Personalized service recommendation method based on knowledge graph and behavior sequence

Abstract

This invention discloses a personalized service recommendation method based on knowledge graphs and behavioral sequences. The method includes: constructing a global static topological potential energy surface of a service knowledge graph in a target domain; extracting the temporal prediction bias of user interaction behavior as an observation residual and updating the user intent state vector; dynamically adjusting the graph space decay constant according to the observation residual, calculating the local dynamic potential energy and fusing it with the global static topological potential energy surface to generate a real-time comprehensive topological potential energy field; starting from the updated intent state vector, performing node walk optimization along the negative gradient direction in the real-time comprehensive topological potential energy field, extracting candidate service entities in the local minimum domain to generate a recommended service list. This invention combines a state observation mechanism with physical field theory, and by adaptively adjusting the potential energy diffusion range through the observation residual, it realizes the dynamic switching between local precise mining and global cross-domain exploration of the recommendation path, alleviating the problem of response lag when user interests shift.

Description

Technical Field

[0001] This invention relates to the field of computer software and recommendation algorithm application technology, specifically to a personalized service recommendation method based on knowledge graphs and behavioral sequences. Background Technology

[0002] Personalized service recommendation systems are core components in modern internet applications used to alleviate information overload and improve user experience. Their basic function is to analyze users' historical behavior, preference attributes, and service content characteristics to filter out target services that match user needs from a massive dataset of candidate data and display them in a targeted manner. In multiple digital economy fields such as e-commerce, content distribution, and lifestyle services, recommendation systems have become a key bridge connecting users and services.

[0003] In existing recommendation system engineering practices, collaborative filtering or content-based recommendation algorithms are usually used. These algorithms construct user-item rating matrices or extract text feature vectors, and use mathematical tools such as matrix factorization, deep neural networks or attention mechanisms to calculate the similarity between user feature vectors and item feature vectors. The system sorts candidate items according to the calculated similarity scores, thereby generating the final recommendation list.

[0004] However, existing recommendation algorithms suffer from significant response lag when dealing with the evolution of user interests. Traditional models mostly rely on static learning from historical data spanning long periods, solidifying user intent into a fixed-dimensional representation vector. When a user's actual intent shifts rapidly or jumps across domains in a short period, existing systems cannot capture this dynamic change in real time and quantify the degree of deviation. Due to the lack of a closed-loop feedback mechanism for users' real-time interactive behavior and the ability to dynamically adjust the recommendation space, the system struggles to correct the recommendation logic in a timely manner. As a result, when user needs change abruptly, the recommendation results remain limited to a local range of historical interests, severely affecting the timeliness and accuracy of recommendations. Summary of the Invention

[0005] This invention provides a personalized service recommendation method based on knowledge graphs and behavioral sequences, comprising the following steps:

[0006] Step S100: Construct a global static topological potential energy surface for the target domain service knowledge graph. Obtain business data from the target service domain and extract service entities and business relationships from it, instantiating them into a directed graph structure in storage space. Calculate the in-degree, out-degree, and weighted degree of each service entity in the directed graph structure, determine the corresponding baseline topological potential energy, and thereby generate a global static topological potential energy surface data structure covering the knowledge graph space.

[0007] Step S200: Extract the temporal prediction bias of user interaction behavior, update and output the corrected user intent state vector. Collect user interaction behavior records uploaded by the client and convert them into multi-dimensional observation vectors. Perform forward evolution calculation through the state observer component to output the predicted intent state vector. Calculate the difference between the multi-dimensional observation vector and the predicted intent state vector to obtain the observation residual, and perform closed-loop compensation on the predicted intent state vector based on the observation residual to output the corrected user intent state vector.

[0008] Step S300: Dynamically adjust the spatial curvature of the spectral map based on the observed residuals to generate a real-time integrated topological potential energy field under the updated timestamp. Calculate the spatial decay constant for the current time step based on the observed residuals. Calculate the local dynamic potential energy centered on the interactive behavior trigger node based on the spatial decay constant, and fuse the baseline topological potential energy with the local dynamic potential energy to generate a real-time integrated topological potential energy field.

[0009] Step S400: Perform node gradient optimization based on the real-time integrated topological potential energy field to generate and distribute a recommended service list. Using the corrected user intent state vector as the optimization starting point, perform node walks along the negative gradient direction in the real-time integrated topological potential energy field to determine candidate service entities in local minima. Sort and truncate the candidate service entities, generate and distribute the recommended service list to the remote user equipment.

[0010] The present invention also provides an electronic device, including a processor and a memory. The memory is used to store a computer program, and the processor is used to call and run the computer program to execute the personalized service recommendation method described in the first aspect. The processor is connected to the memory and a communication interface via a bus to read and process business data, intent state vectors, and potential energy field data, supporting the generation and distribution of recommendation results.

[0011] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the personalized service recommendation method described in the first aspect above. The storage medium stores the logical instructions of the computer program in physical sectors, so that a computing device can reproduce the processing flow from knowledge graph potential construction to intent closed-loop correction and gradient optimization distribution.

[0012] The present invention, by adopting the above technical solution, can bring the following beneficial effects:

[0013] 1. This invention introduces a state observer component from cybernetics to perform closed-loop tracking of the user's implicit intent. It extracts the observation residual by utilizing the deviation between the collected real-time interactive behavior observation vector and the system's predicted intent state vector, and adjusts the decay constant of the knowledge graph physical space in real time based on the residual to change the curvature distribution of the potential energy field. This enables the recommendation system to adaptively respond to the degree of user interest shift, effectively alleviating the problems of model update lag and untimely recommendation feedback in traditional algorithms when facing the evolution of user intent.

[0014] 2. This invention constructs a global static topological potential energy surface of a service knowledge graph in the target domain, and establishes the baseline energy distribution of each node by combining the in-degree, out-degree, and association weights of service entities in the graph structure. It also transforms the user's real-time interactive behavior into a local dynamic potential energy component with time decay characteristics and nonlinearly superimposes and fuses it with the baseline potential energy surface. This achieves a deep unification of the semantic association features of service entities and the user's current interest intensity, thereby ensuring that the recommendation results can more accurately hit the user's deep business needs within a specific observation time window while complying with the logical constraints of the knowledge graph.

