive semantic and pragmatic representation of the edge, containing static frequency intensity and dynamic temporal sequence information, used to identify the edge's influence and neighborhood characteristics.

[0061] Based on the semantic vector representation of the agent's edge, the distribution of the agent's node structure in the neighborhood can be identified.

[0062] In this embodiment of the invention, edge semantic vectors are used to represent data and perform topological analysis on the neighborhood structure of agent nodes. Specifically, the adjacency matrix construction method in graph theory is adopted. Adjacency weights are calculated based on the semantic vector similarity of edges, and cosine similarity is used as the criterion for determining edge weights. For each node, the similarity between the semantic vectors of all its incoming and outgoing edges and other edge vectors is calculated to construct a weighted adjacency matrix. Based on the adjacency matrix, the structural neighborhood features of the nodes are extracted, including the node's degree, clustering coefficient, and the distribution of adjacent edge weights, forming a node neighborhood distribution vector. The node neighborhood distribution vector characterizes the local connectivity density and semantic consistency of nodes in the graph structure, providing key parameters for determining edge adjacency structure features.

[0063] The edge adjacency structure feature data is determined based on the agent edge semantic vector representation data and the agent node structure neighborhood distribution.

[0064] In this embodiment of the invention, edge adjacency structure features are constructed by combining the semantic vector of an edge with the neighborhood distribution of nodes. Specifically, for each edge, the neighborhood distribution vector of its connected nodes is obtained, and the difference in neighborhood features between the two endpoints is calculated (using Euclidean distance). The edge adjacency structure features include the semantic vector of the edge, the neighborhood distribution vectors of the two endpoints, and the neighborhood difference. These feature vectors are concatenated into a unified high-dimensional vector, which serves as the edge adjacency structure feature data. This data reflects the fusion state of the edge's semantic attributes and the structural environment of the connected nodes, providing a multi-dimensional information foundation for subsequent edge relationship context fusion modeling.

[0065] Edge relationship context fusion modeling is performed based on edge adjacency structure feature data to obtain edge relationship context representation data;

[0066] In this embodiment of the invention, a multi-head attention mechanism is employed to achieve context fusion modeling for edge adjacency structure feature data. Specifically, multiple attention heads are used in parallel to calculate the correlation weights between edge adjacency feature vectors and neighbor edge vectors, capturing the multidimensional interaction relationships between edges. Each attention head generates a weighted feature representation, and the outputs of multiple heads are concatenated and merged before being input into a feedforward neural network for nonlinear transformation to obtain a unified edge relationship context representation vector. This process integrates the ontology semantics of edges, adjacency structure information, and neighbor relationship features, with the output dimension set to 512 dimensions. This edge relationship context representation data constitutes the context-aware features between graph edges, endowing edges with a semantic understanding of the graph structure.

[0067] The initial attention weights are assigned based on the edge relation context representation data to obtain the initial edge attention weight data.

[0068] In this embodiment of the invention, initial edge attention weights are calculated based on edge relation context representation data. A dot product operation is performed between the edge relation context vector and a training-initialized weight vector to obtain the weight score for each edge, reflecting the importance of the edge in the current graph structure and semantic environment. Softmax normalization is applied to all edge weight scores to ensure the weight distribution satisfies probability distribution properties. The output initial edge attention weight data is a standardized set of weight values ​​in the form of {edge ID, attention weight value}, which serves as the input for subsequent edge weight enhancement graph construction and multi-hop higher-order propagation.

[0069] Preferably, in step S2, multi-hop higher-order relation propagation processing is performed based on the edge weight enhancement graph data to obtain multi-hop higher-order relation propagation data, including:

[0070] Extract multi-order adjacency information of agent nodes based on edge weighted graph data;

[0071] In this embodiment of the invention, a multi-order adjacency matrix of nodes is constructed based on edge-weighted enhanced graph data, i.e., the attention weights of each edge in the graph have been enhanced. Multi-order adjacency relationships refer not only to the direct neighbors of a node, but also to its neighbors at hop counts of 2, 3, or even higher. A set of multi-order adjacency matrices is obtained by exponentiation of the weighted adjacency matrix of the edge-weighted enhanced graph. Specifically, the original adjacency matrix is ​​exponentiated, and the non-zero elements in the matrix represent the connection strength between nodes at the corresponding order, ranging from order 1 to a preset maximum order N (e.g., N=3). In the multi-order adjacency matrix, the element values ​​represent the connection strength between nodes. Combined with the attention values ​​after edge weight enhancement, the multi-order connection relationships of nodes are weighted and described to form multi-order adjacency relationship information. This information is stored in a sparse matrix format, supporting fast retrieval of a node's multi-order neighbors and their connection strengths.

[0072] Jump distance marking is performed based on the multi-level adjacency relationship information of the agent nodes to obtain jump distance marked edge structure data;

[0073] In this embodiment of the invention, the aforementioned multi-level adjacency relationship information of nodes is used for jump distance marking. Jump distance marking assigns a jump count identifier to each edge within the multi-level connection of nodes, distinguishing whether an edge belongs to a one-hop, two-hop, or three-hop relationship. Specifically, the method involves traversing the multi-level adjacency matrix and marking the jump distance of the corresponding edge based on the position of the non-zero element and its corresponding order. The original edge set is expanded, retaining the original edges (1-hop edges) while introducing 2-hop and 3-hop edges, calculated by connecting pairs of nodes with path lengths of 2 or 3. Each jump distance marked edge is appended with a jump count attribute field to clarify the jump distance level. The jump distance marked edge structure data is stored in the form of an edge list containing the edge's start point, end point, and jump distance attribute, providing basic data for multi-hop path candidate generation.

[0074] Candidate data for multi-hop paths of nodes are determined by using hop distance marked edge structure data and multi-order adjacency relationship information of agent nodes;

[0075] In this embodiment of the invention, a multi-hop path candidate set is constructed based on the hop distance marked edge structure data and combined with the multi-level adjacency relationship information of nodes. A multi-hop path candidate is defined as a path starting from one node and connecting to another node via multi-hop edges according to the hop distance hierarchy. Specific steps include: sequentially traversing the hop distance marked edge set, using a depth-first search (DFS) algorithm to traverse all paths within the hop distance constraint, extracting the edge weights on the paths (provided by the edge weight enhancement graph), and calculating the path weight as the product or weighted sum of all edge weights on the path. Paths with higher path weights are selected as candidate paths. The multi-hop path candidate data includes the path start point, end point, path hop count, path weight, and path edge sequence, stored using a linked list or adjacency list data structure, providing input for behavioral similarity calculation.

