Cognitive behavior evolution prediction and path explainability modeling method, device and system

Abstract

This invention discloses a method, apparatus, and system for predicting cognitive behavior evolution and modeling path interpretability, belonging to the field of consumer behavior prediction. The method includes the following steps: S1, inputting features of structured behavioral chain data into a neural network for training, and using the neural network to model the temporal evolution of user behavior to predict the dynamic evolution of cognitive states; S2, based on the user's current cognitive state, historical behavioral chains, and predicted evolutionary direction, deducing possible evolutionary paths on the cognitive chain graph, and modeling and predicting potential decision trajectories; S3, based on the future path deduction results, combined with cognitive motivation analysis and goal relevance assessment, generating a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation. This invention can achieve high-precision prediction and transparent output from the user's current attention map to future behavioral motivations.

Description

Technical Field

[0001] This invention relates to the field of consumer behavior prediction, and more specifically, to a method, apparatus, and system for predicting cognitive behavior evolution and modeling path interpretability. Background Technology

[0002] With the development of big data and artificial intelligence technologies, user behavior analysis has been widely applied in recommendation systems, travel navigation, and operational decision-making. However, the current field of user behavior modeling lacks interpretable behavior modeling mechanisms. Existing models struggle to integrate temporal and cognitive characteristics of behavior and lack modeling methods that consider the evolution of user attention pathways, resulting in limited interpretability and generalization capabilities.

[0003] In user behavior modeling schemes based on graph neural networks, existing solutions typically rely solely on network structure for information propagation, failing to effectively integrate the cognitive evolutionary drivers and semantic features inherent in the user behavior chain. This results in the message propagation mechanism being unable to dynamically adjust the influence of key nodes and struggles to guarantee the cognitive consistency and temporal rationality of node state updates. Furthermore, existing models often lack explicit constraints on causal relationships, leading to weak interpretability of prediction results and difficulty in extracting logically explainable strategies from the deduction path. Ultimately, this limits the practicality and transparency of the models in behavioral understanding and intelligent decision support. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a method, device and system for predicting cognitive behavior evolution and modeling path interpretability, which can achieve high-precision prediction and transparent output from the user's current attention map to the motivation for future behavior.

[0005] The objective of this invention is achieved through the following solution: A method for predicting cognitive behavior evolution and modeling path interpretability includes the following steps: S1. The features of the structured behavioral chain data are input into the neural network for training. The neural network is used to model the temporal evolution of user behavior to predict the dynamic evolution of cognitive state. The neural network includes a time-series extended graph neural network (T-GNN). When it is a T-GNN, the features of the structured behavioral chain data are used as the features of graph nodes and edges. An externally interpretable weight combining the cognitive dwell characteristics and time decay characteristics of nodes is introduced to guide the aggregation process of nodes to neighbor nodes. The information propagation weights of self-loops and neighbor edges are defined using a weight function based on dwell strength and time decay. A causal consistency regularization mechanism is introduced during training to constrain the changing trends of node self-loop weights and neighbor weights, so as to guide the model to learn predictive behaviors that conform to the user's real cognitive evolution path. S2, based on the user's current cognitive state, historical behavior links and predicted evolution direction, deduces the possible evolution path on the cognitive link graph, and models and predicts the potential decision trajectory; S3, based on the results of future path projection, combined with cognitive motivation analysis and goal relevance assessment, generates a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation.

[0006] Furthermore, in step S1, the introduction of externally interpretable weights that combine node cognitive dwell characteristics and time decay characteristics to guide the node's aggregation process of neighbor node messages specifically includes the message aggregation phase process: The state vectors of all predecessor nodes j pointing to node i in the previous time step According to attention weight After weighting and summarizing, the summary message is obtained. The specific message aggregation formula is as follows: ; in, This represents the set of all predecessor nodes that point to i in the graph; This represents the attention shift weight from node j to i, reflecting the strength of the user's attention shifting from j to i; This represents a learnable projection matrix used to transform the feature vectors of neighboring nodes. Mapped to a message space of the same dimension as the target node; The feature vector of node j at time t-1; attention weights The definition is as follows: ; ratio Indicates the measurement of target node n j The dwell strength relative to the current node n i Differences in residence intensity; Time decay factor It is used to weaken the intensity of attention shift over long time intervals, representing the natural decay of attention over time, and simulating the natural decay process of user attention over time; represents the time interval between the behaviors of nodes ni and nj; μ is the attention decay coefficient over time, used to control the decay rate, which can be adaptively adjusted according to the user's behavior rhythm or system response time; δ is the normalization factor, used to unify the weight scale of all links.

[0007] Furthermore, in step S1, the introduction of externally interpretable weights that combine node cognitive dwell characteristics and time decay characteristics to guide the node's aggregation process of neighbor node messages specifically includes the node state update phase process: When a node is updated, its previous time-to-time characteristics are included. News gathered from neighbors Linear combination, then through activation function A nonlinear mapping is used to obtain the new hidden state of node i at time t. The specific node state update formula is as follows: ; in, This represents a self-circulating projection matrix that evolves co-evolves with neighbor aggregation weights and is subject to causal consistency constraints. Represents a non-linear activation function; This represents the hidden state vector of node i at time t, which will be used for prediction at the next time step or for downstream tasks.

[0008] Further, in step S1, the step of defining the information propagation weights of self-loops and neighbor edges using a weighting function based on dwell strength and time decay specifically includes the following sub-steps: When node i receives a message from its neighbor node j, the weight function is designed as follows: ; Where Res(i) represents the dwell intensity characteristic of node i; Δt ij This represents the time difference between node i and node j; μ is the time decay control parameter; for a self-loop edge, i.e., when i=j, the time interval is zero, and the formula naturally degenerates to: ; The weighting function is designed to ensure that self-looping edges also carry the residence strength characteristics of cognitive node i rather than simple structural connections, and also to make the propagation strength of neighbor messages dynamically affected by the cognitive evolution process.

[0009] Furthermore, in step S1, a causal consistency regularization mechanism is introduced during training. This mechanism constrains the changing trends of node self-loop weights and neighbor weights to guide the model in learning predictive behaviors that align with the user's actual cognitive evolution path. Specifically, this includes the following sub-steps: A causal consistency correction term is introduced into the total loss function, which, together with the basic task loss, constitutes the total loss function for joint optimization. This constraint regularizes the evolution trend of the node's self-loop weights and neighbor weights, guiding the model to learn transition paths that conform to causal priors. The specific definition of the total loss function is as follows: ; Among them, L task Loss due to basic task; L causal β represents the causal consistency loss, used to measure the consistency between the predicted path and the expected causal evolution trend; L is the weighting coefficient of the causal consistency constraint.causal Defined as the deviation between the node state change and the causal prior path, it is expressed as: ; in, Represents a pair of nodes on the predicted path; h i This represents the actual state of node i. This state is the true value obtained from user behavior data extracted from the input data, used to infer the causal relationships within the link; This represents a causal inference function. The state h of node i i After making a prediction, the predicted state of node j is obtained; this state is deduced through the causal relationship network, reflecting the possible evolution path from node i to node j. This indicates that the state h of node i is inferred based on causal priors. i The causal mapping function of the predicted evolution direction feature of its child node j is derived. This mapping is implemented by affine transformation, multilayer perceptron (MLP) or neural network module based on conditional generation mechanism. L is the causal credibility weight coefficient for edge (i,j), used to characterize the importance of link (i,j). This weight is set based on the attention distribution of the link in the original cognitive link graph, historical frequency statistics, or rules based on domain priors, thereby dynamically adjusting the contribution of different links to the overall causal consistency optimization objective; when the model predicts a path that conforms to a predetermined causal evolution trend, L... causal The loss is relatively small, but it increases when non-causal jumps or anomalous transitions occur in the prediction, thus effectively constraining the training process.