[0015] 3. This invention executes a minimum action path optimization algorithm in a real-time generated comprehensive topological potential energy field. It uses the corrected user intent state vector as the initial walking starting point and traverses nodes along the negative gradient direction of potential energy decrease. It uses the potential energy difference value to establish the search path until it captures the candidate service entity in the local minimum value domain. This achieves efficient screening and positioning of target recommendation results from massive heterogeneous service nodes, greatly reduces the hardware resource overhead caused by full node matching calculation in a large-scale knowledge graph environment, and significantly improves the execution efficiency of recommendation decision issuance. Attached Figure Description

[0016] Figure 1 This is a schematic diagram of the module architecture of the personalized service recommendation system of the present invention;

[0017] Figure 2 This is a flowchart illustrating the overall workflow of the personalized service recommendation method of the present invention.

[0018] Figure 3 This is a flowchart of the field curvature adjustment and potential field generator based on observation residuals of the present invention.

[0019] Figure 4 This is a schematic diagram of the hardware structure of the electronic device of the present invention. Detailed Implementation

[0020] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0021] Example 1

[0022] See attached document Figure 1-4 This invention provides a system for a personalized service recommendation method based on knowledge graphs and behavioral sequences, deployed on computing devices or distributed computing clusters with data processing capabilities. The system includes a baseline topological potential energy construction module, an intent state observation module, a field curvature adaptive adjustment module, and a recommendation decision delivery module. All four modules can be implemented through software programs, hardware logic circuits, or a combination of software and hardware. It should be noted that the physical or geometric terms such as "potential energy," "curvature," and "gradient" used in this application are general concepts for easier understanding of the overall solution. In the specific code implementation at the system's underlying level, they strictly correspond to floating-point weight values ​​in memory, scalar scaling factors, and numerical difference results between nodes, respectively.

[0023] The baseline topological potential energy construction module is used to calculate and allocate the initial energy values ​​of each service entity in the service domain knowledge graph. The baseline topological potential energy construction module extracts service entity and associated edge data from the underlying graph database or relational database. For graph storage and query of multi-source heterogeneous data, those skilled in the art can use a conventional graph database architecture. The graph data traversal and reading mechanism is a well-known technology in the field and will not be described in detail here. The baseline topological potential energy construction module performs quantitative calculation based on the weight of each service entity in the knowledge graph to obtain the baseline topological potential energy of each service entity, and summarizes the calculated baseline topological potential energy of all service entities to output global static topological potential energy distribution data.

[0024] The intent state observation module is used to track the user interaction behavior sequence returned by the client and calculate the state prediction deviation. The intent state observation module maintains a continuous state vector reflecting the user intent characteristics in the memory space and has a built-in cybernetics state observer component to receive external interaction behavior input and perform system state updates. The lower-level implementation scheme of the state observer component adopts Kalman filter, extended Kalman filter or Romberg observer. The intent state observation module generates observation residual by calculating the L2 norm of the deviation between the actual received behavior observation value and the predicted value output by the state observer component, and transmits the observation residual to the field curvature adaptive adjustment module.

[0025] The field curvature adaptive adjustment module is used to receive the observation residuals transmitted by the intention state observation module and dynamically adjust the spatial decay constant of the map and the superimposed dynamic potential energy. The field curvature adaptive adjustment module calculates the updated spatial decay constant based on the input observation residuals and transforms the user interaction behavior that occurs in the current time window into a dynamic potential energy injection source. Combining the updated spatial decay constant and time decay factor, it calculates the local dynamic potential energy centered on the interaction node. Furthermore, it superimposes the local dynamic potential energy with the benchmark topological potential energy output by the benchmark topological potential energy construction module to generate real-time comprehensive topological potential energy field data.

[0026] The recommendation decision delivery module is used to perform gradient optimization of discrete nodes in the real-time integrated topological potential energy field. Based on the updated intent state vector, the module matches the corresponding service entity in the knowledge graph as the starting point for the walk, calculates the integrated potential energy difference between adjacent service entities to establish the negative gradient direction, and traverses nodes along the negative gradient direction within the constraint depth. The module numerically sorts the service entities corresponding to the potential energy minima reached during the traversal, extracts service entities that meet the set quantity threshold to generate a recommendation result dataset, and sends it to the remote user device through the communication interface.

[0027] Example 2

[0028] See attached document Figure 1-4 This embodiment provides a personalized service recommendation method based on knowledge graphs and behavioral sequences, running in a system with data processing capabilities. The method utilizes the system's internal hardware processor and memory to collaboratively execute service data processing and recommendation decisions. The method includes the following steps:

[0029] Step S100: The system constructs the global static topological potential surface of the target domain service knowledge graph;

[0030] The system obtains structured business data from the service domain through a data interface. The system extracts the service entities and the objective semantic relationships between the service entities from the structured business data and instantiates them into a directed graph structure in the storage space.

[0031] The system traverses each service entity in the directed graph structure, extracts and counts the topological structure parameters of the service entities. The specific features of the topological structure parameters include the in-degree, out-degree and weighted degree of the service entity in the knowledge graph. Based on the extracted weighted degree of the service entity, the system calculates the baseline topological potential corresponding to all service entities in the knowledge graph.

[0032] The system numerically aggregates and structurally encapsulates the baseline topological potential energy of all service entities, generating a global static topological potential energy surface data structure covering the entire knowledge graph space. The system stores the global static topological potential energy surface data structure in memory as the baseline surface parameters for subsequent calculations.

[0033] Step S200: The system extracts the temporal prediction deviation of user interaction behavior, updates and outputs the corrected user intent state vector.

[0034] The system establishes observation time windows divided according to discrete time steps. Within each discrete time step, the system continuously collects user interaction behavior records uploaded by the client. The user interaction behavior records include the interaction behavior type, the service entity identifier that triggered the interaction, and the corresponding timestamp parameter. The system maintains a state vector representing the user's implicit intent in memory. The system inputs the state vectors of historical time steps into the internal cybernetics state observer component. The cybernetics state observer component performs forward evolution calculation of the system state based on the node transition probability of the knowledge graph structure and outputs the predicted intent state vector of the current time step. The system converts the actual user interaction behavior records collected within the current time step into observation vectors.