[0076] Calculate the node behavior similarity factor based on the candidate data of multi-hop paths of nodes;

[0077] In this embodiment of the invention, the behavioral similarity factor is used to measure the similarity of behavioral features of nodes on multi-hop paths, reflecting the behavioral consistency of multi-hop path connections between nodes. Specifically, the method involves calculating the similarity of behavioral feature vectors of the starting and ending nodes in the candidate multi-hop path data. The behavioral feature vectors originate from statistical features of previously extracted multimodal interaction data of the nodes, such as interaction frequency, interaction timing, and node functional attributes. Using the cosine similarity calculation formula, the behavioral vectors of the starting and ending nodes are mapped to a unit sphere, and the similarity value is obtained through vector dot product, with a value range of [-1, 1]. The calculated node behavioral similarity factor is bound to the corresponding multi-hop path, forming a set of node behavioral similarity factors, which is used to weight the propagation influence between nodes during subsequent multi-hop propagation aggregation.

[0078] Multi-hop propagation aggregation data is obtained by performing multi-hop propagation aggregation processing based on node behavior similarity factors and node multi-hop path candidate data;

[0079] In this embodiment of the invention, the multi-hop propagation aggregation process combines node behavior similarity factors to perform weighted propagation aggregation on multi-hop path candidate data. The specific steps are as follows: Taking each node as the center, traverse all its multi-hop path candidates, and calculate the propagation weight as the product of the node behavior similarity factor and the path weight, reflecting the combined influence of path weight and behavior similarity. Through weighted summation, the feature information of neighboring multi-hop path candidates is aggregated to the central node, forming an enhanced feature vector of the node under multi-hop propagation. Batch weighted aggregation is achieved using matrix multiplication, outputting multi-hop propagation aggregated data. This data is a high-dimensional dense vector representing the node dimension, reflecting the comprehensive expression of the node's behavior and structural information based on multi-hop path propagation.

[0080] Cross-node relationship coupling calculations are performed based on multi-hop propagation aggregated data to obtain cross-node relationship coupling data;

[0081] In this embodiment of the invention, cross-node relationship coupling calculation aims to capture the deep relationship coupling features generated between different nodes through multi-hop paths. Specifically, multi-hop propagation aggregated data is used as input, and the coupling degree between node pairs is calculated using correlation coefficients. The coupling strength between nodes is obtained by calculating the Pearson correlation coefficient matrix between node feature vectors. The node pair coupling strength matrix is ​​further combined with multi-hop path weights and weighted fusion is used to generate a cross-node relationship coupling matrix. This matrix represents the potential high-order coupling relationships between nodes in the agent network and is stored in a sparse matrix format for subsequent high-order semantic expansion.

[0082] Determine higher-order semantic extended data based on cross-node relationship coupling data;

[0083] In this embodiment of the invention, higher-order semantic extension is performed using cross-node relationship coupling data. Specifically, the coupling matrix is ​​subjected to spectral decomposition to extract the implicit semantic structure between nodes. Eigenvalue decomposition (EVD) is used to decompose the coupling matrix into eigenvectors and eigenvalues, and the eigenvectors corresponding to the K largest eigenvalues ​​are selected as the higher-order semantic representation basis. Higher-order semantic extension data is constructed by linearly combining the eigenvectors, forming the semantically enhanced representation of the nodes. This higher-order semantic extension data reveals the deep semantic connections between nodes, supporting the propagation of higher-order relationships between agents.

[0084] Multi-hop higher-order relation propagation data is obtained by performing multi-hop higher-order relation propagation processing based on higher-order semantic extended data and cross-node relation coupling data.

[0085] In this embodiment of the invention, multi-hop high-order relationship propagation is performed by combining high-order semantic extension data and cross-node relationship coupling matrix. An iterative propagation algorithm is employed, with initial node features based on the high-order semantic extension data, and information between nodes is transmitted through the coupling matrix. In each iteration, the node feature vector is multiplied by the coupling matrix to update the node representation. The number of iterations is set to M (e.g., M=5) to ensure sufficient information diffusion. After the iterations are completed, the output set of node feature vectors constitutes the multi-hop high-order relationship propagation data, representing the node-level semantic fusion and structural propagation results, which serves as input for subsequent task decision path construction.

[0086] Preferably, in step S2, the agent node representation enhancement processing is performed based on the multi-hop higher-order relation propagation data to obtain agent node representation enhancement data, including:

[0087] Extract upstream and downstream interaction path information of agent nodes from multi-hop high-order relationship propagation data;

[0088] In this embodiment of the invention, multi-hop higher-order relationship propagation data is used. This data, organized by node, includes enhanced feature vectors obtained by nodes through multi-hop propagation. The upstream and downstream interaction path information of nodes is extracted. Specifically, the upstream and downstream connection paths of each node are identified by parsing the path transmission relationships during the multi-hop higher-order propagation process. A graph traversal algorithm is used to perform directed path scanning on the higher-order relationship graph, extracting all upstream and downstream path sequences containing the target node. The path sequence includes the path start point, end point, path length, and the sequence of nodes along the path. The path information is stored in a path list structure, providing basic data input for statistically analyzing the distribution of node interaction frequencies. After this step, a set of upstream and downstream interaction paths corresponding to all nodes in the agent is obtained, serving as input for subsequent node interaction frequency analysis.

[0089] Statistical distribution data of node interaction frequency is compiled based on the upstream and downstream interaction path information of intelligent agent nodes;

[0090] In this embodiment of the invention, the frequency of each node in the path is statistically analyzed using the obtained upstream and downstream interaction path information to obtain node interaction frequency distribution data. The statistical process employs a counting algorithm to accumulate the number of times each node appears in all paths, and then calculates a weighted frequency by combining the path weights (calculated by multiplying edge weights during multi-hop higher-order relationship propagation). The node interaction frequency distribution data is stored as a pair of node identifiers and corresponding weighted frequency values, reflecting the node's activity and importance in multi-hop paths. After statistical analysis, this data is used to identify high-frequency interaction paths, providing a basis for path reduction and node feature fusion.