[0010] Furthermore, the features of the structured behavioral chain data specifically include node dwell intensity, attention shift weight, and semantic label information.

[0011] Furthermore, in step S2, the step of deducing the possible evolutionary path of the user on the cognitive link graph based on the user's current cognitive state, historical behavioral links, and inferred evolutionary direction, and modeling and predicting potential decision trajectories, specifically includes the following sub-steps: S21, based on the current node state h output by the cognitive state evolution prediction model in step S1. i Based on the inferred evolutionary characteristics and combined with the historical link information in the original cognitive link diagram, the downstream node set {j} is inferred and filtered to generate a local link structure. S22, based on the partial link reconstruction, conduct multi-step path deduction to predict the medium- and long-term trajectory of the possible evolution of the user's cognitive link; S23, After the deduction is completed, the future paths are selected based on their rationality, relevance to the objective, and consistency with the context, and the selected set of future paths P is retained. future.

[0012] Furthermore, in step S21, during the link restoration process, the following mechanism is introduced: Evolutionary direction matching mechanism: based on h i The feature vectors and candidate child nodes h in the original link j Similarity is evaluated based on the feature orientation, and cosine similarity, Euclidean distance, or matching score based on attention mechanism is used for screening. Causal consistency constraint mechanism: combining causal consistency loss L causal The evolution direction of the node state in the restored link must conform to the causal deduction logic in the original link diagram; State succession prediction mechanism: For the selected candidate node j, its potential evolutionary state is predicted by the state succession predictor. This is used in subsequent path deduction and score calculation stages; the following link reconstruction loss function is designed: ; in, h represents the set of real edges in the original cognitive link graph. j Let j be the true state vector of node j. The predicted state λ is obtained based on the reduction and deduction mechanism. ij The importance weight coefficient for link (i,j) is set based on link frequency, attention weight, or empirical rules. By minimizing the link reconstruction error, the model can improve the accuracy of reconstructing existing cognitive paths and provide a reliable foundation for subsequent future inference stages.

[0013] Furthermore, in step S22, the deduction process includes the following sub-processes: Path expansion strategy: Use breadth-first search (BFS) or heuristic depth-first search (DFS) to expand candidate links sequentially from the current node, and limit the maximum number of iterations L. max With branch width B max To avoid an explosive increase in the number of deduction paths; Evolutionary confidence accumulation mechanism: An evolutionary confidence score is accumulated for each inference path. The confidence score is composed of node state transition matching degree and link causality confidence γ. ij The calculation is weighted by the reasonableness of state continuity, and those with a confidence level below the threshold θ are filtered out. path The path; To quantify the overall rationality and credibility of each deduction path's evolution, the path deduction scoring function is defined as follows: ; Among them, h i h jLet [hi;hj] represent the state vectors of node i and node j respectively, and [hi;hj] be the concatenation of the two; W and b are the learnable weight matrix and bias term, respectively; σ( ) is the activation function used to output the rationality score of the evolution from node i to node j; based on this score, multi-branch prediction of cognitive paths, rational path ranking and inference of potential motivations are realized, thereby supporting high-precision prediction of users' future behavior in complex situations.

[0014] Furthermore, in step S22, the deduction process also includes the following sub-processes: For new cognitive stages that are missing, skipped, or not yet covered, potential intermediate nodes are generated based on state generation models to complete the deduction path and improve the integrity and continuity of the link.

[0015] Furthermore, in step S23, the screening criteria include stage rationality verification, target consistency verification, and context dynamic adaptation mechanism; The stage rationality verification checks whether the order of cognitive stages corresponding to the internal nodes of each deduction path conforms to the general logic of cognitive evolution. The goal consistency check evaluates the goal orientation consistency of each derivation path by calculating the semantic relevance score between its terminal node state and the initial goal state; the goal relevance score function is defined as follows: ; in, h represents the state vector of the terminal node of path p. goal This represents the target state vector corresponding to the starting point of the deduction; the function uses cosine similarity, and the higher the score, the more the path matches the user's original goal; a target relevance threshold θ is set during the filtering process. target Paths below the threshold are removed; The context-based dynamic adaptation mechanism dynamically adjusts the future path prediction results based on real-time changes in the external environment, enabling continuous updates and iterative optimization of the deduced path. When a new key behavioral event is detected, the future path deduction and filtering are re-executed based on the current node state, dynamically correcting P. future A set of data is used to ensure the real-time nature and relevance of the prediction results.

[0016] Furthermore, step S23 includes the following sub-steps: Based on the confidence scores at each step of the comprehensive deduction process, the overall confidence score function for the path is defined as follows, which is used to further filter and rank the set of future paths: ; Where (i,j) represents the continuous node transitions in the derivation path p, (i,j)∈p represents each pair of continuous nodes in the path, and Spredict (j|i) is the derivation scoring function; a path confidence threshold θ is set during screening. path High-confidence paths with scores above the threshold are retained.

[0017] Further, in step S3, based on the future path deduction results, combined with cognitive motivation analysis and goal relevance assessment, a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation is generated, specifically including the following sub-steps: S31, Cognitive Motivation Extraction and Intent Attribution Analysis, specifically includes path motivation extraction and intent clustering and induction; the path motivation extraction is based on the state vectors and evolution directions of each node in the deduced path, analyzes potential cognitive motivations, and performs attribution modeling by combining node semantic labels, stage features, and contextual information; the intent clustering and induction clusters the deduced paths with similar motivational features, summarizes the changing trends of users' potential goals, forms motivational labels, and supports subsequent strategy generation; S32, Strategy candidate generation and optimization, specifically includes strategy rule base matching and a multi-dimensional optimization mechanism; the strategy rule base matching generates a corresponding set of strategy candidates based on the matching of motivation tags with the strategy rule base; the multi-dimensional optimization mechanism comprehensively scores the candidate strategies based on target relevance scores, user acceptance predictions, and path rationality indicators, and selects the optimal strategy set; the strategy comprehensive scoring function is defined as follows: ; Among them, S target (s) represents the consistency score between the strategy and the initial goal, S accept (s) is the predicted probability that the strategy will be accepted or responded to by the user, S path (s) represents the overall rationality score of the inference path corresponding to the strategy, where α, β, and γ are adjustable weight parameters that are set according to the actual application scenario. S33, Policy interpretability expression and output, specifically including motivational policy link generation and dynamic update mechanism; the motivational policy link generation is to generate a corresponding motivational explanation link for each output policy; the dynamic update mechanism dynamically adjusts the policy library and generation logic based on the user's subsequent actual behavior feedback.

