A knowledge graph-based user portrait construction method and system

Abstract

This invention discloses a user profile construction method and system based on knowledge graphs, belonging to the field of knowledge graph user modeling technology. The method includes generating a directed acyclic graph of user interest evolution based on multi-source raw behavioral data, reverse-engineering feasible evolutionary chains from current interests to historical trigger nodes, calculating the pattern confidence and evolutionary rationality score of each chain and sorting and filtering them, integrating the filtered evolutionary chains to form a hierarchical interest tracing network, obtaining interest clusters with single core themes through subgraph segmentation, performing cross-time window influence propagation analysis on each interest cluster, quantifying the contribution weight of interest trigger nodes, and generating a structured dynamic user profile report by combining the domain knowledge graph. This method improves the dynamism and tracing accuracy of user profiles through dual-dimensional filtering of evolutionary chains and quantification of interest cluster influence, constructing a logically clear and thematically focused dynamic user profile.

Description

Technical Field

[0001] This invention belongs to the field of knowledge graph user modeling technology, specifically a method and system for constructing user profiles based on knowledge graphs. Background Technology

[0002] Existing user profiling technologies based on knowledge graphs mostly employ forward association mapping to process multi-source raw behavioral data of target users. They only construct static interest associations, failing to build a directed acyclic graph of user interest evolution or perform reverse reasoning from the current interest state to historical trigger nodes, thus failing to form a complete interest evolution chain system. Current technologies only use a single metric to evaluate the effectiveness of interest associations, without simultaneously calculating the pattern confidence and evolutionary rationality scores of the evolutionary path. They also fail to perform ranking and filtering of the evolutionary chain based on two-dimensional metrics, making it difficult to distinguish between effective evolutionary paths and redundant, noisy associations.

[0003] Existing technologies fail to integrate the filtered evolutionary chains into a hierarchical interest tracing network, do not perform subgraph segmentation on the overall network structure, cannot separate interest clusters focusing on a single core interest topic, and do not conduct cross-time window influence propagation analysis for interest clusters, thus failing to quantify the contribution weight of each interest trigger node to the final interest state. User profiles constructed using such technologies only present static correlation characteristics, lacking the tracing logic of interest evolution and the ability to segment topics; the profile structure is chaotic and lacks quantitative basis. This invention aims to address the problems of incomplete reverse derivation of interest evolution chains and the lack of dual-dimensional evaluation and screening, while also resolving the deficiencies in hierarchical tracing network construction, subgraph segmentation, and cross-time window influence quantification. Summary of the Invention

[0004] This invention aims to solve at least one of the technical problems existing in the prior art; Therefore, this invention proposes a user profile construction method based on knowledge graphs, including: Based on the target user's original behavioral data across multiple information sources, a directed acyclic graph of user interest evolution is generated. Based on the aforementioned directed acyclic graph of user interest evolution, all feasible evolutionary chains from the current interest state to the historical trigger node are derived in reverse. For each of the feasible evolutionary chains, calculate the pattern confidence and evolutionary rationality score of its path, and sort and filter all feasible evolutionary chains based on the pattern confidence and evolutionary rationality score; Integrate all feasible evolutionary chains after sorting and filtering to form a hierarchical interest tracing network. In the hierarchical interest tracing network, each node represents an interest entity or behavioral state, and each edge represents an evolutionary driving relationship. The hierarchical interest tracing network is segmented into subgraphs to decompose the complex network structure into multiple interest clusters, each focusing on a core interest topic. For each of the interest clusters, perform cross-time window influence propagation analysis to quantify the contribution weight of each interest triggering node to the final interest state; Based on the contribution weight of each interest-triggered node and combined with the domain knowledge graph, a structured dynamic user profile report is generated.

[0005] Furthermore, the generation of a directed acyclic graph of user interest evolution based on the target user's original behavioral data across multiple information sources includes: Receive raw behavioral data of target users on multiple information sources, including search queries, page browsing, content interaction, social relationships and transaction records; Perform time-series cleaning and entity linking operations on the raw behavioral data to identify core entities and associated events in the user behavior sequence; Based on the identified core entities and associated events, a multidimensional behavioral feature tensor is constructed. The multidimensional behavioral feature tensor is used to represent the potential associations between user behavior, entity attributes and contextual information. The multidimensional behavioral feature tensor is input into a learnable behavioral pattern discovery network. The learnable behavioral pattern discovery network generates a directed acyclic graph of user interest evolution by iteratively pruning accidental associated edges and strengthening stable pattern paths. The step of performing time-series cleaning and entity linking operations on the original behavioral data to identify core entities and associated events in the user behavior sequence includes: Natural language processing is performed on the search query and the content interaction to extract named entities, keywords and sentiment tendencies to form a standardized intent encoding sequence; The page browsing and transaction records are segmented and pattern recognized to extract meaningful session fragments and behavioral patterns, and key behavioral events are located from them. User nodes, contact nodes, and relationship types are extracted from the social relationships to construct a local social subgraph; The standardized intent encoding sequence, the key behavioral events, and the local social subgraph are aligned and merged according to a unified timeline; Using the domain knowledge graph, entity disambiguation and linking operations are performed on the merged data, mapping text mentions and behavioral events to unique entity identifiers in the knowledge graph, thereby completing the identification of the core entities and the associated events.