[0091] High-frequency path reduction processing is performed based on the interaction frequency distribution data to obtain high-frequency path structure reduction data;

[0092] In this embodiment of the invention, a threshold screening method is used to reduce paths with high interaction frequencies based on the node interaction frequency distribution. The reduction operation constructs a simplified high-frequency path structure and reduces path redundancy by merging paths with highly overlapping node sequences and high interaction frequencies. During the reduction process, path similarity indices (such as the Jaccard similarity coefficient of path node sequences) are calculated, and paths with similarity exceeding a set threshold are merged. The weights of the merged paths are calculated by weighted averaging of the original path weights. The reduced paths are stored as path node sequences and weighted weights, forming high-frequency path structure reduction data. This data significantly reduces the size of the path set, retains key interaction structures, and supports efficient node feature fusion.

[0093] Node feature fusion processing is performed based on high-frequency path structure reduction data and upstream and downstream interaction path information of agent nodes to obtain node feature fusion data.

[0094] In this embodiment of the invention, the node feature fusion processing takes high-frequency path structure reduction data and complete upstream and downstream interaction path information as input, and fuses the multi-dimensional features of nodes in different paths. The fusion method includes weighted averaging of node feature vectors on the path, with the weights determined based on path weights and a decreasing function of the node's position in the path. Multiple path features of a node are integrated using a combination of vector concatenation and weighted aggregation to obtain the fused node feature representation. During the fusion process, normalization is applied to ensure that the node feature dimensions and numerical ranges are consistent. The output node feature fusion data is a high-dimensional feature matrix at the node level, which reflects in detail the comprehensive interaction characteristics of nodes in multi-hop paths, serving as the basis for subsequent stability assessment.

[0095] Node state stability is evaluated based on the fused data of node features to obtain node state stability data.

[0096] In this embodiment of the invention, node state stability assessment uses node feature fusion data as input to calculate the state stability index for each node. This index employs a time-series statistical method to analyze the amplitude and variance of node feature changes during multi-round, multi-hop propagation. Specifically, it calculates the Euclidean distance and standard deviation of the node feature vector over consecutive time steps or iteration rounds. Nodes with smaller stability values ​​indicate fewer state changes and exhibit higher stability. The stability calculation results are stored as node identifiers paired with corresponding stability values, forming node state stability data. This data provides a quantitative basis for anomaly feature screening and reflects the dynamic changes in node features during multi-hop propagation.

[0097] Based on node state stability data, abnormal node feature dimensions are filtered to obtain abnormal node feature data.

[0098] In this embodiment of the invention, the abnormal node feature dimension screening identifies dimensions with large fluctuations or instability in the feature vector by analyzing node state stability data. A variance thresholding method is used to calculate the variance across time steps for each node's feature dimensions; dimensions exceeding a preset threshold are identified as abnormal dimensions. All abnormal dimensions are aggregated to form a subset of node abnormal features. This step ensures that unstable feature dimensions are removed from high-dimensional node features, highlighting abnormal information. The screening results are stored as node identifiers and their corresponding sets of abnormal feature dimensions, serving as input for higher-order semantic correction and improving the accuracy of subsequent representations.

[0099] High-order semantic correction is performed based on node anomaly feature data to obtain high-order semantic correction node data.

[0100] In this embodiment of the invention, the high-order semantic correction operation uses a local reconstruction method to correct the selected abnormal feature dimensions of nodes. The feature values ​​of the abnormal dimension are corrected by calculating a weighted average of the abnormal feature dimensions of the node and the features of the same dimension of its neighboring nodes. The weighting coefficients are determined based on the interaction frequency of neighboring nodes and edge weights. The correction process avoids the negative impact of outliers on the overall node representation and enhances the coherence of semantic expression. After correction, the node feature vector is updated to form high-order semantically corrected node data. This data is stored in the form of a node feature matrix, providing a stable and semantically coherent node representation to support subsequent representation enhancement.

[0101] The agent node representation enhancement data is obtained by performing intelligent agent node representation enhancement processing based on high-order semantic correction node data and node feature fusion data.

[0102] In this embodiment of the invention, the high-order semantic correction node data is finally combined with the node feature fusion data to perform node representation enhancement processing. The combination method involves weighted fusion of feature vectors from the two data sources, with the weights dynamically adjusted based on the stability and semantic coherence of the node features. Feature vector normalization and linear mapping are applied during the fusion process to ensure numerical consistency of the fusion results. The fusion result is the agent node representation enhancement data, which represents a comprehensive expression of multimodal, high-order semantic, and behavioral features at the node level. This data serves as an important input for subsequent steps such as constructing the agent's task decision path, completing the high-order relational information representation at the node level.

[0103] Preferably, step S3 includes the following steps:

[0104] Step S31: Construct the task decision path based on the agent node representation enhancement data and multi-hop high-order relationship propagation data to obtain the agent task decision path;

[0105] In this embodiment of the invention, agent node representation enhancement data and multi-hop higher-order relationship propagation data are used as inputs. The agent node representation enhancement data includes a fused feature vector of each node after multi-hop higher-order semantic correction, and the multi-hop higher-order relationship propagation data includes multi-hop propagation information between nodes after edge weight enhancement. Combining node features and higher-order relationship information between nodes, an agent task decision path is constructed. Specifically, a directed graph path search algorithm is used to traverse the higher-order relationship graph. The path search rules are determined based on a comprehensive evaluation function of node feature similarity and edge weights, selecting paths with high feature similarity and strong edge weights as candidate decision paths. The path construction process involves path length constraints, a multi-path candidate mechanism, and node priority ranking to ensure that the path covers key agent nodes and key behavior sequences. Path information includes path node sequences, node representation features, and cumulative path weights. After construction, an agent task decision path dataset is formed, providing input for subsequent path feasibility evaluation.