[0018] Furthermore, step S33 also includes the following sub-steps: Generate user-understandable policy hints or illustrations based on template filling, retrieval generation, or lightweight NLG models.

[0019] Furthermore, it also includes the following steps: Configure one or more of the following enhancement components: Multi-scenario adaptation component: By recoding and mapping the semantic labels of nodes, behavioral stage attributes and inference path rules of the cognitive link graph, the model and inference logic can be migrated and adapted to different business domains; Personalized Explanation Generation Component: During the strategy generation and explanation output phase, personalized explanation content is generated for different user profiles. Multimodal Enhancement Component for Strategy Output: Presents the inference results and strategy explanations in a multimodal manner, enhancing users' perception and understanding of future cognitive evolution trends.

[0020] Furthermore, the implementation of the multi-scenario adaptation component specifically includes: introducing a scenario adaptation layer, automatically redefining node categories, path stages, and inference strategy parameters based on scenario tags; and supporting preset adaptation templates and online fine-tuning mechanisms to quickly adapt to new business requirements. The implementation of the personalized explanation generation component specifically includes: introducing a personalized explanation generator, combining user feature vectors, adaptively adjusting motivation extraction, strategy text templates, and language style, and supporting online personalized generation using lightweight style modeling; The implementation of the strategy output multimodal enhancement component specifically includes: generating a dynamic visualization based on the existing cognitive link graph structure and inference path; and forming a richly illustrated multimodal recommendation interface by combining natural language output.

[0021] A cognitive behavior evolution prediction and path interpretability modeling apparatus includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method described in any of the preceding methods.

[0022] A cognitive behavior evolution prediction and path interpretability modeling system includes the cognitive behavior evolution prediction and path interpretability modeling device as described above.

[0023] The beneficial effects of this invention include: (1) In this invention, the message propagation strength of each edge is not only determined by the structural connection, but also reflects the cognitive evolution motivation in the user behavior link, so that the node state update process can naturally maintain cognitive consistency and temporal rationality. Through this mechanism, the model can dynamically adjust the message flow direction and strength when transmitting information, emphasize the resident nodes in important cognitive stages, and reasonably attenuate the influence of non-critical stages or outdated nodes, laying the foundation for subsequent causal consistency modeling.

[0024] (2) The model of the present invention can naturally integrate the semantic features of the cognitive link when disseminating information, thereby improving the model's ability to capture the evolution law of the user link.

[0025] (3) While ensuring the basic prediction accuracy of the model, the present invention further promotes the consistency between the reasoning results of the cognitive link and the actual causal relationship, and enhances the interpretability and practical value of the model.

[0026] (4) In this invention, the TG NN can achieve collaborative updating of self-loop information and neighbor information in cognitive semantics and causal logic, which not only improves the expressive power of node representation, but also significantly enhances the interpretability and causal rationality of prediction results.

[0027] (5) In this invention, by analyzing the structural features, evolutionary motivations and target matching degree of the deduction path, we can extract strategy prompts that are interpretable and guiding, which can improve the transparency and adaptability of the system in intelligent decision support.

[0028] (6) In this invention, by analyzing the structural features, evolutionary motivations and target matching degree of the deduction path, strategy prompts with interpretability and guidance can be extracted, thereby improving the transparency and adaptability of the system in intelligent decision support.

[0029] (7) In this invention, the path-driven interpretable strategy generator can realize the automatic conversion of the deduction path to the intervention strategy, while ensuring the clarity and logic of the interpretation chain, providing solid support for user behavior prediction, recommendation optimization and decision support.

[0030] (8) Other technical effects of the present invention will be described in detail in the specific embodiments. Attached Figure Description

[0031] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0032] Figure 1 This is a flowchart illustrating the overall steps of the method in an embodiment of the present invention. Detailed Implementation

[0033] All features disclosed in all embodiments of this specification, or steps in all methods or processes implied in the disclosure, may be combined and / or extended or replaced in any way, except for mutually exclusive features and / or steps.

[0034] In a preferred embodiment, the present invention proposes a method for predicting cognitive behavior evolution and modeling path interpretability, comprising the following steps: S1, Constructing a Cognitive State Evolution Prediction Model: The features of structured behavioral chain data are input into a neural network for training. The model uses the neural network to model the temporal evolution of user behavior to predict the dynamic evolution of cognitive states. The neural network includes a time-series extended graph neural network (T-GNN). When using T-GNN, the features of structured behavioral chain data are used as the features of graph nodes and edges. An externally interpretable weight combining node cognitive dwell characteristics and time decay characteristics is introduced to guide the aggregation process of nodes on neighbor nodes. The information propagation weights of self-loops and neighbor edges are defined using a weight function based on dwell strength and time decay. A causal consistency regularization mechanism is introduced during training to constrain the changing trends of node self-loop weights and neighbor weights, thereby guiding the model to learn predictive behaviors that conform to the user's actual cognitive evolution path. S2, Cognitive Link Reconstruction and Future Path Deduction: Based on the user's current cognitive state, historical behavioral links, and predicted evolution direction, the possible evolution path of the user on the cognitive link diagram is deduced, and potential decision trajectories are modeled and predicted. S3, Path-Driven Explainable Strategy Generation: Based on the results of future path projection, combined with cognitive motivation analysis and goal relevance assessment, a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy explanation is generated.

[0035] It should be noted that the above-described embodiments can achieve high-precision prediction and transparent output of the user's current attention graph and future behavioral motivations. Specifically, by performing behavioral stage identification, semantic unification, and timeline mapping on the original multi-source behavioral data, a semantically consistent and temporally continuous user behavior chain can be constructed. For example, the user behavior chain uses fields such as "user ID – behavioral stage – timestamp – behavioral tag" as its core structure, thus providing structured input support for cognitive path modeling and attention link analysis. The various neural networks mentioned in the above embodiments (such as LSTM, GRU, GCN, GAT, etc.) can be used as implementation methods to model the temporal evolution of user behavior. While temporal modeling neural networks (such as time-series extended graph neural networks T-GNN) are particularly suitable for the cognitive link reconstruction and path inference tasks of this invention, other solutions, such as conventional graph neural network (GNN) methods or traditional machine learning models, can also be used as the basis for implementing this module. The specific method used can be flexibly selected according to data characteristics, task requirements, and model accuracy requirements.

[0036] In an optional implementation, step S1, which introduces an externally interpretable weight that combines node cognitive dwell characteristics and time decay characteristics to guide the node's aggregation process of neighbor node messages, specifically includes a message aggregation phase process and a node state update phase process.

[0037] Furthermore, the message aggregation phase combines the state vectors of all predecessor nodes j pointing to node i from the previous time step. According to a certain attention transfer weight (e.g., attention transfer weight) After weighting and summarizing, the summary message is obtained. This allows each node to dynamically obtain information from its neighbors based on its own "attention intensity" with those of its neighbors.