[0035] The system performs a difference operation on the behavioral expectation data generated by the observation vector and the predicted intention state vector, calculates and extracts the deviation between the two as the observation residual, calls the built-in observer gain matrix, and combines the observation residual to compensate and update the predicted intention state vector, thereby obtaining the corrected user intention state vector at the current time step.

[0036] In step S300, the system dynamically adjusts the curvature of the spectral space based on the observed residuals to generate a real-time integrated topological potential field under the updated timestamp;

[0037] The system reads the observation residual values ​​output in step S200, substitutes the observation residual values ​​into the mapping logic, calculates and outputs the corresponding spatial decay constant, and the spatial decay constant characterizes the steepness of the energy outward diffusion gradient in the physical space of the map.

[0038] Based on the collected user interaction behavior records, the system injects local dynamic potential energy that follows an exponential decay law into the knowledge graph, with the service entity that triggered the interaction as the origin.

[0039] The system reads the global static topological potential energy surface data structure from memory. The system performs algebraic subtraction between the local dynamic potential energy values ​​corresponding to each service entity and the reference topological potential energy values ​​in the global static topological potential energy surface to obtain the above-mentioned real-time integrated topological potential energy, thereby forming the real-time integrated topological potential energy field.

[0040] In step S400, the system performs node gradient optimization based on the real-time integrated topological potential energy field, and generates and distributes a list of recommended services.

[0041] The system obtains the corrected user intent state vector. The system adopts the nearest neighbor retrieval mechanism to map the user intent state vector to the node vector space of the knowledge graph. The system extracts several service entity nodes within the distance threshold and adds the extracted service entity nodes to the optimization starting point set.

[0042] The system uses each service entity in the optimization starting point set as the reference position, reads the real-time comprehensive topological potential energy of the adjacent service entities in the knowledge graph that have a connection edge with the reference position, calculates the comprehensive potential energy difference between the optimization starting point and the adjacent service entities, establishes the negative gradient direction that makes the potential energy decrease, and controls the optimization path to extend in the negative gradient direction within the maximum graph walking depth range of the constraint. The system records the service entities corresponding to the boundaries of zero or positive potential energy difference in the walking path, and marks the recorded service entities as candidate service entities in the local minimum value domain.

[0043] The system arranges candidate service entities in ascending order based on real-time integrated topological potential energy values. The system extracts a preset number of candidate service entities at the top of the ranking to form a personalized recommendation service list. The system generates message data containing the personalized recommendation service list and sends it to the remote user equipment through the network communication interface to complete the interactive display.

[0044] Example 3

[0045] See attached document Figure 1-4 The method provided in this embodiment is executed by the benchmark topological potential energy construction module in the computing device, which transforms discrete business data into a graph structure. The graph space mapping and topological feature extraction process includes the following steps:

[0046] In step S101, the personalized recommendation system acquires business data from the target service domain, extracts service entities and business relationships, accesses the data storage node of the business server, and reads structured form data and unstructured text data containing service information. The personalized recommendation system calls a named entity recognition model to parse the unstructured text data, extracting independent service entities and the interaction relationships between them. For the construction and training of the named entity recognition model, those skilled in the art can use a network architecture based on bidirectional long short-term memory networks and conditional random fields. The entity extraction and sequence labeling process of the named entity recognition model is a well-known technology in the field and will not be elaborated here. The personalized recommendation system merges the inherent business logic in the structured form data with the parsed and extracted interaction relationships to generate a unified set of business relationships.

[0047] In step S102, the personalized recommendation system instantiates a directed graph structure in memory space, configures topological nodes and directed edges, maps each extracted service entity to a discrete node in the directed graph structure, maps business relationships to directed edges connecting different discrete nodes, and counts the frequency of occurrence of each business relationship in the original business data, and calculates the weight value of the directed edges based on the frequency of occurrence. The specific weight value calculation formula is as follows:

[0048]

[0049] In the formula, Indicates from discrete nodes Pointing to discrete nodes The directed edge weight values, Represents discrete nodes With discrete nodes The frequency of business relationships between nodes in the original business data is used to calculate the weight values ​​of the nodes and assign them to the corresponding directed edges. The personalized recommendation system writes the dataset containing discrete nodes, directed edges and weight values ​​into the graph database to establish a service knowledge graph.

[0050] Step S103: The personalized recommendation system traverses each discrete node in the directed graph structure, calculates the weighted degree parameters of the service entities, and for a single service entity in the knowledge graph, extracts all input edges pointing to the current service entity and all output edges pointed to by the current service entity. The personalized recommendation system reads the weight values ​​bound to each input edge and output edge, obtains the preset in-degree adjustment coefficient and the preset out-degree adjustment coefficient, and performs a weighted summation calculation on the input edge weights and output edge weights. The specific weighted degree calculation formula is as follows:

[0051]

[0052] In the formula, Indicates service entity The weighting degree, Indicates pointing to the service entity The set of predecessor nodes, Indicates from the predecessor node Pointing to service entity The directed edge weights are calculated by summing the weights of all incoming predecessor edges. Indicates that by the service entity The set of successor nodes pointed to. Indicates from service entity Point to the successor node The directed edge weights are calculated by summing the weights of all subsequent outgoing edges. This represents the preset in-degree adjustment coefficient. This represents the preset output adjustment coefficient;

[0053] The personalized recommendation system updates the calculated weighted degree to the corresponding service entity attribute field in the knowledge graph. The personalized recommendation system completes the traversal and calculation of all service entities in the knowledge graph, stores the weighted degree of each service entity as a parameter, and passes the weighted degree parameter to the association calculation process to calculate the baseline topological potential of each service entity.

[0054] Furthermore, the method provided in this embodiment is executed by a personalized recommendation system to assign initial numerical distributions to each service entity in the knowledge graph. The mechanism for establishing the static baseline potential includes the following steps:

[0055] Step S104: The personalized recommendation system obtains the preset potential energy reference surface adjustment constant. The potential energy reference surface adjustment constant is a floating-point number that is greater than zero and is used to maintain the non-negative mathematical properties of the potential energy calculation results. The personalized recommendation system reads the weighting parameters of each service entity in the knowledge graph from the associated calculation process.