[0006] Furthermore, the multidimensional behavioral feature tensor is input into a learnable behavioral pattern discovery network. This network generates a directed acyclic graph (DAG) of user interest evolution by iteratively pruning accidental association edges and reinforcing stable pattern paths, including: Initialize a fully connected behavior association graph structure, wherein each feature dimension in the multidimensional behavior feature tensor corresponds to a node in the graph; Using an attention-based structure learner, the existence probability and directional weight of each edge in the fully connected behavioral graph structure are learned by optimizing an objective function that includes sparsity constraints and acyclicity penalties. During training, the structure learner outputs an adjacency matrix in each round and calculates the directed acyclicity of the current graph structure based on the adjacency matrix; The acyclicity penalty is dynamically increased based on the degree of violation of directed acyclicity, and edges that form loops are removed, while stable associated edges with weights exceeding a threshold are retained. When the output of the structure learner changes below the preset tolerance, the model is determined to have converged, and the final adjacency matrix structure is transformed into the user interest evolution directed acyclic graph, where the edge weights represent the evolution driving strength.

[0007] Furthermore, based on the aforementioned directed acyclic graph of user interest evolution, all feasible evolutionary chains from the current interest state to historical trigger nodes are derived in reverse, including: In the directed acyclic graph of user interest evolution, the node representing the user's interest state at the most recent time point is marked as the current state node; Starting from the current state node, the reverse breadth-first search algorithm is used to traverse the directed acyclic graph of user interest evolution in the opposite direction of the directed edges; During the traversal, the system records the complete sequence of nodes and edges traversed from the current state node back to any source node representing the initial interest or external trigger. Each such backtracking path is a feasible evolutionary chain. For each of the feasible evolutionary chains, verify whether the timestamp order of its nodes is consistent with the order of occurrence of the original behavioral data, and filter out paths with contradictory time logic.

[0008] Furthermore, for each of the aforementioned feasible evolutionary chains, a pattern confidence score and an evolutionary rationality score are calculated for its path, including: For a single feasible evolutionary chain, extract the evolution driving strength weights of all directed edges on the feasible evolutionary chain; A basic mode strength value is obtained by continuously multiplying all evolution driving strength weights on the feasible evolution chain. By combining the global prior distribution probability of each node in the directed acyclic graph of user interest evolution, Bayesian inference is used to correct the basic pattern strength value to obtain the pattern confidence of the feasible evolution chain. Obtain the actual time difference of adjacent nodes on the feasible evolutionary chain in the original behavioral data; The actual time difference of this behavior is compared with the common time intervals of similar interest shifts statistically derived from historical group behavior, and the matching degree is calculated as the evolutionary rationality score of the feasible evolutionary chain.

[0009] Furthermore, the integration of all feasible evolutionary chains after sorting and filtering forms a hierarchical interest tracing network, including: Merge all feasible evolutionary chains in which both the pattern confidence score and the evolutionary rationality score are higher than their respective set thresholds; During the merging process, identify and merge identical nodes and edges between different feasible evolutionary chains; Using shared high-frequency nodes as the network hub, multiple feasible evolution chains are woven into a directed acyclic mesh topology. Based on the typical time stages in which nodes appear in the evolutionary chain, the nodes in the mesh topology are automatically classified into a potential interest layer, a development and evolution layer, and a stable state layer, thereby forming the hierarchical interest tracing network with a clear hierarchy.

[0010] Furthermore, the hierarchical interest tracing network is subjected to subgraph segmentation, decomposing the complex network structure into multiple interest clusters, including: Calculate the pattern correlation degree between any two nodes in the hierarchical interest tracing network. The pattern correlation degree is determined by the edge weight on the shortest path between the nodes and the number of shared neighbors. A modularity-based community detection algorithm is adopted to divide network nodes according to the pattern correlation degree, so that after the division, the nodes within the same community are closely related, while the nodes between different communities are sparsely related. Each subgraph corresponding to a community is extracted from the hierarchical interest tracing network, and the subgraph constitutes an interest cluster. For each interest cluster, a descriptive topic tag is automatically assigned based on the entity type distribution of its internal nodes.

[0011] Furthermore, for each of the interest clusters, a cross-time-window influence propagation analysis is performed to quantify the contribution weight of each interest-triggered node to the final interest state, including: Within a single interest cluster, the interest node located in the stable state layer is designated as the evaluation target node; An independent cascade model is applied to simulate the propagation of causal influence. Starting from each candidate trigger node in the interest cluster located in the potential interest layer or the development and evolution layer, the process of its influence spreading along the directed edge to the evaluation target node is simulated. During the simulated diffusion, the strength of an influence is multiplied by the evolution driving strength weight of the directed edge for each directed edge it passes through to decay. Finally, all influence values ​​that successfully reach the evaluation target node are accumulated. The cumulative influence value contributed by each candidate triggering node is normalized to the proportion of the total influence value contributed by all candidate triggering nodes. This proportion is the contribution weight of the candidate triggering node.

[0012] Furthermore, based on the contribution weight of each interest-triggered node and combined with the domain knowledge graph, a structured dynamic user profile report is generated, including: Filter out interest-triggered nodes whose contribution weight is greater than a set contribution weight threshold, and mark them as key interest-driving factors; In the domain knowledge graph, query the superordinate concepts, synonyms, attributes, and typical application scenarios associated with each key interest-driven factor; All key interest drivers are sorted from highest to lowest according to their contribution weight; For each key interest-driven factor, the interest cluster to which it belongs, the calculated contribution weight, the related concepts and attributes retrieved from the knowledge graph are recorded in a structured manner, and potential user intentions or needs are derived accordingly. The selected key interest drivers and their related information are organized and populated according to the preset user profile report template to generate the final structured dynamic user profile report.

[0013] Furthermore, the present invention also includes a user profile construction system based on a knowledge graph, the system including a memory, a processor, and a computer program stored in the memory and running on the processor, wherein when the processor executes the computer program, it implements the steps of the user profile construction method based on a knowledge graph as described above.