[0106] Step S32: Based on the agent's task decision path, conduct an agent path feasibility assessment to obtain agent path feasibility data;

[0107] In this embodiment of the invention, path feasibility assessment is based on the agent task decision path generated in step S31. For each path, a feasibility score is calculated by combining path node features, edge weights, path length, and the density of multi-hop relationships between nodes in the path. The scoring method employs a multi-index weighted comprehensive model, with indices including path connectivity, node state stability, edge weight strength, and path behavior consistency. Specifically, this is achieved through the following operations: extracting stability data for each node in the path to evaluate the overall stability of the path; combining edge weight enhancement information to calculate the weighted average of edges in the path; and calculating path behavior consistency through node behavior similarity. The scoring results are stored as path identifiers and their corresponding feasibility values, forming an agent path feasibility dataset. This data quantitatively expresses the rationality and effectiveness of path execution in the current agent state and higher-order relationship network.

[0108] Step S33: Perform a three-level cognitive reflection operation on the agent's path feasibility data to obtain three-level cognitive reflection data, including result reflection data, process reflection data, and strategy reflection data;

[0109] In this embodiment of the invention, the three-level cognitive reflection operation, based on the agent path feasibility data from step S32, performs in-depth analysis from three dimensions: result, process, and strategy. Result reflection data assesses the achievement of the path execution goal by comparing the actual execution results with the expected effects, and quantifies result deviations using error analysis and deviation detection methods. Process reflection data analyzes the path selection logic, node switching frequency, and dynamic changes in edge weights for key steps in the path construction and execution process, and identifies abnormal fluctuations or bottleneck nodes using time series analysis techniques. Strategy reflection data integrates path selection strategies and weight allocation mechanisms to evaluate the adaptability and adjustment effects of the strategy in multi-hop propagation and edge weight allocation, and quantifies them using strategy effectiveness metrics. The three types of reflection data are recorded in structured tables, recording path identifiers, reflection index values, and corresponding analysis results, and are synthesized into three-level cognitive reflection data.

[0110] Step S34: Perform fusion processing on the three-level cognitive reflection data to obtain the agent reflection fusion data.

[0111] In this embodiment of the invention, the result reflection data, process reflection data, and strategy reflection data generated in step S33 are fused to comprehensively reflect the cognitive state of the agent's task path. The fusion process employs a weighted fusion algorithm. Based on the weight allocation of each reflection data point, the three dimensions of reflection indicators are normalized, and a weighted sum is calculated to form a single comprehensive reflection indicator. During the fusion process, data alignment technology is used to ensure consistency in time and path identification across different dimensions, avoiding information misalignment. The fused agent reflection data includes path identification, comprehensive cognitive score, and a multi-dimensional reflection indicator matrix. The data format facilitates subsequent task decision adjustments and path optimization. This data serves as the input for the agent's next step of optimizing the path based on an edge-state reinforcement learning model, completing a systematic summary of the path's cognitive state.

[0112] Preferably, step S31 includes the following steps:

[0113] Step S311: Extract node task perception feature data based on the agent node representation enhancement data;

[0114] In this embodiment of the invention, enhanced agent node representation data is used as the foundation. This data includes fused feature information of nodes based on multi-hop higher-order relationship propagation. A feature selection algorithm is employed to extract task-related dimensions from the node representation data. Principal component analysis (PCA) is performed on the node representation feature matrix to reduce dimensionality, remove noise and redundant information, and extract the dominant features representing the node's task perception. Subsequently, based on feature importance scores, key feature dimensions reflecting the node's task state are selected, including node activity, node correlation, and edge weight influence. The extraction results form a node task perception feature dataset, whose data structure includes a unique node identifier and its corresponding multi-dimensional perception feature vector. This data provides an accurate description of the task state for subsequent path boundary determination and connectivity calculation.

[0115] Step S312: Determine the path node boundary information data based on the node task perception feature data and the multi-hop high-order relationship propagation data;

[0116] In this embodiment of the invention, step S312 combines node task-aware feature data with multi-hop higher-order relationship propagation data to define the boundary attributes of path nodes. Multi-order adjacency information between nodes is extracted from the multi-hop higher-order relationship propagation data, including hop count, edge weight, and semantic correlation between nodes. Based on node task-aware features, a threshold determination method is used to determine the boundary state of nodes. Boundary nodes are defined as nodes whose task-aware feature indicators are in the edge range, or nodes whose propagation weights differ significantly from those of their neighboring nodes. By comparing the adjacency matrix and the task-aware vector, the starting node, ending node, and key segmentation nodes in the path are selected. The node boundary information data consists of node identifiers, boundary types, and relevance indicators, providing structured input for path connectivity evaluation.

[0117] Step S313: Calculate path connectivity based on path node boundary information data and multi-hop higher-order relationship propagation data to obtain path connectivity data;

[0118] In this embodiment of the invention, for defined path node boundaries, the connectivity index of the path is calculated by combining the inter-node propagation weights and hop distance information in the multi-hop higher-order relationship propagation data. The specific process is as follows: Based on the path node boundary information, a path subgraph is constructed, including the connection relationships between boundary nodes and intermediate nodes; a weighted graph connectivity algorithm is used to calculate the connectivity of each pair of nodes within the path. The connectivity calculation is based on the weighted sum of edge weights and the reciprocal of the node hop count, expressed as connectivity score = Σ(edge ​​weight × reciprocal of hop count), reflecting the density and information flow efficiency of the path. During the calculation process, the weight distribution of node adjacency relationships is dynamically adjusted to ensure that the true multi-hop relationship structure of the path is reflected. Path connectivity data is represented by path identifiers and corresponding connectivity values, providing a core basis for path ranking optimization.

[0119] Step S314: Perform path optimization and sorting processing based on path connectivity data to generate task path optimization and sorting data;

[0120] In this embodiment of the invention, based on the path connectivity data obtained in step S313, a sorting algorithm is used to optimize and sort all candidate paths. The sorting is based on connectivity scores, arranged from highest to lowest, ensuring that paths with strong connectivity and high information propagation efficiency are given priority. The sorting algorithm employs efficient algorithms such as quicksort or heapsort to guarantee computational efficiency when processing massive amounts of path data. The sorting results generate a task path optimization sorting dataset, whose data structure includes path identifiers, connectivity scores, and sorting numbers. This data provides high-quality path candidates for subsequent task decision-making path construction, ensuring that path selection has strong inherent connectivity and execution rationality.

[0121] Step S315: Construct task decision paths based on task path optimization and sorting data to obtain the agent's task decision paths.