[0056] Step S105: The personalized recommendation system performs algebraic operations based on the weighted degree parameters and the potential energy reference surface adjustment constant. The personalized recommendation system increments the weighted degree parameters of the service entities by one and calculates the natural logarithm to obtain the cardinality of the topological connectivity. The specific formula for calculating the cardinality of the topological connectivity is as follows:

[0057]

[0058] In the formula, Indicates service entity Cardinality of topological connectivity Indicates service entity The weighted degree is calculated, and the reciprocal of the topological connectivity cardinality is calculated. The calculated reciprocal is added to the potential energy reference surface adjustment constant, and the reference topological potential energy of the corresponding service entity is output. The personalized recommendation system traverses the knowledge graph to complete the task of calculating the reference topological potential energy of all service entities in the knowledge graph.

[0059] In step S106, the personalized recommendation system summarizes the baseline topological potential energy of all service entities. The personalized recommendation system allocates a one-dimensional array structure in the memory space of the computing device. Following the ascending index order of the globally unique identifiers of the service entities in the knowledge graph, the personalized recommendation system sequentially writes the baseline topological potential energy of each service entity into the one-dimensional array structure. This one-dimensional array structure constitutes a global static topological potential energy surface data structure. The mathematical expression formula for the global static topological potential energy surface data structure is as follows:

[0060]

[0061] In the formula, This represents the column vector corresponding to the global static topological potential surface data structure. This represents the baseline topological potential of the service entity whose global unique identifier index is first in order. The baseline topological potential represents the service entity whose globally unique identifier is the Nth order. This represents the total number of service entities in the knowledge graph. This represents the matrix transpose operation;

[0062] Personalized recommendation systems store global static topological potential energy surface data structures in memory;

[0063] Example 4

[0064] See attached document Figure 1-4 The method provided in this embodiment is executed by the intent state observation module in the computing device, and is used to transform the dynamic user interaction process into a system evolution model in cybernetics. The execution process of state space definition and forward evolution prediction includes the following steps:

[0065] Step S201: The personalized recommendation system initializes a discrete state space model in the memory of the computing device. The personalized recommendation system divides the observation time window according to the preset time step. The personalized recommendation system establishes an intention state vector that represents the user's implicit intention. The dimension of the intention state vector is equal to the total number of service entities in the knowledge graph. The personalized recommendation system extracts the user's interaction behavior records in the current observation time window and uses one-hot encoding or multi-hot encoding to convert the service entity identifier that triggers the interaction into a multi-dimensional observation vector and a control input vector with the same dimension as the intention state vector.

[0066] Step S202: The personalized recommendation system constructs a system matrix based on the topological structure data of the knowledge graph. The personalized recommendation system extracts the directed edge weights between various service entities in the knowledge graph. Based on these directed edge weights, the personalized recommendation system calculates the random walk transition probabilities between nodes and generates a system state transition matrix. The specific formula for calculating the elements of the system state transition matrix is ​​as follows:

[0067]

[0068] In the formula, Represents the system state transition matrix The Middle Line 1 The matrix elements of the columns represent the user's intent from the service entity. Transfer to service entity The probability, Indicates from service entity Pointing to service entity The directed edge weight values, Indicates that by the service entity The set of successor nodes that it points to;

[0069] The personalized recommendation system synchronously initializes the input control matrix and the observation matrix. The personalized recommendation system configures the input control matrix as a diagonal matrix, and the diagonal elements of the diagonal matrix correspond to the gain coefficients of each service entity receiving external stimuli. The personalized recommendation system configures the observation matrix as an identity matrix to realize the equal-dimensional mapping transformation from the intent state vector space to the actual interaction behavior observation space.

[0070] Step S203: The personalized recommendation system performs forward evolution calculation of the user's intent. The personalized recommendation system obtains the corrected intent state vector output from the previous discrete time step and the control input vector received from the previous discrete time step. The personalized recommendation system substitutes the system state transition matrix, input control matrix, corrected intent state vector, and control input vector into the forward inference equation of the state observer component to calculate the predicted intent state vector for the current discrete time step. The specific forward inference equation formula is as follows:

[0071]

[0072] In the formula, Indicates based on The intention of the time-may-time state vector is calculated. Predict the intention state vector at any time. Represents the system state transition matrix. express The intentional state vector at time step. Indicates the input control matrix. express The control input vector at each time step;

[0073] The personalized recommendation system writes the calculated predicted intention state vector into a buffer memory. The predicted intention state vector serves as the basis for comparison parameters for subsequent extraction of observation residuals and execution of state closed-loop correction.

[0074] Furthermore, the method provided in this embodiment is executed by the intent state observation module in the computing device, and is used to adjust the internal state prediction results of the system in conjunction with the actual interaction data. The execution process of prediction residual calculation and intent state closed-loop correction includes the following steps:

[0075] Step S204: The personalized recommendation system extracts the multi-dimensional observation vector and the predicted intent state vector generated at the current discrete time step. The personalized recommendation system uses a preset observation matrix to perform a linear transformation on the predicted intent state vector, outputting a predicted behavior vector. The personalized recommendation system performs element-wise subtraction between the multi-dimensional observation vector and the predicted behavior vector to obtain the residual vector. The specific formula for calculating the residual vector is as follows:

[0076]

[0077] In the formula, This represents the residual vector at the current discrete time step. The multidimensional observation vector at the current discrete time step. Represents the observation matrix. Represents the predicted intention state vector;

[0078] The personalized recommendation system calculates the L2 norm of the residual vector and outputs the observed residual values ​​in scalar form. Stored in a computing device;

[0079] Step S205: The personalized recommendation system dynamically calculates the observer gain matrix based on the local topological entropy of the knowledge graph. The personalized recommendation system parses the multi-dimensional observation vector to locate the target service entity that the user actually triggered the interaction in the current discrete time step. The personalized recommendation system extracts the target service entity and its neighbor nodes within a preset number of hops, and calculates the local topological entropy of the target service entity. The specific formula for calculating the local topological entropy is as follows:

[0080]

[0081] In the formula, Indicates the target service entity Local topological entropy, Indicates from the target service entity Transfer to successor node The transition probability, Indicates that the target service entity The set of successor nodes that it points to;

[0082] Personalized recommendation systems calculate the gain adjustment coefficient based on local topological entropy. The specific formula for calculating the gain adjustment coefficient is as follows:

[0083]

[0084] In the formula, This represents the gain adjustment coefficient for the current discrete time step. and These represent the preset constant adjustment coefficients;

[0085] The personalized recommendation system reads the basic observer gain matrix, multiplies it by the gain adjustment coefficient, and outputs the dynamic observer gain matrix for the current discrete time step. The specific formula for calculating the dynamic observer gain matrix is ​​as follows:

[0086]

[0087] In the formula, Represents the dynamic observer gain matrix. This represents the gain adjustment coefficient for the current discrete time step. Represents the basic observer gain matrix;

[0088] When the local topological entropy of the target service entity is high, it indicates complex node connections and strong uncertainty in the current interaction behavior (i.e., corresponding to a high level of observation noise). The gain adjustment coefficient decreases, and the personalized recommendation system reduces the weight of the dynamic observer gain matrix to weaken the sensitivity to observation noise. When the local topological entropy of the target service entity is low, it indicates strong determinism in the current interaction behavior (i.e., corresponding to a low level of observation noise). The gain adjustment coefficient increases, and the personalized recommendation system increases the weight of the dynamic observer gain matrix to enhance the correction magnitude of the intent state.

[0089] Step S206: The personalized recommendation system performs closed-loop correction of the user's intent state. The personalized recommendation system obtains the residual vector and the dynamic observer gain matrix. It then performs matrix multiplication of the dynamic observer gain matrix and the residual vector to obtain the state compensation vector. Finally, the personalized recommendation system adds the state compensation vector to the predicted intent state vector to complete the closed-loop intervention on the predicted state. The specific formula for the closed-loop correction of the intent state is as follows:

[0090]

[0091] In the formula, This represents the intentional state vector after closed-loop correction at the current moment. Represents the predicted intention state vector. Represents the dynamic observer gain matrix. Represents the residual vector;

[0092] The personalized recommendation system writes the closed-loop corrected intent state vector into the memory space, and the closed-loop corrected intent state vector is used to drive the forward inference calculation in the next observation time window.

[0093] Example 5

[0094] See attached document Figure 1-4The method provided in this embodiment is executed by the field curvature adaptive adjustment module in the computing device, which is used to transform the observation residuals output by the intention state observation process into constraint parameters for energy diffusion in the knowledge graph. The spatial curvature adaptive mapping process based on the observation residuals includes the following steps:

[0095] Step S301: The personalized recommendation system reads the observation residual corresponding to the current observation time window. The personalized recommendation system obtains the preset initial spatial attenuation benchmark value and the preset curvature adjustment parameter from the configuration storage area. The initial spatial attenuation benchmark value is used to characterize the set energy transmission attenuation benchmark rate between knowledge graph nodes under the condition of no external interactive behavior intervention. The curvature adjustment parameter is used to control the numerical influence weight of the observation residual on the transmission attenuation benchmark rate.

[0096] In step S302, the personalized recommendation system calculates the curvature scaling denominator based on the observed residuals and curvature adjustment parameters. The personalized recommendation system performs a multiplication operation between the curvature adjustment parameters and the observed residuals, adds a constant 1 to the product of the multiplication operation, and outputs the curvature scaling denominator. The specific formula for calculating the curvature scaling denominator is as follows:

[0097]

[0098] In the formula, express The curvature scaling denominator at time step, This indicates the preset curvature adjustment parameters. express The observation residual at any given moment;

[0099] Step S303: The personalized recommendation system calculates the spatial decay constant for the current time step using the initial spatial decay benchmark value and the curvature scaling denominator. The personalized recommendation system divides the initial spatial decay benchmark value by the curvature scaling denominator and outputs the corresponding spatial decay constant. The specific formula for calculating the spatial decay constant is as follows:

[0100]

[0101] In the formula, express The spatial decay constant at time , This represents the preset initial spatial attenuation reference value. express The curvature scaling denominator at time, the spatial decay constant As the negative exponential multiplier parameter in the subsequent exponential decay function, the smaller its value, the gentler the energy decay gradient with spatial distance (i.e., the flatter the potential energy field), which is used to define the steepness of the potential energy diffusion gradient.

[0102] When the observed residual value increases, it indicates that the user's current interaction behavior deviates from the system's historical state prediction results. The curvature scaling denominator increases accordingly, and the calculated spatial decay constant value decreases. The reduced spatial decay constant decreases the decay rate of the subsequently generated local dynamic potential energy in the knowledge graph (i.e., the potential energy field becomes flatter), expands the coverage hop range of potential energy diffusion, and increases the probability of cross-domain service nodes being included in the recommendation list. When the observed residual value decreases, it indicates that the user's current interaction behavior conforms to the system's historical state prediction results. The calculated spatial decay constant value increases, and the increased spatial decay constant accelerates the decay of potential energy diffusion (i.e., the potential energy field becomes steeper). The potential energy diffusion range shrinks, and the mining accuracy of adjacent nodes of the service entity that triggered the interaction is improved.

[0103] In step S304, the personalized recommendation system writes the calculated spatial decay constant into memory, and the spatial decay constant is provided to the local dynamic potential energy calculation process as a constant input parameter.

[0104] Furthermore, the method provided in this embodiment is executed by the field curvature adaptive adjustment module in the computing device, which is used to convert the user's real-time interactive behavior into energy distribution in physical space and generate a real-time integrated topological potential energy field for recommendation decision-making. The dynamic potential energy trap construction and global field fusion process includes the following steps:

[0105] Step S305: The personalized recommendation system extracts the user's interaction behavior sequence within the current observation time window. The personalized recommendation system parses the interaction behavior sequence to obtain the trigger node identifier, behavior occurrence time, and interaction type parameter of each interaction behavior. The personalized recommendation system matches preset intensity weights based on the interaction type parameter. The interaction type parameter includes click behavior, favorite behavior, and purchase behavior. The personalized recommendation system configures incremental intensity weight values ​​for click behavior, favorite behavior, and purchase behavior in sequence.