[0014] Compared with the prior art, the beneficial effects of the present invention are: Based on the directed acyclic graph of user interest evolution, all feasible evolutionary chains from the current interest state to historical trigger nodes are derived in reverse. For each feasible evolutionary chain, the pattern confidence and evolutionary rationality scores are calculated. All feasible evolutionary chains are sorted and filtered based on the dual-dimensional scores, which can completely cover all evolutionary paths of user interests from historical nodes to the current state. Pattern confidence measures the degree of pattern matching of the evolutionary path, and evolutionary rationality score can judge the logical compliance of the path. The combination of dual-dimensional indicators can accurately distinguish between effective evolutionary paths and invalid noise paths. The sorting and filtering operations can eliminate redundant links with low confidence and low rationality, and retain the core paths that conform to the real interest evolution logic, so that the related links of interest evolution have clear logical support.

[0015] The integrated and screened feasible evolutionary chains form a hierarchical interest tracing network. This network is then subgraph-segmented into multiple interest clusters focusing on a single core interest theme. For each interest cluster, cross-time-window influence propagation analysis is performed to quantify the contribution weight of each interest trigger node to the final interest state. This process can organize the chaotic and scattered evolutionary relationships into a hierarchical structure. Subgraph segmentation enables the independent splitting of different interest themes, avoiding structural chaos caused by multiple themes. Cross-time-window influence propagation analysis can adapt to the temporal changes in user interests, accurately defining the degree of influence of each historical trigger node in the interest formation process. The quantified contribution weights can directly support subsequent profile generation, giving user profiles the characteristics of dynamic tracing and precise quantification. Attached Figure Description

[0016] Figure 1 This is a flowchart illustrating the steps of a knowledge graph-based user profile construction method according to the present invention. Figure 2 A flowchart for the network processing of learnable behavior pattern discovery; Figure 3 A trend chart showing the weights of nodes in the evolution of user interests; Figure 4 A hierarchical interest tracing network node distribution diagram; Figure 5 Weighting analysis diagram for candidate triggering nodes. Detailed Implementation

[0017] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0018] See Figure 1This invention provides a user profile construction method based on knowledge graphs, the method comprising: Based on the target user's raw behavioral data across multiple information sources, a directed acyclic graph (DAG) of user interest evolution is generated. This graph uses nodes to represent interest entities or behavioral states and directed edges to represent evolutionary driving relationships. From this graph, starting with the current interest state node, all feasible evolutionary chains tracing back to historical trigger nodes are derived. For each feasible evolutionary chain, a pattern confidence score and an evolutionary rationality score are calculated, and all feasible evolutionary chains are ranked and filtered based on these scores. The filtered feasible evolutionary chains are integrated to form a hierarchical interest tracing network with clear layers of potential interests, developmental evolution, and stable states. This network is then subgraphed into multiple interest clusters focusing on specific themes. For each interest cluster, cross-time-window influence propagation analysis is performed to quantify the contribution weight of each historical interest trigger node to the final stable interest state. Based on these contribution weights and combined with a domain knowledge graph, a structured dynamic user profile report is generated.

[0019] In one embodiment of the present invention, raw behavioral data of a target user from multiple information sources, such as search engines, content platforms, e-commerce platforms, and social networks, is received. This raw behavioral data includes search queries, page browsing, content interaction, social relationships, and transaction records. Temporal cleaning and entity linking operations are performed on the raw behavioral data to identify core entities and related events in the user's behavioral sequence. This process includes natural language processing of search query and content interaction text to extract named entities, keywords, and sentiment tendencies, forming a standardized intent encoding sequence; session segmentation and pattern recognition of page browsing and transaction records to extract meaningful session fragments and behavioral patterns and locate key behavioral events; parsing user nodes, contact nodes, and relationship types from social relationship data to construct a local social subgraph; aligning and merging the obtained standardized intent encoding sequence, key behavioral events, and local social subgraph along a unified timeline; and using a pre-built domain knowledge graph to perform entity disambiguation and linking operations on the merged data, mapping text mentions and behavioral events to unique entity identifiers in the knowledge graph. Based on the identified core entities and related events, a multidimensional behavioral feature tensor is constructed. This tensor integrates multidimensional features such as behavior type, entity attributes, timestamps, and contextual information to represent the potential relationships between them. The multidimensional behavioral feature tensor is then input into a learnable behavior pattern discovery network. This network iteratively prunes accidental association edges and strengthens stable pattern paths, ultimately outputting a directed acyclic graph of user interest evolution.

[0020] In specific implementation, raw behavioral data of target users from multiple information sources is received. This raw behavioral data includes search queries entered by users on search engines, page browsing records generated in news applications, content interaction behaviors on video platforms, social relationship data on social networks, and transaction records on e-commerce platforms. These multi-source heterogeneous data together constitute a comprehensive record of the target user's behavior. Temporal cleaning and entity linking operations are performed on the raw behavioral data. Temporal cleaning includes standardizing and aligning messy timestamps and correcting logical errors. The purpose of entity linking is to identify core entities and associated events in the user's behavioral sequence, transforming discrete behavioral records into structured entity-event sequences. In some embodiments, natural language processing is performed on the text content accompanying search queries and content interaction behaviors. The processing includes word segmentation, part-of-speech tagging, named entity recognition, keyword extraction, and sentiment analysis, extracting named entities, keywords, and sentiment tendencies to form a standardized intent encoding sequence. Optionally, session segmentation and pattern recognition are performed on page browsing and transaction records based on time and event continuity. Sessions are segmented according to the topic coherence of the browsing pages or the transaction interval, and meaningful session fragments and behavioral patterns are extracted from the session fragments. Behavioral patterns, such as the "search-browse-favorite" sequence, are used to locate key behavioral events, such as "purchase decision clicks." From social relationship data, the central node representing the user, neighboring nodes representing contacts, and the relationship types between nodes are parsed to construct a user-centric local social subgraph, which reflects the user's direct social network structure. Essentially, the standardized intent encoding sequence, key behavioral events, and local social subgraph are aligned and merged along a unified timeline to form a timestamped multimodal behavioral trajectory. Timestamp alignment ensures the consistency of the temporal order of events from different sources. By utilizing a pre-built domain knowledge graph, entity disambiguation and linking operations are performed on the merged behavioral trajectory data. Entity disambiguation solves the problem of entities with the same name or ambiguous references. Entity linking operations map the references in the text and the objects involved in the behavioral events to unique entity identifiers in the knowledge graph, thus completing the identification of core entities and related events. For example, the text "Python" is mapped to the entity "Python (programming language)" in the knowledge graph, and the "add to cart" event is mapped to the "purchase intention" behavioral event type.