[0122] In this embodiment of the invention, task path optimization ranking data is used as input, and a task decision path for the intelligent agent is constructed by combining the path ranking sequence number and connectivity index. The task decision path construction follows a strategy of importing paths one by one, starting with the path with the highest connectivity. Paths with common nodes and edges are merged using path node sequence recombination technology to form a more complete and efficient task path chain. During the construction process, path node and edge weight information is integrated into the path structure, forming a comprehensive description of the path node sequence, node feature fusion data, and edge weight distribution. The generated intelligent agent task decision path data structure includes path identifiers, node sequences, cumulative path weight values, and comprehensive path connectivity indexes, laying the data foundation for subsequent task execution and path optimization.

[0123] Preferably, step S32 includes the following steps:

[0124] Step S321: Statistically analyze the load status of agent path nodes based on the agent's task decision path;

[0125] In this embodiment of the invention, intelligent agent task decision path data is used as input. This data includes the sequence information of each node in the path and its related task characteristics. Load status is defined as the amount of tasks undertaken by a node in the path and its resource usage. For each node in the path, resource consumption data during actual task execution is collected, including key performance indicators such as CPU utilization, memory utilization, and task request queue length. Resource consumption data is collected in real time through a performance monitoring system during node operation to ensure data timeliness and accuracy. During the statistical process, the above performance indicators are normalized to form a load indicator with unified dimensions. The load indicators of all nodes in the path are statistically analyzed one by one to generate a node load status data table, including node ID, resource consumption indicator, and its normalized load value, providing basic data for subsequent response time and load volatility analysis.

[0126] Step S322: Calculate the average response time of the agent nodes based on the load status of the agent path nodes;

[0127] In this embodiment of the invention, the average response time of each node is calculated based on the node load status data collected in step S321. Response time is defined as the time interval from when a node receives a task request to when it completes a response. The node response latency is calculated by extracting timestamp information from the task execution log. Specifically, for each node, task request response time records within a certain time window are collected, the response latency is calculated using the timestamp difference, outliers are removed, and the average is calculated. This average response time reflects the node's processing capacity and load pressure. The calculation results are structured data consisting of the node ID and the corresponding average response time, for subsequent evaluation of load volatility and response stability.

[0128] Step S323: Detect the load fluctuation rate of the agent node based on the load status of the agent path node;

[0129] In this embodiment of the invention, the node load volatility index is calculated based on the node load status time series data in step S321. Load volatility is defined as the ratio of the standard deviation to the mean of the node load status, used to measure the stability of the node load. The specific operation process is as follows: calculate the standard deviation σ and the mean μ of the load value sequence of the node within a given time period, and calculate the volatility index as σ / μ. A higher load volatility value indicates a more unstable node load. The node load volatility data includes the node ID and its corresponding volatility value, providing a basis for evaluating the stability of the node response. This data structure can be automatically extracted using time series analysis methods.

[0130] Step S324: Evaluate the response stability of the agent node based on the agent node load volatility and the agent node average response time;

[0131] In this embodiment of the invention, the average response time calculated in step S322 and the load volatility calculated in step S323 are comprehensively considered to quantitatively evaluate the node response stability. Response stability is defined as an index of the persistence and volatility of the node response time. Specifically, a weighted composite function is used to combine the average response time and load volatility to calculate the node response stability score, ensuring that response time and volatility have equal weight or adjusting the weight allocation according to actual needs. The evaluation result consists of a node ID and a response stability score, used to represent the node's response reliability and stability during task execution.

[0132] Step S325: Determine the collaborative efficiency of the agent operation based on the stability of the agent node response;

[0133] In this embodiment of the invention, the overall collaborative efficiency of the agent is calculated by aggregating the response stability scores of all path nodes. Collaborative efficiency is defined as the weighted average of the response stability of nodes in the path, with weights allocated according to the task importance of each node in the path. The specific calculation process is as follows: The response stability scores of all nodes in the path are calculated, each node's score is multiplied by its corresponding task weight, the sum is divided by the total weights, and the collaborative efficiency index E is obtained. This index reflects the overall collaborative performance of nodes in the path. Collaborative efficiency data includes path identifiers and corresponding efficiency values, serving as key parameters for path feasibility assessment.

[0134] Step S326: Based on the agent's operational coordination efficiency and the agent node response stability, conduct an agent task path feasibility assessment to obtain agent path feasibility data.

[0135] In this embodiment of the invention, the node response stability index from step S324 and the operational coordination efficiency index from step S325 are integrated to construct a comprehensive path feasibility assessment model. A multi-factor evaluation method is used to weight and synthesize the node response stability score and path coordination efficiency to generate a path feasibility score P. This ensures a comprehensive reflection of path performance and stability. Path feasibility data includes path identifiers, feasibility scores, and corresponding evaluation levels, used for subsequent path selection and optimization. This data forms an agent path feasibility dataset, ensuring the reliability and efficiency of task paths during execution.

[0136] Of particular importance, step S33 includes the following steps:

[0137] Step S331: Perform path target deviation analysis based on the agent's path feasibility data to obtain path target deviation data;

[0138] In this embodiment of the invention, based on the obtained agent path feasibility data, the numerical offset between the target state and the actual achieved state in each task path is compared node by node. A path target attribute consistency calculation method is used to perform field-level error calculation between the task weight factor and key output parameter values ​​in the expected task completion state and the state parameters after actual path execution, resulting in a target offset vector. For each task node in the path, the difference between its target task description and the actual achieved index is recorded, constructing a path target deviation matrix. Furthermore, Euclidean distance and attribute matching degree function are used to cluster the degree of target deviation, outputting path target deviation data representing the overall deviation of the agent's task paths for use in subsequent result reflection stages.

[0139] Step S332: Perform basic processing of the results reflection based on the path target deviation data to obtain the basic data for results reflection;

[0140] In this embodiment of the invention, based on path target deviation data, a target difference hierarchical attribution analysis algorithm is used to classify reflection trigger points of different severity levels according to the numerical distance between the node target achievement index and the task specification benchmark. A task completion rate threshold setting rule is introduced; when the node target completion rate is lower than this threshold, it is marked as a result reflection trigger point, and its location index, deviation vector, task label, etc., are extracted to form a structured reflection trigger record. Combining various target deviation labels, such as "path interruption," "boundary transition failure," and "critical task delay," a classification record is formed for the trigger nodes, and a basic reflection information set is constructed, resulting in result reflection basic data containing information such as target offset value, deviation trend type, and affected path segment number.