[0106] Step S306: The personalized recommendation system calculates the spatial distance between each service entity and the trigger node in the knowledge graph. Starting from the service entity corresponding to the trigger node identifier, the personalized recommendation system traverses the target service entities in the knowledge graph and calculates the shortest path distance between the trigger node and the target service entity in the directed graph structure.

[0107] Step S307: The personalized recommendation system combines the time decay factor and the spatial decay constant to calculate the local dynamic potential energy centered on the trigger node. The personalized recommendation system reads the updated spatial decay constant from memory. Based on the time of the behavior, the current time step, the shortest path distance, and the spatial decay constant, the personalized recommendation system calculates the dynamic potential energy value received by each service entity. The specific formula for calculating the local dynamic potential energy is as follows:

[0108]

[0109] In the formula, Indicates service entity exist Local dynamic potential energy at time t. Indicates the length of the interaction sequence within the observation time window. Indicates the first The intensity weight corresponding to each interaction behavior This indicates the preset time decay rate. Indicates the current time step. Indicates the first The timing of the next interaction. express The spatial decay constant at time , Indicates service entity With the The node that triggers this interaction behavior Shortest path distance in a knowledge graph;

[0110] Step S308: The personalized recommendation system performs a fusion calculation of the global topological potential energy field. The personalized recommendation system reads the global static topological potential energy surface data structure stored in memory, extracts the baseline topological potential energy of each service entity, and subtracts the baseline topological potential energy from the local dynamic potential energy to output the real-time comprehensive topological potential energy of each service entity. The specific potential energy fusion calculation formula is as follows:

[0111]

[0112] In the formula, Indicates service entity exist Real-time integrated topological potential energy at any given moment. Indicates service entity The baseline topological potential energy;

[0113] Indicates service entity exist Regarding the local dynamic potential energy at any given time, it should be clarified that in this embodiment of the invention, the lower the potential energy value, the higher the recommendation priority of the service entity. Therefore, by subtracting the local dynamic potential energy from the baseline topological potential energy (i.e., the subtraction operation), the comprehensive potential energy of the service entities adjacent to the trigger node decreases, thereby increasing their "attractiveness" to users and their recommendation priority. The personalized recommendation system numerically combines the real-time comprehensive topological potential energy of all service entities in the knowledge graph to generate real-time comprehensive topological potential energy field data. The personalized recommendation system outputs the real-time comprehensive topological potential energy field data to the recommendation decision distribution module for performing gradient optimization of the recommendation nodes.

[0114] Example 6

[0115] See attached document Figure 1-4 The method provided in this embodiment is executed by the recommendation decision distribution module in the computing device, and is used to perform gradient optimization of discrete nodes in the real-time integrated topological potential energy field. The execution process of the minimum action path optimization algorithm includes the following steps:

[0116] Step S401: The personalized recommendation system determines the set of starting points for path optimization;

[0117] The personalized recommendation system obtains the intent state vector after closed-loop correction. The personalized recommendation system reads the state values ​​corresponding to each dimension of the intent state vector. Each dimension of the intent state vector corresponds one-to-one with each service entity in the knowledge graph. The personalized recommendation system sorts each service entity in descending order of state value. The personalized recommendation system extracts a preset number of service entities in the sorted order and writes the extracted service entities as the initial service entities into the optimization starting point set.

[0118] Step S402: The personalized recommendation system calculates the potential energy difference values ​​between adjacent service entities in the real-time integrated topological potential energy field.

[0119] The personalized recommendation system starts from the initial service entity in the optimization starting set, extracts adjacent service entities in the knowledge graph that have connecting edges with the initial service entity, and reads the real-time comprehensive topological potential energy of the adjacent service entities and the initial service entity at the current moment. The personalized recommendation system subtracts the real-time comprehensive topological potential energy of the initial service entity from the real-time comprehensive topological potential energy of the adjacent service entities, and outputs the potential energy difference value. The specific formula for calculating the potential energy difference is as follows:

[0120]

[0121] In the formula, Indicates starting from the current walking node To adjacent service entities The potential energy difference value, Indicates adjacent service entities exist Real-time integrated topological potential energy at any given moment. Indicates the current walking node exist Real-time integrated topological potential energy at any given moment, and the current wandering node The initial walk corresponds to the initial service entity;

[0122] Step S403: The personalized recommendation system performs node walks along the negative gradient direction.

[0123] The mathematical symbol of the potential energy difference value calculated in step S402 of the personalized recommendation system. When there is an adjacent service entity with a potential energy difference value less than zero, the personalized recommendation system determines that there is a negative gradient direction of potential energy decrease. Among all adjacent service entities with potential energy difference values ​​less than zero, the personalized recommendation system selects the adjacent service entity corresponding to the largest absolute value of potential energy difference value as the next hop node. The personalized recommendation system advances the hop path to the next hop node and updates the hop depth recording parameters in the system.

[0124] In step S404, the personalized recommendation system determines the walk termination condition and outputs candidate service entities. The personalized recommendation system uses the updated next-hop walk node as the new current walk node and continues to execute the logic of steps S402 and S403 until the preset walk termination condition is triggered. The walk termination condition includes the walk depth recording parameter reaching the preset maximum depth threshold, or the potential energy difference value of all adjacent service entities of the current walk node being greater than or equal to zero. When the potential energy difference value of all adjacent service entities of the current walk node is greater than or equal to zero, it indicates that the current walk node is in a local minimum region on the potential energy surface. The personalized recommendation system extracts the resident node when the walk termination condition is triggered and writes the resident node as a candidate service entity into the candidate set. After the personalized recommendation system completes the walk traversal of all initial service entities in the optimization starting point set, it outputs a candidate set containing multiple candidate service entities.