[0021] In practical implementation, based on the identified core entities and related events, and combined with the time, frequency, and contextual information of the behavior, a multidimensional behavioral feature tensor is constructed. The multidimensional behavioral feature tensor can be formalized as a high-dimensional array, where each dimension corresponds to a feature type, such as a behavior type dimension, entity attribute dimension, temporal context dimension, and contextual label dimension. The multidimensional behavioral feature tensor is used to represent the potential associations between user behavior, entity attributes, and contextual context. It can be understood that the behavior type dimension encodes discrete action types such as search, click, play, purchase, share, and follow; the entity attribute dimension encodes attribute values ​​such as category, tag, and description of the corresponding entity in the knowledge graph; the temporal context dimension encodes information such as the absolute timestamp, relative time interval, and position in the behavior sequence; and the contextual label dimension may encode information such as device type, network environment, geographical location, or application channel. In some embodiments, the construction of the multidimensional behavioral feature tensor can be represented as a feature combination function, for example, for a behavior instance... Its corresponding feature vector It can be built in the following ways: in: It is a type of behavior Embedded vector, It is an entity associated with behavior. The attribute aggregation embedding vector, It's a timestamp. Time-encoded vector, It is a multi-hot encoded vector of context labels. This represents a vector concatenation operation. The constructed multidimensional behavioral feature tensor is input into a learnable behavioral pattern discovery network. This network iteratively prunes accidental connections and strengthens stable pattern paths, ultimately generating a directed acyclic graph (DAG) of user interest evolution. The learnable behavioral pattern discovery network can be an autoencoder structure or a causal discovery network structure based on graph neural networks. Its training objective is to learn stable and directed causal or temporal relationships between behavioral features. In some embodiments, during training, the behavioral pattern discovery network evaluates the statistical significance and temporal rationality of the edges between nodes of different feature dimensions in the multidimensional behavioral feature tensor. Through gradient backpropagation and weight updates, it gradually removes edges that do not meet the conditions and strengthens the weights of pattern paths that repeatedly appear in multiple user behavior sequences until the network structure stabilizes and converges, outputting a weighted, acyclic directed graph as the DAG of user interest evolution.

[0022] In one embodiment of the present invention, when generating a directed acyclic graph of user interest evolution, a multidimensional behavioral feature tensor is input into a learnable behavioral pattern discovery network, see [reference]. Figure 2The network first initializes a fully connected behavior graph structure, where each feature dimension of the tensor corresponds to a node in the graph. The network uses an attention-based structure learner, which learns the existence probability and directional weight of each edge in the fully connected graph by optimizing an objective function that includes sparsity constraints and a directed acyclicity (DAG) penalty term. During training, the structure learner outputs an adjacency matrix representing the current graph structure in each iteration and calculates the corresponding DAG metric. Based on the degree of violation of the DAG constraint, the weight of the DAG penalty term is dynamically increased, and edges that lead to loops are removed, while stable edges with weights exceeding a preset threshold are retained. When the change in the adjacency matrix output by the structure learner is lower than a preset tolerance threshold, the model is considered to have converged. The structure represented by the final converged adjacency matrix is ​​transformed into a directed acyclic graph of user interest evolution, where the weight of each directed edge represents the driving strength of the evolution from the source node's interest to the target node's interest.

[0023] In practice, the multidimensional behavioral feature tensor is input into a learnable behavior pattern discovery network. The first step of this network is to initialize a fully connected behavior association graph structure. In this graph, each feature dimension of the multidimensional behavioral feature tensor corresponds to an independent node. In some embodiments, if the multidimensional behavioral feature tensor has N feature dimensions, the initialized fully connected graph structure contains N nodes, and there are two directed edges in opposite directions between any two different nodes, forming a fully directed graph topology. The network uses an attention-based structure learner. This learner receives the multidimensional behavioral feature tensor as input and learns to assign an existence probability value and a directional weight representing causality or evolutionary strength to each directed edge in the fully connected graph. The structure learner is trained by optimizing a predefined objective function. This function typically includes two constraints: a sparsity constraint that promotes the sparsity of the graph structure, and a directed acyclic penalty that penalizes the appearance of cycles in the graph. The directed acyclicity penalty term ensures that the final learned graph structure does not contain cyclic dependencies, thus conforming to the temporal irreversibility of interest evolution.