[0141] Step S333: Based on the results, reflect on the basic data and perform path node operation tracking processing to obtain path node operation tracking data;

[0142] In this embodiment of the invention, based on the path target deviation node index marked in the result reflection basic data, the path execution record is traced back to perform operation tracking on the execution time sequence log of the corresponding agent node. Behavioral data such as upstream and downstream task nodes, trigger edge information, jump distance, and response delay involved in the path execution process of each target deviation node are extracted. A directed operation record chain construction mechanism is adopted to establish an operation sequence containing "node number—trigger edge number—trigger timestamp—feedback status," obtaining path node operation tracking data. To facilitate causal chain analysis in the subsequent process reflection stage, this tracking data structure must retain execution order information and state change trajectory information.

[0143] Step S334: Perform path decision causal chain verification based on path node operation tracking data to obtain process reflection feature data;

[0144] In this embodiment of the invention, based on path node operation tracking data, a decision causal chain verification graph is constructed to track the upstream triggering conditions and downstream response paths corresponding to the target deviation generation nodes. A link consistency verification method based on offset factor propagation is applied to perform hop-by-hop causal consistency verification of the triggering relationships between nodes, edge weight changes, and node state changes, identifying execution misalignments, edge weight distortions, or higher-order semantic jump errors in the path. Node segments with structural problems such as abnormal jumps, invalid feedback loops, or path breaks in the causal chain are marked, and causal break labels and correlation factors are extracted to constitute process reflection feature data.

[0145] Step S335: Perform strategy reflection correlation processing on the process reflection feature data and the result reflection basic data to obtain strategy reflection correlation data;

[0146] In this embodiment of the invention, process reflection feature data and result reflection basic data are fused to establish a policy causal chain from task goal deviation to behavioral path deviation. A policy factor co-occurrence analysis method is applied to extract the policy units with the highest frequency and largest fluctuations in edge adjacency strength in the deviation trigger path. Further identification of structural imbalance points and cognitive gaps in the policy usage process is performed, such as the lack of intermediary node support during policy switching or policy failure due to inconsistent edge semantics. Through a path-policy reasoning trajectory mapping mechanism, a causal integration chain from result deviation to process tracing to policy attribution is constructed, forming structured policy reflection association data, including information such as policy failure nodes, associated control instructions, and policy coupling failure sections between nodes.

[0147] Step S336: Perform a three-level cognitive reflection operation based on the strategy reflection correlation data to obtain three-level cognitive reflection data, including result reflection data, process reflection data and strategy reflection data.

[0148] In this embodiment of the invention, based on strategy reflection correlation data, result deviation factors, process execution misalignment features, and strategy failure modes are extracted to construct a three-level cognitive reflection structure map of "result-process-strategy". Using a third-order cognitive chain mapping algorithm, each deviation path is covered with three-dimensional cognitive labels, marking its target off-target area at the result level, causal jump nodes at the process level, and control strategy conflict relationships at the strategy level. Each cognitive dimension constructs a reflection data structure composed of structural labels, parameter labels, logical associations, and edge weight contexts, forming a three-level cognitive reflection data set including result reflection data, process reflection data, and strategy reflection data. This data structure will provide causal chains, structural focuses, and strategy correction targets for subsequent path deviation correction.

[0149] Preferably, step S4 includes the following steps:

[0150] Step S41: Perform cognitive distillation on the agent's reflection and fusion data to obtain lightweight knowledge unit data;

[0151] In this embodiment of the invention, agent reflection fusion data is used as input. This data integrates multi-level cognitive reflection information of the task path, including result reflection, process reflection, and strategy reflection information. Feature extraction is performed on the reflection fusion data using matrix factorization or dimensionality reduction techniques (such as principal component analysis, singular value decomposition, etc.) to remove redundant information and extract core cognitive feature vectors. Next, knowledge distillation technology is used to map complex high-dimensional cognitive feature vectors into low-dimensional lightweight knowledge units. During knowledge distillation, a teacher-student network framework is constructed. The high-dimensional cognitive features of the teacher network are used as a reference, and the student network minimizes the reconstruction error through a loss function to achieve feature compression. The resulting lightweight knowledge unit data has a small data volume and strong semantic expressive power. Its structure is a set of node-edge knowledge pairs, containing cognitive weight information and edge state feature parameters, providing an input foundation for the subsequent construction of an edge-state reinforcement learning model.

[0152] Step S42: Construct an edge-state reinforcement learning model based on lightweight knowledge unit data;

[0153] In this embodiment of the invention, a side-state reinforcement learning model is constructed using the lightweight knowledge unit data output in step S41. This model takes the edge state features in the knowledge unit as state input, defines the action space as a set of edge weight adjustment policies, initializes the reinforcement learning environment, and the state space corresponds to the current attention weights of the edges and their cognitive distillation features. The action space is preset to a series of weight increase and decrease operations. A value function or policy function is used to evaluate the value of state-action pairs and calculate the expected reward. The reward function is designed based on the edge weight optimization objective, such as improving path propagation efficiency or enhancing propagation accuracy. During model training, temporal difference (TD learning) or Monte Carlo methods are used for policy updates to ensure that the model optimizes edge state performance by iteratively adjusting edge weight policies. The side-state reinforcement learning model forms a comprehensive system including state transition rules, policy functions, and value estimation modules. Model parameters and state policy data are used as outputs, ready for path iterative optimization.

[0154] Step S43: Perform iterative optimization of the higher-order propagation path based on the edge state reinforcement learning model to obtain the higher-order propagation path iterative optimization data.

[0155] In this embodiment of the invention, based on the edge-state reinforcement learning model trained in step S42, iterative optimization of high-order propagation paths is performed. The input is the current propagation path structure and corresponding edge weight states. Through the policy function of the reinforcement learning model, weight adjustment decisions are made for each edge state, and the adjusted edge weight distribution is calculated. Then, the propagation capability of the path is recalculated based on the new weight distribution, including path connectivity, propagation efficiency, and node coverage indicators. Based on the calculation results, the degree of path performance improvement is evaluated. If the path performance does not reach a preset threshold, the next round of weight adjustment and path performance evaluation continues, completing the iterative optimization. During the iteration process, the path weight configuration and propagation indicators of each round are saved, forming a historical data sequence for path optimization. After multiple iterations, high-order propagation path iterative optimization data is output, including the optimized edge weight distribution, path structure information, and performance indicators, for subsequent execution and evaluation of agent task decision paths.