[0038] The message aggregation formula is: ; in, This represents the set of all predecessor nodes that point to i in the graph; This represents the attention shift weight from node j to i, reflecting the strength of the user's attention shifting from j to i; This represents a learnable projection matrix used to transform the feature vectors of neighboring nodes. Mapped to a message space of the same dimension as the target node; This represents the feature vector of node j at time t-1, which includes dwell strength, semantic embedding, contextual information, etc. Optional attention transfer weights It can be defined as follows: ; ratio Indicates the measurement of target node n j The dwell strength relative to the current node n i Differences in residence intensity; time decay factor It is used to weaken the intensity of attention shift over long time intervals, representing the natural decay of attention over time, and simulating the natural decay process of user attention over time; represents the time interval between the behaviors of nodes ni and nj; μ is the attention decay coefficient over time, used to control the decay rate, which can be adaptively adjusted according to the user's behavior rhythm or system response time; δ is the normalization factor, used to unify the weight scale of all links.

[0039] It should be noted that the above embodiments propose an attention-weight-guided message flow aggregation mechanism. In traditional graph neural networks (such as GCN and GAT), node feature updates are usually based on feature aggregation from neighboring nodes. Weight design is either fixed or entirely dependent on internal model learning, lacking guidance from external cognitive priors. This results in limited expressiveness and insufficient interpretability of reasoning results in user behavior reasoning tasks. To address these issues, the above embodiments of this invention propose an attention-weight-guided message flow mechanism in temporal graph neural networks (T-GNN).

[0040] Furthermore, the node update phase incorporates the node's features from the previous time step. News gathered from neighbors Linear combination, then through activation function A nonlinear mapping is used to obtain the new hidden state of node i at time t. By preserving the "self-looping" branch The model can inherit its own historical information and also absorb external influences from its neighbors. Activation function (For example, ReLU or tanh can be used) to give the model stronger expressive power and be able to capture nonlinear relationships.

[0041] Node state update formula: ; in, The self-circular projection matrix is ​​a trainable weight matrix that "maps" or "transforms" the historical state of a node to change its own features. This represents a nonlinear activation function, used to introduce nonlinearity and maintain gradient stability; This represents the hidden state vector of node i at time t, which will be used for prediction at the next time step or for downstream tasks.

[0042] In other embodiments, step S1, which defines the information propagation weights of self-loops and neighbor edges using a weighting function based on dwell strength and time decay, specifically includes the following sub-steps: A unified weighting function based on dwell strength and time decay is adopted to define the information propagation weights for self-loops and neighbor edges. Specifically, when node i receives a message from neighbor node j, the weighting function is uniformly designed as follows: ; Where Res(i) represents the dwell intensity characteristic of node i; Δt ij This represents the time difference between node i and node j; μ is the time decay control parameter; for a self-loop edge, i.e., when i=j, the time interval is zero, and the formula naturally degenerates to: ; This weighting function is designed to ensure that self-looping edges also carry the residence strength characteristics of cognitive node i, rather than simple structural connections, and also to dynamically influence the propagation strength of neighbor messages through cognitive evolution. More specifically, this weighting design not only guarantees that self-looping edges also carry cognitive residence characteristics rather than simple structural connections, but also dynamically influences the propagation strength of neighbor messages through cognitive evolution. Therefore, the model can naturally integrate the semantic features of cognitive links when propagating information, thereby improving the model's ability to capture the evolutionary patterns of user links.

[0043] Furthermore, it should be noted that during information propagation, an externally interpretable weight combining node cognitive dwell characteristics and time decay properties is introduced to guide the aggregation process of messages from neighboring nodes. The message propagation strength of each edge is not only determined by structural connections but also reflects the cognitive evolutionary drivers in the user behavior chain, thus enabling the node state update process to naturally maintain cognitive consistency and temporal rationality. Through this mechanism, the model can dynamically adjust the message flow and intensity when transmitting information, emphasizing dwelling nodes in important cognitive stages and reasonably decaying the influence of non-critical stages or outdated nodes, laying the foundation for subsequent causal consistency modeling.

[0044] Furthermore, to further enhance the model's explanatory power and causal reasoning ability in prediction tasks, this invention introduces a causal consistency regularization mechanism during training. By constraining the changing trends of node self-loop weights and neighbor weights, the model is guided to learn predictive behaviors that conform to the user's actual cognitive evolution path, reducing non-causal jumps or anomalous migrations. Because in step S1, the causal consistency regularization mechanism is introduced during training, constraining the changing trends of node self-loop weights and neighbor weights to guide the model to learn predictive behaviors that conform to the user's actual cognitive evolution path, specifically including the following sub-steps: A causal consistency correction term is introduced into the total loss function, which, together with the basic task loss, constitutes the total loss function for joint optimization. This constraint regularizes the evolution trend of the node's self-loop weights and neighbor weights, guiding the model to learn transition paths that conform to causal priors. The specific definition of the total loss function is as follows: ; Among them, L task The loss is based on the fundamental task (such as cross-entropy for node classification or mean squared error for node state regression); L causal β represents the causal consistency loss, used to measure the consistency between the predicted path and the expected causal evolution trend; L is the weighting coefficient of the causal consistency constraint. causal Defined as the deviation between the node state change and the causal prior path, it is expressed as: ; in, Represents a pair of nodes on the predicted path; h i This represents the actual state of node i. This state is the true value obtained from user behavior data (such as current cognitive state, historical behavior, etc.) extracted from the input data, and is used to infer the causal relationships within the chain. This represents a causal inference function. The state h of node i i After making a prediction, the predicted state of node j is obtained; this state is deduced through the causal relationship network, reflecting the possible evolution path from node i to node j. This indicates that the state h of node i is inferred based on causal priors. i The causal mapping function of the predicted evolution direction feature of its child node j is obtained by deduction. For example, the reasonable evolution direction of the node state can be inferred based on the stage sequence, transfer preference or association pattern in the cognitive link. This mapping is implemented by affine transformation (such as linear layer), multilayer perceptron MLP or neural network module based on conditional generation mechanism to improve the expressive flexibility and adaptability of the link deduction process at different granularities. L is the causal credibility weight coefficient for edge (i,j), used to characterize the importance of link (i,j). This weight is set based on the attention distribution of the link in the original cognitive link graph, historical frequency statistics, or rules based on domain priors, thereby dynamically adjusting the contribution of different links to the overall causal consistency optimization objective; when the model predicts a path that conforms to a predetermined causal evolution trend, L... causal While the initial loss is relatively small, it increases when non-causal jumps or anomalous shifts occur in the prediction, thus effectively constraining the training process. By introducing this loss, while ensuring the model's basic prediction accuracy, it further promotes the consistency between the cognitive link inference results and the actual causal relationship, enhancing the model's interpretability and practical value.

[0045] Furthermore, the above-described embodiment introduces a causal consistency correction term into the total loss function. This term, along with the basic task loss (such as node classification or state regression error), forms the total loss function for joint optimization, enabling a deep integration of structure awareness and causal reasoning capabilities. This constraint, by regularizing the evolution trend of the self-loop weights and neighbor weights of nodes, guides the model to learn transition paths that conform to causal priors. For example, when there is a clear stage progression in the user behavior chain (such as from intention formation to path planning to in-store behavior), the model will encourage the predicted path to remain consistent with the actual causal evolution trend during training, reducing the interference of non-causal transition paths.