[0024] In practice, the attention-based structure learner outputs an adjacency matrix representing the current graph's connectivity state in each training iteration. Each element of this matrix corresponds to the weight of the directed edge from node i to node j. The structure learner calculates the directed acyclicity (DAG) metric of the current graph structure based on this adjacency matrix; a common method is to check the exponential trace of the adjacency matrix. During training, the system dynamically monitors the DAG metric. If a cycle exists in the current graph structure, violating the DAG constraint, the strength of the DAG penalty term in the objective function is dynamically increased based on the degree of violation, forcing the structure learner to adjust parameters in the next iteration to reduce cycles. Simultaneously, after each iteration, the adjacency matrix output by the structure learner is post-processed. Edges forming cycles are directly removed, and edges with weights below a preset threshold are pruned as accidental connections, while stable connections with weights exceeding the threshold are retained. Optionally, the structure learner is trained through multiple iterations. The model is considered converged when the norm of the adjacency matrix output by the structure learner in two adjacent training iterations falls below a preset tolerance. After convergence, the graph structure represented by the post-processed adjacency matrix output in the final iteration is transformed into a final directed acyclic graph (DAG) of user interest evolution. In this DAG, the weight of each directed edge represents the evolutionary driving strength from the source node's interest to the target node's interest.

[0025] In practical implementation, the objective function optimized by the attention-based structure learner can be formally represented as minimizing a loss function, the structure of which can be designed as follows: in: It is the reconstruction loss that measures the ability of a structure learner to reconstruct the multidimensional behavioral feature tensor of the input. It is the parameter matrix representing the edge weights and directions that the structure learner is to learn. These are sparsity constraint coefficients that control the sparsity of parameters. It is the L1 norm of the parameter matrix, used to promote the sparsity of graph structures. It is a coefficient that controls the strength of the directed acyclicity penalty. It is a parameter matrix Matrices of the same dimension, their elements The prior or basic weights of the probability that an edge exists from node i to node j are encoded. This represents the Hadamard product of a matrix (element-by-element multiplication). It is a matrix The matrix index, It is the trace of the matrix. The trace of the matrix exponent acts as a directed acyclicity penalty term here, if and only if the matrix... When the graph corresponding to the weighted adjacency matrix is ​​a directed acyclic graph, the value of this term reaches its theoretical minimum.

[0026] In one embodiment of the present invention, the process of reverse deducing the evolutionary chain based on the generated directed acyclic graph of user interest evolution is as follows: In the graph, the node representing the user's most recent interest state is marked as the current state node. Starting from the current state node, a reverse breadth-first search algorithm is used to traverse the entire graph in the opposite direction of the directed edges. During the traversal, the system records the complete sequence of nodes and edges traversed from the current state node back to any source node representing the initial interest or external trigger. Each such backtracking path is a feasible evolutionary chain. For each recorded feasible evolutionary chain, it is necessary to verify whether the timestamp order of the nodes on the path is consistent with the actual occurrence order in the original behavioral data, filtering out paths with contradictory temporal logic. For each validated feasible evolutionary chain, calculate its pattern confidence and evolutionary rationality score. During the calculation, extract the evolution driving strength weights of all directed edges on the evolutionary chain, perform continuous multiplication of these weights to obtain a basic pattern strength value, combine the global prior distribution probability of each node in the directed acyclic graph of user interest evolution based on historical data, and use Bayesian inference to correct the basic pattern strength value to obtain the pattern confidence of the evolutionary chain. At the same time, obtain the actual behavior time difference of adjacent nodes on the evolutionary chain in the original behavior data, compare this actual time difference with the common time interval of similar interest transfers statistically obtained from the historical group behavior database, and calculate its matching degree as the evolutionary rationality score of the evolutionary chain.

[0027] In practice, the process of reverse derivation based on the directed acyclic graph (DAG) of user interest evolution begins by locating a specific node in the graph. In the DAG, the node representing the user's most recent interest state is marked as the current state node. The current state node typically corresponds to the node with the latest behavior timestamp or a node generated from aggregated recent behavior data. Starting from the current state node, a reverse breadth-first search algorithm traverses the DAG along the reverse direction of the directed edges. The reverse breadth-first search algorithm begins at the current state node, first visiting all predecessor nodes that directly point to the current state node, then visiting the predecessor nodes of these predecessor nodes in sequence, backtracking layer by layer. During the traversal, the system records the complete sequence of nodes and edges traversed from the current state node back to any source node representing the initial interest or external trigger. Source nodes are typically defined as nodes with an in-degree of zero or nodes that satisfy specific origin determination conditions. Each complete backtracking path from the current state node to the source node is a feasible evolutionary chain. It is understandable that for each recorded feasible evolutionary chain, it is necessary to verify whether the timestamp order of the nodes on the path is consistent with the actual occurrence order of the original behavioral data. The verification logic is to ensure that the timestamp of each predecessor node in the path is not later than the timestamp of its successor node, and to filter out paths with contradictory time logic.

[0028] In practice, for each validated feasible evolutionary chain, the pattern confidence and evolutionary rationality score of its path are calculated. For a single feasible evolutionary chain, the evolutionary driving strength weights of all directed edges on the chain are extracted. These weights originate from the edge weights in the directed acyclic graph of user interest evolution. A basic pattern strength value is obtained by continuously multiplying all evolutionary driving strength weights on the feasible chain. This basic pattern strength value represents the product of continuous probabilities of evolution along this path. The basic pattern strength value is then corrected using Bayesian inference, based on the global prior distribution probability of each node in the directed acyclic graph of user interest evolution, to obtain the pattern confidence of the feasible evolutionary chain. The global prior distribution probability can be statistically derived from historical data, representing the prior probability of each interest node appearing among all users. In some embodiments, the pattern confidence... It can be calculated using the following formula: in: The likelihood of evolution along this feasible evolutionary chain can be approximated by the basic mode strength value; This represents the conditional probability of the endpoint interest occurring given this evolutionary path, which can be statistically estimated from historical path data; This represents the global prior distribution probability of the endpoint interest nodes. It can be understood that this formula combines the strength of the path itself with the universality of the endpoint interest to correct and obtain the pattern confidence.