[0125] Furthermore, the method provided in this embodiment is executed by the recommendation decision distribution module in the computing device, used to extract the final recommendation data from the candidate set output by the walk-and-optimize process and complete the network interaction. The execution process of candidate result truncation and recommendation list generation includes the following steps:

[0126] Step S405: The personalized recommendation system performs deduplication and sorting operations on the candidate service entities in the candidate set. The personalized recommendation system reads the globally unique identifier of each candidate service entity in the candidate set. The personalized recommendation system compares the globally unique identifiers and removes redundant candidate service entities with the same globally unique identifier to obtain the deduplicated candidate service entity set. The personalized recommendation system extracts the real-time comprehensive topological potential of each candidate service entity in the deduplicated candidate service entity set.

[0127] The personalized recommendation system sorts the deduplicated candidate service entities in ascending order according to their real-time comprehensive topological potential energy values, generating a ranked service entity queue. The ranked service entity queue satisfies the following mathematical relationship:

[0128]

[0129] In the formula, This indicates that the entity in the sorting service queue is at position 1. Candidate service entities in order of priority Indicates candidate service entities exist Real-time integrated topological potential energy at any given moment. This represents the total number of service entities contained in the candidate service entity set after deduplication;

[0130] Step S406: Extract the target recommended entity from the self-ranking service entity queue of the personalized recommendation system;

[0131] The personalized recommendation system obtains a preset recommendation quantity threshold. The personalized recommendation system compares the total number of service entities in the deduplicated candidate service entity set with the recommendation quantity threshold. When the total number of service entities is greater than or equal to the recommendation quantity threshold, the personalized recommendation system extracts candidate service entities with a quantity equal to the recommendation quantity threshold from the starting position of the sorted service entity queue. When the total number of service entities is less than the recommendation quantity threshold, the personalized recommendation system extracts all candidate service entities in the sorted service entity queue. The personalized recommendation system defines the extracted candidate service entities as target recommendation entities.

[0132] The personalized recommendation system reads the business attribute data of the target recommended entity. The business attribute data includes the service name, service network link, and service description text.

[0133] Step S407: The personalized recommendation system generates a recommendation list message and executes network delivery;

[0134] The personalized recommendation system encodes the target recommended entity and its business attribute data according to a preset structured language format and encapsulates them to generate a personalized recommendation service list.

[0135] The personalized recommendation system calls the network communication interface of the computing device to send a communication message containing a list of personalized recommendation services to the remote user device. The remote user device receives the communication message and parses the service attribute data to display the list of personalized recommendation services on the display screen of the remote user device.

[0136] Example 7

[0137] See attached document Figure 1-4 To execute the aforementioned personalized service recommendation method, this method is performed using the electronic device provided in this embodiment.

[0138] Electronic devices include processors, memory, communication interfaces, and buses. The processor, memory, and communication interface are electrically connected and communicate data through the bus. The bus is responsible for transmitting information between the various components of the electronic device. The specific implementation features of the bus include industry standard architecture bus, peripheral component interconnection standard bus, or extended peripheral component interconnection standard bus.

[0139] The processor connects to various hardware components of electronic devices through physical interfaces and internal buses. The processor runs or executes software program instructions stored in memory, calls data stored in memory, and executes the core operation logic of the personalized recommendation system. The specific lower-level implementation features of the processor include central processing unit, graphics processing unit, tensor processing unit, application-specific integrated circuit, digital signal processor or field-programmable gate array.

[0140] The memory is used to store computer-executable instructions and business operation data. The memory includes high-speed random access memory and non-volatile memory. The specific implementation characteristics of non-volatile memory include disk storage devices, solid-state drives, flash memory devices or read-only memory. The memory is internally divided into a program storage area and a data storage area. The program storage area is used to store operating system code and application software programs required to execute personalized recommendation methods. The data storage area is used to cache directed graph structure data of knowledge graphs, global static topological potential energy surface data structures, user interaction behavior records, observation residual values, and real-time integrated topological potential energy field data.

[0141] The processor reads computer instructions from memory and executes the aforementioned personalized recommendation method. The processor extracts structured business data to construct a knowledge graph, calculates the baseline topological potential energy of service entities to generate a global static topological potential energy surface, runs the control theory state observer component, extracts observation residuals by combining the received user interaction behavior sequence, completes the closed-loop correction of the intent state vector, dynamically adjusts the spatial decay constant based on the observation residuals, constructs local dynamic potential energy and generates a real-time comprehensive topological potential energy field, and performs walk optimization along the negative gradient direction of the potential energy of adjacent nodes in the real-time comprehensive topological potential energy field, outputs candidate service entities in the local minimum range and generates a personalized recommendation service list.

[0142] The communication interface is used to establish a data transmission channel between electronic devices and external business servers and clients. The communication interface receives service entity data sent by the business server and user interaction behavior records uploaded by the client. The communication interface sends communication messages generated by the processor, which contain a personalized recommendation service list, to the remote user equipment. The specific lower-level implementation features of the communication interface include local area network card interface, fiber optic communication interface or wireless baseband radio frequency module.

[0143] Furthermore, embodiments of the present invention provide a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the aforementioned personalized recommendation method.

[0144] The specific implementation types of computer-readable storage media include non-volatile storage media and volatile storage media. The substructure of non-volatile storage media includes read-only memory, programmable read-only memory, electrically erasable programmable read-only memory, magnetic storage media, solid-state drive or optical disk. The substructure of volatile storage media includes dynamic random access memory or static random access memory.

[0145] A computer program contains multiple computer-executable instructions. When these instructions are loaded and executed by the processor, they control the processor to perform the following personalized recommended steps:

[0146] Control the processor to construct the global static topological potential surface of the target domain service knowledge graph;

[0147] The control processor extracts the temporal prediction deviation of user interaction behavior, updates and outputs the corrected user intent state vector;

[0148] The control processor dynamically adjusts the curvature of the spectral space based on the observed residuals to generate a real-time integrated topological potential field with updated timestamps;

[0149] The control processor performs node gradient optimization based on the real-time integrated topological potential energy field, and generates and distributes a list of recommended services.