[0156] Of particular importance is that step S41 includes the following steps:

[0157] Step S411: Perform cognitive granularity segmentation processing on the agent reflection fusion data to obtain cognitive granularity segmentation data;

[0158] In this embodiment of the invention, based on the agent reflection fusion data obtained in the preceding steps, result reflection data, process reflection data, and strategy reflection data are extracted to construct three sets of cognitive information units. The above reflection information is processed hierarchically using cognitive granularity calibration rules, with granularity division dimensions including the task objective abstraction level, path structure complexity, and strategy control decision chain length. Using a task decomposition degree function, the execution behavior of task nodes is layer-by-layered to construct an initial set of cognitive fragments composed of a quadruple of "task category - control method - edge trigger type - reflection tag". Then, an information redundancy elimination algorithm filters out redundant path edges and highly overlapping reflection factors. Furthermore, combining the edge weight fluctuation amplitude and node state perturbation frequency, a threshold is set to cluster and merge the cognitive units. Cognitive granularity division data is output according to three dimensions: "high-granularity target deviation", "medium-granularity process misalignment", and "low-granularity control conflict", for use in the next stage of cognitive structure chain construction.

[0159] Step S412: Construct the cognitive reflection chain structure based on the cognitive granularity segmentation data to obtain the cognitive reflection chain structure data;

[0160] In this embodiment of the invention, based on the acquired cognitive granularity segmentation data, a graph structure mapping construction method is used to topologically connect the logical dependencies between various granularity reflection units according to time sequence, path segment structure, and execution state transition chain. High-granularity task deviation nodes are set as the starting nodes of the graph structure. Then, medium-granularity process operation cognitive segments are connected layer by layer to construct relay sub-paths, and the degree of causal influence on the strategy result is determined by a weight threshold function. For low-granularity segments of strategy conflict, a dynamic path association discrimination mechanism is used to connect them to the end of the process chain, forming a complete multi-level cognitive reflection chain graph. To ensure structural connectivity and directional consistency, the constructed chain graph structure undergoes direction verification, duplicate path merging, and isolated node removal processing to obtain cognitive reflection chain structure data with causal continuity, strategy relevance, and path integrity, serving as the structural basis for subsequent matching and distillation operations.

[0161] Step S413: Perform cross-matching of related factors based on the cognitive reflection chain structure data to obtain cross-matching data of related factors;

[0162] In this embodiment of the invention, based on the cognitive reflection chain structure data, the factor information carried by each node (i.e., cognitive segment) in the chain graph is cross-compared to extract key common factors. A factor matching rule base is set, and the matching fields include: task type code, path node number, strategy scope, semantic jump identifier, etc. Through field pointer cross-mapping technology, node pairs with an overlap rate greater than a set threshold (e.g., 85%) in the path of the chain graph are marked, and a factor matching mapping table is established to output the factor cross-degree matrix. Further, through factor weight similarity clustering, highly coupled cognitive node segments are compressed into logically equivalent groups, redundant factor cross nodes are removed, and a simplified key related factor set is constructed. The output related factor cross-matching data is represented by a four-element structure of "related factor group ID-factor category-mapping range-strategy conflict label", providing a basis for factor compression in the cognitive distillation stage.

[0163] Step S414: Perform cognitive distillation based on the cross-matching data of the correlation factors to obtain lightweight knowledge unit data.

[0164] In this embodiment of the invention, based on the cross-matching data of related factors, a factor coupling degree weighting mechanism is used to screen out key cognitive factor groups that appear frequently, have a wide path breadth and deep influence, and have significant semantic conflict risks. A lightweight cognitive compression rule tree is constructed to compress fragment nodes with high similarity in causal chain propagation paths and consistent policy influence directions in the related factor groups. Each compression unit includes: a set of influencing task nodes, a reflection label merging rule, and a policy correction direction vector. During distillation compression, the high-weight nodes in the reflection chain graph are used as anchors to pass policy influence factors downstream, and paths without new contributions are pruned; lightweight knowledge unit data with compact structure, clear expression, and reusable logical rules is generated. The obtained lightweight knowledge unit data is stored in the form of a four-element structure of "knowledge unit ID - influencing policy - structural scope - path edge constraint", which serves as the core input data for subsequent edge state reinforcement learning modeling and as the direct basis for generating "interpretable path deviation correction strategies".

[0165] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for modeling high-order relationships between agents based on edge attention weights, characterized in that, Includes the following steps: Step S1: Obtain multimodal interaction data of the agents; construct an initial relationship graph of the agents based on the multimodal interaction data; Determine the edge attribute features of the agent based on the initial relation graph of the agent; Step S2: Perform initial attention weight allocation based on the agent's edge attribute features to obtain initial edge attention weight data, which includes: Edge semantic encoding is performed based on the edge attribute features of the agent to obtain the agent edge semantic encoding data; Extracting the frequency and intensity of agent-side interactions based on agent-side semantic encoding data; Extracting temporal features of agent-side interactions based on agent-side semantic encoding data; The semantic vector representation data of the agent's side interaction is determined based on the temporal characteristics of the agent's side interaction and the frequency and intensity of the agent's side interaction. Based on the semantic vector representation of the agent's edge, the distribution of the agent's node structure in the neighborhood can be identified. The edge adjacency structure feature data is determined based on the agent edge semantic vector representation data and the agent node structure neighborhood distribution. Edge relationship context fusion modeling is performed based on edge adjacency structure feature data to obtain edge relationship context representation data; The initial attention weights are assigned based on the edge relation context representation data to obtain the initial edge attention weight data. Construct an edge weight enhancement graph based on edge attention weight data; perform multi-hop higher-order relation propagation processing on the edge weight enhancement graph data to obtain multi-hop higher-order relation propagation data; perform agent node representation enhancement processing on the multi-hop higher-order relation propagation data to obtain agent node representation enhancement data. Step S3: Construct a task decision path based on agent node representation enhancement data and multi-hop high-order relationship propagation data to obtain the agent task decision path; perform a three-level cognitive reflection operation on the agent task decision path to obtain three-level cognitive reflection data; and fuse the three-level cognitive reflection data to obtain agent reflection fusion data. Step S4: Construct a side-state reinforcement learning model based on the agent's reflection and fusion data; perform iterative optimization of the higher-order propagation path based on the side-state reinforcement learning model to obtain the higher-order propagation path iterative optimization data.