[0029] In practical implementation, the actual behavior time difference corresponding to each pair of adjacent nodes on the feasible evolutionary chain is obtained in the original behavior data. The actual behavior time difference refers to the time interval between two consecutive actions by a user that are associated with these two nodes. This actual behavior time difference is compared with the common time intervals for similar interest transfers statistically derived from the historical group behavior database. The common time intervals for similar interest transfers refer to the statistical distribution (such as mean and variance) of the time intervals spent by a large number of users in history to transfer from the same source interest node to the same target interest node. The matching degree is calculated as the evolutionary rationality score of the feasible evolutionary chain. The matching degree can be calculated based on the probability density of the actual time difference falling in the statistical distribution or the normalized distance from the statistical mean. Optionally, for a feasible evolutionary chain containing k edges, its evolutionary rationality score is... This can be the geometric mean or weighted average of the temporal matching degree of each adjacent node. In some embodiments, the system independently calculates the pattern confidence and evolutionary rationality score for each feasible evolutionary chain, and sorts and filters all feasible evolutionary chains based on these two scores.

[0030] See Figure 3 This is a trend chart of node weights in the evolution of user interests, visually presenting the weight changes from the initial stage to the stable stage of an interest. The weights monotonically increase with the stage of interest evolution, exhibiting an accelerating S-shaped growth characteristic, consistent with the formation pattern of user interests from weak to strong and from temporary to stable. It can be used to locate the current lifecycle stage of a user's interest, providing a basis for refined operations. The slope of the weight change reflects the speed of interest evolution; a larger slope indicates a more significant leap in user interest at that stage, which can be used to predict future interest trends. It verifies the rationality of node weights in the directed acyclic graph of user interest evolution, providing fundamental data support for subsequent reverse BFS evolution chain derivation and pattern confidence calculation. The faster the weight increases, the higher the user's investment in that interest, allowing for early prediction of their subsequent behavioral preferences and improving the accuracy of recommendations / advertising. Comparing node weight changes under different operational strategies allows for a quantitative evaluation of the effectiveness of operational actions such as "interest activation" and "preference reinforcement."

[0031] In one embodiment of the present invention, when integrating all feasible evolutionary chains after sorting and filtering, feasible evolutionary chains with both pattern confidence and evolutionary rationality scores higher than their respective set thresholds are merged. During the merging process, identical nodes and edges between different evolutionary chains are identified and merged. Using shared high-frequency nodes in the network as connection hubs, multiple feasible evolutionary chains are woven into a directed acyclic mesh topology. Based on the typical time stages in which nodes appear in numerous evolutionary chains, nodes in this mesh topology are automatically classified. For example, nodes appearing in early time windows are classified into the potential interest layer, nodes appearing in mid-term time windows are classified into the development and evolution layer, and nodes appearing in recent time windows are classified into the stable state layer, thereby forming a hierarchical interest tracing network. To perform subgraph segmentation on this hierarchical interest tracing network, it is necessary to calculate the pattern correlation degree between any two nodes in the network. This correlation degree is determined by the product of edge weights on the shortest path between nodes and the number of shared neighbor nodes. A community detection algorithm based on modularity optimization is adopted to divide the network nodes according to the calculated pattern correlation degree. The goal is to make the nodes within the same community closely related and the nodes between different communities sparsely related. The subgraph corresponding to each community is extracted from the original network, and each subgraph constitutes an interest cluster. A descriptive topic label is automatically assigned to each interest cluster based on the entity type distribution of its internal nodes.

[0032] In practical implementation, all feasible evolutionary chains, after being sorted and filtered, are integrated to form a hierarchical interest tracing network. This process first merges feasible evolutionary chains whose pattern confidence and evolutionary rationality scores are both higher than their respective set thresholds. The set thresholds can be dynamically adjusted based on historical data distribution or business needs. During the merging process, the system identifies and merges identical nodes and edges between different feasible evolutionary chains. For example, if two feasible evolutionary chains both contain a directed edge from the node "machine learning" to the node "deep learning," these two edges will be merged into one edge, with its weight possibly being the average or maximum value. High-frequency nodes shared in the network serve as the network hub. High-frequency nodes refer to nodes that appear repeatedly in multiple feasible evolutionary chains. These high-frequency nodes are used as connection points to weave multiple feasible evolutionary chains into a directed acyclic mesh topology. Based on the typical time stages in which nodes appear in numerous feasible evolutionary chains, nodes in the mesh topology are automatically classified. For example, nodes appearing in the early time window are classified into the potential interest layer, nodes appearing in the mid-term time window are classified into the development and evolution layer, and nodes appearing in the recent time window are classified into the stable state layer, thus forming a hierarchical interest tracing network with clear levels.

[0033] In practical implementation, subgraph segmentation of the hierarchical interest tracing network requires calculating the pattern correlation degree between any two nodes in the network. This pattern correlation degree is determined by the edge weights on the shortest path between nodes and the number of shared neighbors. In some embodiments, the pattern correlation degree... It can be calculated using the following formula: in: Represents a node With nodes The degree of correlation between patterns It is a node To the node The shortest path (based on edge weights). It is a path Above Evolutionary driving strength weights This indicates a series of multiplication operations. and Representing nodes respectively and nodes The set of neighboring nodes, It refers to the number of shared neighbors. It is the size of the union. and It is a coefficient that adjusts the weights of the two items, and A modularity-based community detection algorithm is employed to partition network nodes based on calculated pattern correlations, aiming to ensure that nodes within the same community are closely related while nodes between different communities are sparsely related. A subgraph corresponding to each community is extracted from the hierarchical interest tracing network, and each subgraph constitutes an interest cluster. Each interest cluster is automatically assigned a descriptive topic label based on the entity type distribution of its internal nodes. For example, if most nodes in an interest cluster belong to the entity types "programming languages" and "development frameworks," the topic label "software development" might be assigned.