[0046] Furthermore, the features of the aforementioned structured behavioral chain data specifically include node dwell strength, attention shift weights, and semantic label information. Through this approach, the TG NN of this invention can achieve collaborative updates of self-looping information and neighbor information in terms of cognitive semantics and causal logic, thereby improving the expressive power of node representations and significantly enhancing the interpretability and causal rationality of prediction results.

[0047] The above-described embodiments propose a self-looping and neighbor information collaborative update mechanism based on cognitive semantic weight definition and causal consistency correction. In traditional GCNs, node self-loops are typically introduced by adding an identity matrix to the adjacency matrix, with fixed weights primarily to meet structural connectivity requirements, without reflecting the cognitive characteristics or causal motivations of the nodes. While GAT includes the node itself in its attention weight calculation, its weights rely entirely on the model learning process, lacking external cognitive semantic or causal prior guidance, resulting in limited interpretability. The above-described embodiments propose a solution for the self-looping and neighbor information collaborative update mechanism based on cognitive semantic weight definition and causal consistency correction within the message passing mechanism of Temporal Graph Neural Networks (T-GNN).

[0048] In other implementations, based on the user's current cognitive state, historical behavioral links, and inferred evolutionary direction, possible evolutionary paths on the cognitive link graph are deduced, and potential decision trajectories are modeled and predicted. As a natural extension of the cognitive state evolution prediction model, this mainly includes three sub-processes: link reconstruction, path deduction, and feasibility screening, aiming to maintain the causal consistency, evolutionary rationality, and dynamic scalability of the link structure. Therefore, in step S2, the deduction of possible evolutionary paths on the cognitive link graph based on the user's current cognitive state, historical behavioral links, and inferred evolutionary direction, and the modeling and prediction of potential decision trajectories, specifically includes the following sub-steps: S21, Execution Link Local Reconstruction and State Continuation Modeling: Based on the current node state h output by the cognitive state evolution prediction model in step S1. i Based on the inferred evolutionary characteristics and combined with historical link information from the original cognitive link graph, the possible downstream node set {j} is inferred and filtered to generate a local link structure; the following mechanism is introduced during the link reconstruction process: Evolutionary direction matching mechanism: based on h i The feature vectors and candidate child nodes h in the original link j Similarity can be evaluated based on the feature orientation, and cosine similarity, Euclidean distance, or matching score based on attention mechanism can be used for screening. Causal consistency constraint mechanism: combining causal consistency loss L causal The evolution direction of the node state in the restored link must conform to the causal deduction logic in the original link diagram; State succession prediction mechanism: For the selected candidate node j, its potential evolutionary state is predicted by a state succession predictor (which can be a lightweight MLP, attention fusion module, etc.). This is used in subsequent path deduction and score calculation stages; to further ensure the rationality and causal consistency of the node state evolution direction during the local link restoration process, the following link restoration loss function is designed: ; in, h represents the set of real edges in the original cognitive link graph. j Let j be the true state vector of node j. The predicted state λ is obtained based on the reduction and deduction mechanism. ij The importance weight coefficient for link (i,j) is set based on link frequency, attention weight, or empirical rules. By minimizing the link reconstruction error, the model can improve the accuracy of reconstructing existing cognitive paths and provide a reliable foundation for subsequent future inference stages.

[0049] S22, Future Path Deduction and Multi-Step Link Generation: Based on the local link reconstruction, multi-step path deduction is carried out to predict the medium- and long-term trajectory of the user's cognitive link. The deduction process includes the following sub-processes: Path expansion strategy: Use breadth-first search (BFS) or heuristic depth-first search (DFS) to expand candidate links sequentially from the current node, and limit the maximum number of iterations L. max With branch width B max To avoid an explosive increase in the number of deduction paths; Evolutionary confidence accumulation mechanism: An evolutionary confidence score is accumulated for each inference path. The confidence score is composed of node state transition matching degree and link causality confidence γ. ij The calculation is weighted by the reasonableness of state continuity, and those with a confidence level below the threshold θ are filtered out. path The path; Optionally, for new cognitive stages that are missing, skipped, or not yet covered by links, potential intermediate nodes can be generated based on state generation models (such as conditional VAEs or diffusion models) to complete the deduction path and improve the integrity and continuity of the links.

[0050] To quantify the overall rationality and credibility of each deduction path's evolution, the path deduction scoring function is defined as follows: ; Among them, h i h j Let [hi;hj] represent the state vectors of node i and node j respectively, and [hi;hj] be the concatenation of the two; W and b are the learnable weight matrix and bias term, respectively; σ( ) is an activation function (such as sigmoid) used to output the rationality score of the inference from node i to node j; based on this score, multi-branch prediction of cognitive paths, rational path ranking and inference of potential motivations are realized, thereby supporting high-precision inference of users' future behavior in complex situations.

[0051] S23, Path Feasibility Screening and Dynamic Update: After the simulation is completed, paths are screened based on their rationality, goal relevance, and contextual consistency, and the set of future paths P after screening is retained. future The screening criteria include the following three aspects: Stage rationality verification: For each deduction path, verify whether the order of cognitive stages corresponding to its internal nodes conforms to the general logic of cognitive evolution; for example, the motivation decision-making stage usually precedes the path execution stage. If the stage order of nodes in the path is reversed, it is judged as unreasonable and eliminated. This verification helps to ensure the coherence of cognitive evolution in the deduction path and the possibility of conforming to actual behavioral patterns.

[0052] Goal Consistency Verification: For each derivation path, the goal orientation consistency of the path is evaluated by calculating the semantic relevance score between its terminal node state and the initial goal state; the goal relevance score function is defined as follows: ; in, h represents the state vector of the terminal node of path p. goal This represents the target state vector corresponding to the starting point of the deduction; the function uses cosine similarity, and the higher the score, the more the path matches the user's original goal; a target relevance threshold θ is set during the filtering process. target Paths below the threshold are removed.

[0053] Context-driven dynamic adaptation mechanism: Based on real-time changes in the external environment (such as new search behaviors and consumption behaviors), the future path prediction results are dynamically adjusted to achieve continuous updates and iterative optimization of the inference link. To adapt to dynamic changes in the real-time environment and user behavior (such as new search behaviors and in-store behaviors), an event-triggered inference update mechanism is supported. Specifically, when a new key behavioral event is detected, the future path inference and filtering are re-executed based on the current node state, and P is dynamically corrected. future A set of data is used to ensure the real-time nature and relevance of the prediction results.

[0054] Furthermore, step S23 includes the following sub-steps: Based on the confidence scores at each step of the comprehensive deduction process, the overall confidence score function for the path is defined as follows, which is used to further filter and rank the set of future paths: ; Where (i,j) represents the continuous node transitions in the derivation path p, (i,j)∈p represents each pair of continuous nodes in the path, and S predict (j|i) is the derivation scoring function; a path confidence threshold θ is set during screening. path High-confidence paths with scores above the threshold are retained.