2. The method for modeling high-order relationships of agents based on edge attention weights according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Acquire multimodal interaction data of the intelligent agent; Step S12: Perform structural feature fusion processing based on the agent's multimodal interaction data to obtain agent structural feature fusion data; Step S13: Construct an initial relational graph of the agents based on the fusion data of the agents' multimodal interaction data and the agents' structural features; Step S14: Determine the edge attribute features of the agent based on the initial relation graph of the agent.

3. The method for modeling high-order relationships of agents based on edge attention weights according to claim 2, characterized in that, Step S13 includes the following steps: Step S131: Perform interaction semantic extraction processing based on the agent's multimodal interaction data to obtain interaction semantic feature data; Step S132: Perform node functional attribute identification processing based on the fusion data of the agent's structural features to obtain node functional attribute data; Step S133: Perform agent node association analysis based on interaction semantic feature data and node functional attribute data to obtain agent node association relationship data; Step S134: Construct node relationship connection structure data based on agent node association data; Step S135: Perform relation edge generation processing based on node relationship connection structure data and interaction semantic feature data to obtain structural relation edge data; Step S136: Perform node-edge information fusion processing based on node functional attribute data and structural relationship edge data to obtain relation graph structure data; Step S137: Construct the initial relation graph of the agent based on the structural relation edge data and the relation graph structure data.

4. The method for modeling high-order relationships of agents based on edge attention weights according to claim 1, characterized in that, In step S2, multi-hop higher-order relation propagation processing is performed based on the edge weight enhancement graph data to obtain multi-hop higher-order relation propagation data, including: Extract multi-order adjacency information of agent nodes from edge-weighted graph data; Jump distance marking is performed based on the multi-level adjacency relationship information of the agent nodes to obtain jump distance marked edge structure data; Candidate data for multi-hop paths of nodes are determined by using hop distance marked edge structure data and multi-order adjacency relationship information of agent nodes; Calculate the node behavior similarity factor based on the candidate data of multi-hop paths of nodes; Multi-hop propagation aggregation data is obtained by performing multi-hop propagation aggregation processing based on node behavior similarity factors and node multi-hop path candidate data; Cross-node relationship coupling calculations are performed based on multi-hop propagation aggregated data to obtain cross-node relationship coupling data; Determine higher-order semantic extended data based on cross-node relationship coupling data; Multi-hop higher-order relation propagation data is obtained by performing multi-hop higher-order relation propagation processing based on higher-order semantic extended data and cross-node relation coupling data.

5. The method for modeling high-order relationships of agents based on edge attention weights according to claim 1, characterized in that, In step S2, agent node representation enhancement processing is performed based on the multi-hop higher-order relation propagation data, resulting in agent node representation enhancement data including: Extract upstream and downstream interaction path information of agent nodes from multi-hop high-order relationship propagation data; Statistical distribution data of node interaction frequency is compiled based on the upstream and downstream interaction path information of intelligent agent nodes; High-frequency path reduction processing is performed based on the interaction frequency distribution data to obtain high-frequency path structure reduction data; Node feature fusion processing is performed based on high-frequency path structure reduction data and upstream and downstream interaction path information of agent nodes to obtain node feature fusion data. Node state stability is evaluated based on the fused data of node features to obtain node state stability data. Based on node state stability data, abnormal node feature dimensions are filtered to obtain abnormal node feature data. High-order semantic correction is performed based on node anomaly feature data to obtain high-order semantic correction node data. The agent node representation enhancement data is obtained by performing intelligent agent node representation enhancement processing based on high-order semantic correction node data and node feature fusion data.

6. The method for modeling high-order relationships of agents based on edge attention weights according to claim 1, characterized in that, Step S3 includes the following steps: Step S31: Construct the task decision path based on the agent node representation enhancement data and multi-hop high-order relationship propagation data to obtain the agent task decision path; Step S32: Based on the agent's task decision path, conduct an agent path feasibility assessment to obtain agent path feasibility data; Step S33: Perform a three-level cognitive reflection operation on the agent's path feasibility data to obtain three-level cognitive reflection data, including result reflection data, process reflection data, and strategy reflection data; Step S34: Perform fusion processing on the three-level cognitive reflection data to obtain the agent reflection fusion data.

7. The method for modeling high-order relationships of agents based on edge attention weights according to claim 6, characterized in that, Step S31 includes the following steps: Step S311: Extract node task perception feature data based on the agent node representation enhancement data; Step S312: Determine the path node boundary information data based on the node task perception feature data and the multi-hop high-order relationship propagation data; Step S313: Calculate path connectivity based on path node boundary information data and multi-hop higher-order relationship propagation data to obtain path connectivity data; Step S314: Perform path optimization and sorting processing based on path connectivity data to generate task path optimization and sorting data; Step S315: Construct task decision paths based on task path optimization and sorting data to obtain the agent's task decision paths.

8. The method for modeling high-order relationships of agents based on edge attention weights according to claim 6, characterized in that, Step S32 includes the following steps: Step S321: Statistically analyze the load status of agent path nodes based on the agent's task decision path; Step S322: Calculate the average response time of the agent nodes based on the load status of the agent path nodes; Step S323: Detect the load fluctuation rate of the agent node based on the load status of the agent path node; Step S324: Evaluate the response stability of the agent node based on the agent node load volatility and the agent node average response time; Step S325: Determine the collaborative efficiency of the agent operation based on the stability of the agent node response; Step S326: Based on the agent's operational coordination efficiency and the agent node response stability, conduct an agent task path feasibility assessment to obtain agent path feasibility data.

9. The method for modeling high-order relationships of agents based on edge attention weights according to claim 1, characterized in that, Step S4 includes the following steps: Step S41: Perform cognitive distillation on the agent's reflection and fusion data to obtain lightweight knowledge unit data; Step S42: Construct an edge-state reinforcement learning model based on lightweight knowledge unit data; Step S43: Perform iterative optimization of the higher-order propagation path based on the edge state reinforcement learning model to obtain the higher-order propagation path iterative optimization data.

Images