[0034] In practical implementation, calculating pattern correlation may require handling multi-path cases between node pairs. Optionally, the shortest path between nodes can be defined as the path with the largest edge weight product or the path with the highest average edge weight. In some embodiments, the system constructs a pattern correlation matrix for the hierarchical interest tracing network to store the correlation values ​​between all node pairs, allowing the community detection algorithm to access it quickly. It can be understood that after the community detection algorithm divides the network into communities, each interest cluster corresponds to a connected subgraph. This subgraph may span multiple levels, but the internal nodes are highly correlated in terms of pattern correlation. Optionally, when assigning topic labels to interest clusters, the entity types of all nodes within the interest cluster can be statistically analyzed, and the entity type with the highest frequency can be selected as the topic label, or a generalization can be made by combining the hypernyms of entity types in the domain knowledge graph. Refer to Table 1, which shows a simplified example illustrating the data comparison between node pairs when calculating pattern correlation.

[0035] Table 1: Example Table of Node Pair Pattern Association Degree Calculation In Table 1, the shortest path edge weight product is the continuous product of all edge weights on the path, the shared neighbor ratio is the ratio of the number of shared neighbors to the size of the union, and the pattern correlation degree... Calculate using the above formula, where and Table 1 shows the differences in pattern correlation between different node pairs, which will be used for subsequent community segmentation.

[0036] See Figure 4 This is a hierarchical interest tracing network node distribution diagram, showing the node distribution across the three interest levels in the network. Overall, the stable state layer has the most nodes, indicating that the user's current interests have reached a relatively mature and stable state. The development and evolution layer and the potential interest layer have roughly equal numbers of nodes, representing active interest development paths and unactivated potential interest points. The substantial number of nodes in the potential interest layer and the development and evolution layer indicates the existence of expandable interest directions, allowing for targeted interest awakening and scenario extension to uncover new user consumption / interaction potential. Clearly defining the node scale at each level allows for more accurate simulation of the diffusion of influence from the potential / development layer to the stable state layer in independent cascading models. The stable state layer has the highest proportion of nodes, indicating that the user's core interests are highly solidified, making it suitable for in-depth operation and precise conversion.

[0037] In one embodiment of the present invention, an influence propagation analysis across time windows is performed for each segmented interest cluster. Within a single interest cluster, interest nodes located in the stable state layer are designated as evaluation target nodes. An independent cascade model is applied to simulate the propagation process of causal influence. Starting from each candidate trigger node in the interest cluster located in the potential interest layer or the development and evolution layer, the influence is simulated to gradually diffuse along directed edges towards the evaluation target node. During the simulated diffusion, the strength of the influence is multiplied by the evolution driving strength weight of that edge for attenuation as it passes through each directed edge. All influence values ​​that successfully reach the evaluation target node are accumulated. The cumulative influence value contributed by each candidate trigger node is normalized to the proportion of the total influence value contributed by all candidate trigger nodes. This proportion is the contribution weight of that candidate trigger node. When generating a report based on the calculated contribution weights, interest trigger nodes with contribution weights greater than a set contribution weight threshold are selected and marked as key interest drivers. The report queries a pre-defined domain knowledge graph to find the related superordinate concepts, synonyms, attributes, and typical application scenarios for each key interest driver. All key interest drivers are sorted from highest to lowest contribution weight. For each key interest driver, its interest cluster, calculated contribution weight, and related concepts and attributes retrieved from the knowledge graph are recorded in a structured manner, and potential user intentions or needs are deduced accordingly. The selected and sorted key interest drivers and their related information are organized and populated according to a pre-defined user profile report template to generate the final structured dynamic user profile report.

[0038] In practice, a cross-time-window influence propagation analysis is performed for each segmented interest cluster. Within a single interest cluster, interest nodes located in the stable state layer are designated as evaluation target nodes, representing the user's current stable or mature interest state. An independent cascade model is applied to simulate the propagation process of causal influence in the directed network within the interest cluster. Starting from each candidate trigger node in the potential interest layer or the development and evolution layer of the interest cluster, the process of its influence gradually spreading along directed edges towards the evaluation target node is simulated. Candidate trigger nodes are early or mid-stage interest nodes that may influence the final interest state. During the simulated diffusion, the strength of the influence is multiplied by the evolutionary driving strength weight of that directed edge for attenuation as it passes through each directed edge. The total influence values ​​that successfully reach the evaluation target node are accumulated. This accumulated value represents the total influence generated on the evaluation target node from a specific candidate trigger node through all possible paths. It is understandable that simulations of independent cascade models are typically performed in discrete time steps. At each time step, each activated node (i.e., one that has received influence) attempts to activate its unactivated subsequent nodes with a certain probability. In this implementation, the activation probability is defined by the evolution drive strength weight of the directed edges.

[0039] In practice, the process of quantifying contribution weights is based on multiple simulations of the independent cascade model. The cumulative influence value contributed by each candidate triggering node is normalized to its proportion of the total influence value contributed by all candidate triggering nodes; this proportion is the contribution weight of that candidate triggering node. Specifically, for an interest cluster containing m candidate triggering nodes, the cumulative influence value of each candidate triggering node v on the evaluation target node d is obtained through simulation. Then the contribution weight of node v The calculation formula is as follows: in: This represents the average (or sum) of the influence values ​​that successfully propagate from the candidate trigger node v to the evaluation target node d in multiple independent cascading simulations. This represents the sum of the cumulative influence values ​​of all m candidate triggering nodes within this interest cluster. Contribution Weight It is a value between 0 and 1, reflecting the relative share of importance of node v in the formation of the final interest state among all historical triggers.