[0055] Through the above-mentioned multiple screening mechanisms, the evolutionary rationality, target relevance and environmental adaptability of the inference path can be effectively improved, thereby providing reliable support for subsequent path-driven interpretable strategy generation and personalized recommendation.

[0056] Through the above cognitive link reconstruction and future path deduction process, not only can the reasonable link structure under the current cognitive state be restored, but also the cognitive evolution path that the user may form in the future can be predicted, providing rich and dynamic basic information support for the subsequent generation of path-driven explainable strategies.

[0057] In other optional implementations, based on the future path projection results, combined with cognitive motivation analysis and goal relevance assessment, a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation is generated. By analyzing the structural characteristics, evolutionary drivers, and goal matching degree of the projection path, interpretable and guiding strategy prompts are extracted, improving the system's transparency and adaptability in intelligent decision support. Therefore, in step S3, the generation of a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation based on the future path projection results, combined with cognitive motivation analysis and goal relevance assessment, specifically includes the following sub-steps: S31, Cognitive Motivation Extraction and Intention Attribution Analysis, specifically includes: Path motivation extraction: Based on the state vectors and evolution directions of each node in the deduced path, analyze potential cognitive motivations (such as interest shifts, demand evolution, decision wavering, etc.), and combine node semantic labels, stage features and contextual information for attribution modeling. Intent clustering and summarization: Cluster the deduction paths of similar motivational features, summarize the changing trends of users' potential goals, and form motivational tags (such as "preference shift to new brands" and "increased price sensitivity") to support subsequent strategy generation.

[0058] S32, Strategy candidate generation and optimization, specifically includes: Strategy rule base matching: Based on the matching of motivation tags with the strategy rule base (predefined or learned), a corresponding set of strategy candidates is generated, such as recommendation adjustment, reminder prompts, path simplification or intervention guidance, etc. Multi-dimensional optimization mechanism: Candidate strategies are comprehensively scored based on target relevance score, user acceptance prediction, and path rationality index to select the optimal strategy set; the strategy comprehensive scoring function is defined as follows: ; Among them, S target (s) represents the consistency score between the strategy and the initial goal, S accept (s) is the predicted probability that the strategy will be accepted or responded to by the user, S path(s) is the overall rationality score of the inference path corresponding to the strategy (which can be called Spath(p)), and α, β, γ are adjustable weight parameters that can be set according to the actual application scenario; S33, Policy Interpretability Representation and Output, specifically includes: Motivation strategy link generation: Generate a corresponding motivation explanation link for each output strategy, such as: "Because search behavior focuses on low-priced options → infer budget sensitivity → recommend promotional activities", to enhance strategy transparency; Optionally, natural language generation can be based on template filling, retrieval generation, or a lightweight NLG model to automatically generate user-understandable policy hints or illustrations.

[0059] Dynamic update mechanism: The strategy library and generation logic are dynamically adjusted based on the user's subsequent actual behavior feedback (such as clicks, ignores, reaction time, etc.) to improve the system's adaptability and personalization level.

[0060] By driving the explained strategy generation process through the above path, the automatic transformation from deduction path to intervention strategy can be achieved, while ensuring the clarity and logic of the explanation chain, providing solid support for user behavior prediction, recommendation optimization and decision support.

[0061] In other embodiments of the present invention, as optional enhancement components, to further improve the engineering adaptability, personalized processing capability, and cross-scenario migration capability of the system of the present invention, any one or more of the following enhancement components can be configured on top of the above-described embodiments: Multi-Scenario Adaptation Component: By recoding and mapping the semantic labels, behavioral stage attributes, and inference path rules of the cognitive link graph, this component enables the migration and adaptation of the model and inference logic across different business domains. Specifically, through flexible recoding and mapping of the semantic labels, behavioral stage attributes, and inference path rules of the cognitive link graph, it facilitates rapid migration and low-cost adaptation of the model and inference logic across different business domains (such as e-commerce, travel, finance, and education), lowering the barriers to cross-domain deployment and model migration. Implementation methods include, but are not limited to, introducing a Scene Adaptation Layer, which automatically redefines node categories, path stages, and inference strategy parameters based on scene tags. It also supports preset adaptation templates and online fine-tuning mechanisms to quickly adapt to new business needs, improving the model's universality and engineering application efficiency in multi-industry environments.

[0062] Personalized Explanation Generation Component: During the strategy generation and explanation output stages, personalized explanation content is generated tailored to different user profiles. Specifically, during these stages, personalized explanation content is generated based on different user profiles (such as interests, cognitive styles, and decision-making habits) to enhance user comprehension and trust. Implementation methods include, but are not limited to, introducing a Personalized Explanation Generator, which adaptively adjusts motivation extraction, strategy text templates, and language style by combining user feature vectors. It supports the use of lightweight style modeling (such as word vector-based or small generative networks) to achieve online personalized generation, improving the readability and approachability of the inference results and adapting to diverse user needs and multi-scenario communication standards.

[0063] The strategy output multimodal enhancement component presents the deduction results and strategy explanations in a multimodal manner, enhancing users' perception and understanding of future cognitive evolution trends. Specifically, it presents the deduction results and strategy explanations in a multimodal manner, such as text, graphical path diagrams, and dynamic recommendation cards, to enhance users' perception and understanding of future cognitive evolution trends. Implementation methods include, but are not limited to, generating dynamic visualizations based on existing cognitive link diagram structures and deduction paths, such as timeline evolution diagrams, node heatmaps, and strategy recommendation flowcharts; simultaneously, combining natural language output to form a richly illustrated multimodal recommendation interface, improving user experience and strategy acceptance rates, particularly suitable for complex decision support scenarios (such as car purchase, financial management, and travel planning).

[0064] It should be noted that by introducing the above optional enhancement components, this invention can further expand the system's adaptability, personalization level, and user interaction experience in practical engineering applications while ensuring the performance and interpretability of the basic modules, demonstrating good system scalability and engineering implementation value.

[0065] In other embodiments of the present invention, a cognitive behavior evolution prediction and path interpretability modeling apparatus is also provided, including a processor and a memory, wherein the memory stores a computer program, and when the computer program is loaded by the processor, it executes the method described in any of the preceding embodiments.

[0066] In other embodiments of the present invention, a cognitive behavior evolution prediction and path interpretability modeling system is also provided, characterized in that it includes the cognitive behavior evolution prediction and path interpretability modeling device as described above.

[0067] The units described in the embodiments of the present invention can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0068] According to one aspect of the present invention, a computer program product or computer program is provided, the computer program product or computer program including computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium, and executes the computer instructions, causing the computer device to perform the methods provided in the various optional implementations described above.

[0069] In another aspect, embodiments of the present invention also provide a computer-readable medium, which may be included in the electronic device described in the above embodiments; or it may exist independently and not assembled into the electronic device. The computer-readable medium carries one or more programs, which, when executed by the electronic device, cause the electronic device to perform the methods described in the above embodiments.