Claims

1. A personalized service recommendation method based on knowledge graphs and behavioral sequences, characterized in that, The method includes: Step S100: Construct the global static topological potential surface of the target domain service knowledge graph; Step S200: Extract the temporal prediction deviation of user interaction behavior, update and output the corrected user intent state vector; Step S300: Based on the observation residuals, dynamically adjust the curvature of the spectral space to generate a real-time integrated topological potential field under the updated timestamp; Step S400: Node gradient optimization is performed based on the real-time integrated topological potential energy field to generate and distribute a recommended service list.

2. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 1, characterized in that, The personalized service recommendation method is applied to multimedia content distribution or e-commerce recommendation scenarios. The service entity is multimedia content resources or physical goods. The user interaction behavior includes user-triggered clicks, browsing, favorites, adding to cart, or purchasing operations on the client.

3. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 1, characterized in that, Step S100 includes: Step S101: Obtain business data of the target service domain and extract service entities and business relationships; Step S102: Instantiate a directed graph structure in the storage space, and calculate the weight values ​​of the directed edges based on the frequency of occurrence of each of the business relationships. Step S103: Traverse the discrete nodes in the directed graph structure and calculate the weighted degree parameter of the service entity based on the weight value; Step S104: Obtain the preset potential energy reference surface adjustment constant and read the weighting parameters of each of the service entities; Step S105: Perform algebraic operations based on the weighting parameters and the preset potential energy reference surface adjustment constant to output the reference topological potential energy corresponding to the service entity. Step S106: Write all the reference topological potential energy into a one-dimensional array structure to form the global static topological potential energy surface data structure.

4. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 3, characterized in that, In step S103, the weighted parameter of the service entity is calculated, and the specific calculation formula is as follows: In the formula, Indicates service entity The weighting degree, Indicates pointing to the service entity The set of predecessor nodes, Indicates from the predecessor node Pointing to service entity The directed edge weights are summed by adding the weights of all incoming edges belonging to the predecessor node set. Indicates that by the service entity The set of successor nodes pointed to. Indicates from service entity Point to the successor node The directed edge weights are defined by summing the weights of all outgoing edges belonging to the successor node set. This represents the preset in-degree adjustment coefficient. This indicates the preset output adjustment coefficient.

5. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 1, characterized in that, Step S200 includes: Step S201: Establish observation time windows divided according to discrete time steps, collect user interaction behavior records and convert them into multi-dimensional observation vectors with the same dimension as the intent state vector; Step S202: Extract the weight values ​​of the directed edges between each service entity in the knowledge graph to calculate the transition probability between nodes, generate the system state transition matrix, and initialize the input control matrix and observation matrix. Step S203: Input the intention state vector of the historical time step into the cybernetics state observer component to perform forward evolution calculation, and output the predicted intention state vector of the current time step. Step S204: Perform a difference operation between the multidimensional observation vector and the predicted intention state vector to obtain the residual vector and output the observation residual value in scalar form.

6. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 4, characterized in that, Step S200 further includes: Step S205: Calculate the local topological entropy of the target service entity, calculate the gain adjustment coefficient based on the local topological entropy, and multiply the basic observer gain matrix by the gain adjustment coefficient to output the dynamic observer gain matrix. Step S206: Perform matrix multiplication of the dynamic observer gain matrix and the residual vector to obtain the state compensation vector, and add the state compensation vector to the predicted intention state vector to complete the closed-loop correction of the predicted state and output the corrected user intention state vector.

7. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 3, characterized in that, Step S300 includes: Step S301: Obtain the preset initial spatial attenuation reference value and the preset curvature adjustment parameters; Step S302: Calculate the curvature scaling denominator based on the observed residuals and the preset curvature adjustment parameters; Step S303: Divide the preset initial spatial attenuation reference value by the curvature scaling denominator to calculate the spatial attenuation constant for the current time step; Step S305: Extract the interaction behavior sequence within the current observation time window and parse the interaction type parameter to configure the corresponding intensity weight value; Step S306: Calculate the shortest path distance between each service entity and the node that triggers the interaction behavior in the knowledge graph; Step S307: Combine the preset time decay rate with the spatial decay constant to calculate the local dynamic potential energy centered on the interactive behavior trigger node; Step S308: Subtract the reference topological potential energy from the local dynamic potential energy to output the real-time integrated topological potential energy of each service entity and generate the real-time integrated topological potential energy field data. The specific formula for calculating the local dynamic potential energy in step S307 is as follows: In the formula, Indicates service entity exist Local dynamic potential energy at time t. Indicates the length of the interaction sequence. Indicates intensity weight, This indicates the preset time decay rate. Indicates the current time step. Indicates the time when the interaction occurs. Represents the spatial decay constant. Indicates service entity With trigger node The shortest path distance in a knowledge graph.

8. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 1, characterized in that, Step S400 includes: Step S401: Read the service entities whose values ​​are within the previous preset number in the corrected user intent state vector and write them into the optimization starting point set; Step S402: Starting from the initial service entity in the optimization starting point set, calculate the potential energy difference value between the current wandering node and its adjacent service entities; Step S403: When there is an adjacent service entity whose potential energy difference value is less than zero, select the adjacent service entity corresponding to the potential energy difference value with the largest absolute value as the next hop walking node, and perform node walking along the negative gradient direction.

9. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 8, characterized in that, Step S400 further includes: Step S404: Continue traversing until a preset traversal termination condition is triggered. The traversal termination condition includes the traversal depth reaching a preset maximum depth threshold or the potential energy difference values ​​of adjacent service entities being greater than or equal to zero. The resident node is then marked as a candidate service entity in a local minimum range and written into the candidate set.

10. The personalized service recommendation method based on knowledge graphs and behavioral sequences according to claim 9, characterized in that, Step S400 further includes: Step S405: Perform deduplication on the candidate service entities in the candidate set, and generate a sorted service entity queue in ascending order of the real-time comprehensive topological potential energy value. Step S406: Define the candidate service entities in the ranking service entity queue whose number is equal to the recommendation quantity threshold as target recommendation entities; Step S407: Encode and encapsulate the target recommended entity and its business attribute data to generate a personalized recommendation service list, and send it to the remote user equipment through the network communication interface.

Images