Claims

1. A method for predicting cognitive behavior evolution and modeling path interpretability, characterized in that, Includes the following steps: S1. The features of the structured behavioral chain data are input into the neural network for training. The neural network is used to model the temporal evolution of user behavior to predict the dynamic evolution of cognitive state. The neural network includes a time-series extended graph neural network (T-GNN). When it is a T-GNN, the features of the structured behavioral chain data are used as the features of graph nodes and edges. An externally interpretable weight combining the cognitive dwell characteristics and time decay characteristics of nodes is introduced to guide the aggregation process of nodes to neighbor nodes. The information propagation weights of self-loops and neighbor edges are defined using a weight function based on dwell strength and time decay. A causal consistency regularization mechanism is introduced during training to constrain the changing trends of node self-loop weights and neighbor weights, so as to guide the model to learn predictive behaviors that conform to the user's real cognitive evolution path. S2, based on the user's current cognitive state, historical behavior links and predicted evolution direction, deduces the possible evolution path on the cognitive link graph, and models and predicts the potential decision trajectory; S3, based on the results of future path projection, combined with cognitive motivation analysis and goal relevance assessment, generates a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation.

2. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 1, characterized in that, In step S1, the introduction of externally interpretable weights that combine node cognitive dwell characteristics and time decay characteristics to guide the aggregation process of neighbor node messages specifically includes the message aggregation phase process: The state vectors of all predecessor nodes j pointing to node i in the previous time step According to attention weight After weighting and summarizing, the summary message is obtained. The specific message aggregation formula is as follows: ; in, This represents the set of all predecessor nodes that point to i in the graph; This represents the attention shift weight from node j to i, reflecting the strength of the user's attention shifting from j to i; This represents a learnable projection matrix used to transform the feature vectors of neighboring nodes. Mapped to a message space of the same dimension as the target node; The feature vector of node j at time t-1; attention weights The definition is as follows: ； ratio Indicates the measurement of target node n j The dwell strength relative to the current node n i Differences in residence intensity; Time decay factor It is used to weaken the intensity of attention shift over long time intervals, representing the natural decay of attention over time, and simulating the natural decay process of user attention over time; represents the time interval between the behaviors of nodes ni and nj; μ is the attention decay coefficient over time, used to control the decay rate, which can be adaptively adjusted according to the user's behavior rhythm or system response time; δ is the normalization factor, used to unify the weight scale of all links.

3. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 1, characterized in that, In step S1, the introduction of externally interpretable weights that combine node cognitive dwell characteristics and time decay characteristics to guide the node's aggregation process of neighbor node messages specifically includes the node state update phase process: When a node is updated, its previous time-to-time characteristics are included. News gathered from neighbors Linear combination, then through activation function A nonlinear mapping is used to obtain the new hidden state of node i at time t. The specific node state update formula is as follows: ; in, This represents a self-circulating projection matrix that evolves co-evolves with neighbor aggregation weights and is subject to causal consistency constraints. Represents a non-linear activation function; This represents the hidden state vector of node i at time t, which will be used for prediction at the next time step or for downstream tasks.

4. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 2, characterized in that, In step S1, the method of defining the information propagation weights of self-loops and neighbor edges using a weighting function based on dwell strength and time decay specifically includes the following sub-steps: When node i receives a message from its neighbor node j, the weight function is designed as follows: ； where Res(i) represents the residence strength feature of node i; Δt ij denotes the time difference between node i and node j; μ is the time decay control parameter; for the self-loop edge, i.e., i = j, the time interval is zero, and the formula naturally degenerates to: ； The weighting function is designed to ensure that self-looping edges also carry the residence strength characteristics of cognitive node i rather than simple structural connections, and also to make the propagation strength of neighbor messages dynamically affected by the cognitive evolution process.

5. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 1, characterized in that, In step S1, a causal consistency regularization mechanism is introduced during training. This mechanism constrains the changing trends of node self-loop weights and neighbor weights to guide the model in learning predictive behaviors that align with the user's actual cognitive evolution path. Specifically, this includes the following sub-steps: A causal consistency correction term is introduced into the total loss function, which, together with the basic task loss, constitutes the total loss function for joint optimization. This constraint regularizes the evolution trend of the node's self-loop weights and neighbor weights, guiding the model to learn transition paths that conform to causal priors. The specific definition of the total loss function is as follows: ； wherein L task is the base task loss; L causal is the causal consistency loss for measuring the consistency of the predicted path with the expected causal evolution trend; β is the weight coefficient of the causal consistency constraint; L causal is defined as the deviation between the node state change and the causal prior path, and is expressed in the form as: ； in, Represents a pair of nodes on the predicted path; h i This represents the actual state of node i. This state is the true value obtained from user behavior data extracted from the input data, used to infer the causal relationships within the link; This represents a causal inference function. The state h of node i i After making a prediction, the predicted state of node j is obtained; this state is deduced through the causal relationship network, reflecting the possible evolution path from node i to node j. This indicates that the state h of node i is inferred based on causal priors. i The causal mapping function of the predicted evolution direction feature of its child node j is derived. This mapping is implemented by affine transformation, multilayer perceptron (MLP) or neural network module based on conditional generation mechanism. L is the causal credibility weight coefficient for edge (i,j), used to characterize the importance of link (i,j). This weight is set based on the attention distribution of the link in the original cognitive link graph, historical frequency statistics, or rules based on domain priors, thereby dynamically adjusting the contribution of different links to the overall causal consistency optimization objective; when the model predicts a path that conforms to a predetermined causal evolution trend, L... causal The loss is relatively small, but it increases when non-causal jumps or anomalous shifts occur in the prediction, thus effectively constraining the training process.

6. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 1, characterized in that, The specific features of the structured behavioral chain data include node dwell intensity, attention shift weight, and semantic label information.

7. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 1, characterized in that, In step S2, the process of deducing the possible evolutionary path of the user on the cognitive link graph based on the user's current cognitive state, historical behavioral links, and inferred evolutionary direction, and modeling and predicting potential decision trajectories, specifically includes the following sub-steps: S21, the current node state h output by the cognitive state evolution prediction model in step S1 i and the evolution characteristics are inferred and screened in combination with the historical link information in the original cognitive link graph to generate a local link structure. S22, based on the partial link reconstruction, conduct multi-step path deduction to predict the medium- and long-term trajectory of the possible evolution of the user's cognitive link; S23, after the deduction is completed, screening is performed based on the rationality of the path, the relevance of the target and the consistency of the context, and a future path set P after screening is retained future .

8. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 7, characterized in that, In step S21, the following mechanism is introduced during the link restoration process: Evolutionary direction matching mechanism: based on h i The feature vectors and candidate child nodes h in the original link j Similarity is evaluated based on the feature orientation, and cosine similarity, Euclidean distance, or matching score based on attention mechanism is used for screening. Causal consistency constraint mechanism: combining causal consistency loss L causal The evolution direction of the node state in the restored link must conform to the causal deduction logic in the original link diagram; State succession prediction mechanism: For the selected candidate node j, its potential evolutionary state is predicted by the state succession predictor. This is used in subsequent path deduction and score calculation stages; the following link reconstruction loss function is designed: ； in, h represents the set of real edges in the original cognitive link graph. j Let j be the true state vector of node j. The predicted state λ is obtained based on the reduction and deduction mechanism. ij The importance weight coefficient for link (i,j) is set based on link frequency, attention weight, or empirical rules. By minimizing the link reconstruction error, the model can improve the accuracy of reconstructing existing cognitive paths and provide a reliable foundation for subsequent future inference stages.

9. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 7, characterized in that, In step S22, the deduction process includes the following sub-processes: Path expansion strategy: Use breadth-first search (BFS) or heuristic depth-first search (DFS) to expand candidate links sequentially from the current node, and limit the maximum number of iterations L. max With branch width B max To avoid an explosive increase in the number of deduction paths; Evolutionary confidence accumulation mechanism: An evolutionary confidence score is accumulated for each inference path. The confidence score is composed of node state transition matching degree and link causality confidence γ. ij The calculation is weighted by the reasonableness of state continuity, and those with a confidence level below the threshold θ are filtered out. path The path; To quantify the overall rationality and credibility of each deduction path's evolution, the path deduction scoring function is defined as follows: ； Among them, h i h j Let [hi;hj] represent the state vectors of node i and node j respectively, and [hi;hj] be the concatenation of the two; W and b are the learnable weight matrix and bias term, respectively; σ( ) is the activation function used to output the rationality score of the evolution from node i to node j; based on this score, multi-branch prediction of cognitive paths, rational path ranking and inference of potential motivations are realized, thereby supporting high-precision prediction of users' future behavior in complex situations.

10. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 9, characterized in that, In step S22, the deduction process also includes the following sub-processes: For new cognitive stages that are missing, skipped, or not yet covered, potential intermediate nodes are generated based on state generation models to complete the deduction path and improve the integrity and continuity of the link.

11. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 7, characterized in that, In step S23, the screening criteria include stage rationality verification, target consistency verification, and context dynamic adaptation mechanism; The stage rationality verification checks whether the order of cognitive stages corresponding to the internal nodes of each deduction path conforms to the general logic of cognitive evolution. The goal consistency check evaluates the goal orientation consistency of each derivation path by calculating the semantic relevance score between its terminal node state and the initial goal state; the goal relevance score function is defined as follows: ； in, h represents the state vector of the terminal node of path p. goal This represents the target state vector corresponding to the starting point of the deduction; the function uses cosine similarity, and the higher the score, the more the path matches the user's original goal; a target relevance threshold θ is set during the filtering process. target Paths below the threshold are removed; The context-based dynamic adaptation mechanism dynamically adjusts the future path prediction results based on real-time changes in the external environment, enabling continuous updates and iterative optimization of the deduced path. When a new key behavioral event is detected, the future path deduction and filtering are re-executed based on the current node state, dynamically correcting P. future A set of data is used to ensure the real-time nature and relevance of the prediction results.

12. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 11, characterized in that, Step S23 includes the following sub-steps: Based on the confidence scores at each step of the comprehensive deduction process, the overall confidence score function for the path is defined as follows, which is used to further filter and rank the set of future paths: ； Where (i,j) represents the continuous node transitions in the derivation path p, (i,j)∈p represents each pair of continuous nodes in the path, and S predict (j|i) is the derivation scoring function; a path confidence threshold θ is set during screening. path High-confidence paths with scores above the threshold are retained.

13. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 1, characterized in that, In step S3, based on the future path projection results, combined with cognitive motivation analysis and goal relevance assessment, a set of feasible strategies for user behavior intervention, recommendation optimization, or strategy interpretation is generated, specifically including the following sub-steps: S31, Cognitive Motivation Extraction and Intent Attribution Analysis, specifically includes path motivation extraction and intent clustering and induction; the path motivation extraction is based on the state vectors and evolution directions of each node in the deduced path, analyzes potential cognitive motivations, and performs attribution modeling by combining node semantic labels, stage features, and contextual information; the intent clustering and induction clusters the deduced paths with similar motivational features, summarizes the changing trends of users' potential goals, forms motivational labels, and supports subsequent strategy generation; S32, Strategy candidate generation and optimization, specifically includes strategy rule base matching and a multi-dimensional optimization mechanism; the strategy rule base matching generates a corresponding set of strategy candidates based on the matching of motivation tags with the strategy rule base; the multi-dimensional optimization mechanism comprehensively scores the candidate strategies based on target relevance scores, user acceptance predictions, and path rationality indicators, and selects the optimal strategy set; the strategy comprehensive scoring function is defined as follows: ； Among them, S target (s) represents the consistency score between the strategy and the initial objective, S accept (s) is the predicted probability that the strategy will be accepted or responded to by the user, S path (s) represents the overall rationality score of the inference path corresponding to the strategy, where α, β, and γ are adjustable weight parameters that are set according to the actual application scenario. S33, Policy interpretability expression and output, specifically including motivational policy link generation and dynamic update mechanism; the motivational policy link generation is to generate a corresponding motivational explanation link for each output policy; the dynamic update mechanism dynamically adjusts the policy library and generation logic based on the user's subsequent actual behavior feedback.

14. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 13, characterized in that, Step S33 also includes the following sub-steps: Generate user-understandable policy hints or illustrations based on template filling, retrieval generation, or lightweight NLG models.

15. The method for predicting cognitive behavior evolution and modeling path interpretability according to claim 1, characterized in that, It also includes the following steps: Configure one or more of the following enhancement components: Multi-scenario adaptation component: By recoding and mapping the semantic labels of nodes, behavioral stage attributes and inference path rules of the cognitive link graph, the model and inference logic can be migrated and adapted to different business domains; Personalized Explanation Generation Component: During the strategy generation and explanation output phase, personalized explanation content is generated for different user profiles. Multimodal Enhancement Component for Strategy Output: Presents the inference results and strategy explanations in a multimodal manner, enhancing users' perception and understanding of future cognitive evolution trends.

16. The cognitive behavior evolution prediction and path interpretability modeling method according to claim 15, characterized in that, The implementation of the multi-scenario adaptation component specifically includes: introducing a scenario adaptation layer, automatically redefining node categories, path stages, and inference strategy parameters based on scenario tags; and supporting preset adaptation templates and online fine-tuning mechanisms to quickly adapt to new business requirements. The implementation of the personalized explanation generation component specifically includes: introducing a personalized explanation generator, combining user feature vectors, adaptively adjusting motivation extraction, strategy text templates, and language style, and supporting online personalized generation using lightweight style modeling; The implementation of the strategy output multimodal enhancement component specifically includes: generating a dynamic visualization based on the existing cognitive link graph structure and inference path; and forming a richly illustrated multimodal recommendation interface by combining natural language output.

17. A device for predicting cognitive behavior evolution and modeling path interpretability, characterized in that, It includes a processor and a memory, wherein the memory stores a computer program that, when loaded by the processor, executes the method as described in any one of claims 1 to 16.

18. A cognitive behavior evolution prediction and path interpretability modeling system, characterized in that, It includes the cognitive behavior evolution prediction and path interpretability modeling device as described in claim 17.

Images