[0040] In practical implementation, the process of generating a structured dynamic user profile report based on the calculated contribution weights first requires filtering out interest trigger nodes with contribution weights greater than a set contribution weight threshold and marking them as key interest-driving factors. The contribution weight threshold can be set according to the actual application scenario, such as 0.1 or 0.05. In a pre-defined domain knowledge graph, information on the parent concepts, synonyms, attributes, and typical application scenarios associated with each key interest-driving factor is queried. For example, if the key interest-driving factor is "convolutional neural network," its parent concepts might include "deep learning model" and "neural network," its synonyms might be "CNN," its attributes might include "suitable for image recognition," and its typical application scenarios might include "computer vision." All key interest-driving factors are sorted from high to low according to their contribution weights, and the sorted list intuitively shows the importance order of the core factors driving the user's current interest state. For each key interest-driving factor, its interest cluster, calculated contribution weight, associated concepts and attributes retrieved from the knowledge graph are structurally recorded, and potential user intentions or needs are inferred accordingly. For example, a highly weighted "convolutional neural network" node, combined with its "artificial intelligence" interest cluster and "image recognition" attribute, can infer that the user may have the intention of "learning computer vision technology" or "developing image processing projects".

[0041] See Figure 5This is a graph analyzing the contribution weights of candidate trigger nodes, visually presenting the contribution weight of each candidate trigger node to the final stable interest state, with 0.1 as the key screening threshold. Machine learning, neural networks, data mining, and big data all have contribution weights exceeding the threshold of 0.1 and are marked as key interest-driving factors; reinforcement learning and recommendation systems are exactly equal to the threshold and can be considered edge nodes. Overall, it exhibits a "head-concentration" characteristic, with machine learning and neural networks contributing nearly half of the total influence, making them the core source driving the user's current interest. This strictly corresponds to the step of "screening interest trigger nodes with contribution weights greater than the set threshold" in the implementation example, providing direct evidence for generating structured user profile reports. It visually verifies the simulation results of the independent cascade model and quantifies the diffusion effect of the influence of potential / development layer nodes on the stable state layer. High-weight nodes can be directly used as core topic labels for the corresponding interest clusters, improving the interpretability of subgraph segmentation.

[0042] The above embodiments are only used to illustrate the technical methods of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical methods of the present invention without departing from the spirit and scope of the technical methods of the present invention.

Claims

1.A knowledge graph-based user portrait construction method, characterized in that, include: Based on the target user's original behavioral data across multiple information sources, a directed acyclic graph of user interest evolution is generated. Based on the aforementioned directed acyclic graph of user interest evolution, all feasible evolutionary chains from the current interest state to the historical trigger node are derived in reverse. For each of the feasible evolutionary chains, calculate the pattern confidence and evolutionary rationality score of its path, and sort and filter all feasible evolutionary chains based on the pattern confidence and evolutionary rationality score; Integrate all feasible evolutionary chains after sorting and filtering to form a hierarchical interest tracing network. In the hierarchical interest tracing network, each node represents an interest entity or behavioral state, and each edge represents an evolutionary driving relationship. The hierarchical interest tracing network is segmented into subgraphs to decompose the complex network structure into multiple interest clusters, each focusing on a core interest topic. For each of the interest clusters, perform cross-time window influence propagation analysis to quantify the contribution weight of each interest triggering node to the final interest state; Based on the contribution weight of each interest-triggered node and combined with the domain knowledge graph, a structured dynamic user profile report is generated. 2.The knowledge graph-based user portrait construction method of claim 1, wherein, The process of generating a directed acyclic graph of user interest evolution based on the target user's original behavioral data across multiple information sources includes: Receive raw behavioral data of target users on multiple information sources, including search queries, page browsing, content interaction, social relationships and transaction records; Perform time-series cleaning and entity linking operations on the raw behavioral data to identify core entities and associated events in the user behavior sequence; Based on the identified core entities and associated events, a multidimensional behavioral feature tensor is constructed. The multidimensional behavioral feature tensor is used to represent the potential associations between user behavior, entity attributes and contextual information. The multidimensional behavioral feature tensor is input into a learnable behavioral pattern discovery network. The learnable behavioral pattern discovery network generates a directed acyclic graph of user interest evolution by iteratively pruning accidental associated edges and strengthening stable pattern paths. The step of performing time-series cleaning and entity linking operations on the original behavioral data to identify core entities and associated events in the user behavior sequence includes: Natural language processing is performed on the search query and the content interaction to extract named entities, keywords and sentiment tendencies to form a standardized intent encoding sequence; The page browsing and transaction records are segmented and pattern recognized to extract meaningful session fragments and behavioral patterns, and key behavioral events are located from them. User nodes, contact nodes, and relationship types are extracted from the social relationships to construct a local social subgraph; The standardized intent encoding sequence, the key behavioral events, and the local social subgraph are aligned and merged according to a unified timeline; Using the domain knowledge graph, entity disambiguation and linking operations are performed on the merged data, mapping text mentions and behavioral events to unique entity identifiers in the knowledge graph, thereby completing the identification of the core entities and the associated events. 3.The knowledge graph-based user portrait construction method of claim 2, wherein, The multidimensional behavioral feature tensor is input into a learnable behavioral pattern discovery network. This network generates a directed acyclic graph (DAG) of user interest evolution by iteratively pruning accidental edges and reinforcing stable pattern paths, including: Initialize a fully connected behavior graph structure, wherein each feature dimension in the multidimensional behavior feature tensor corresponds to a node in the graph; Using an attention-based structure learner, the existence probability and directional weight of each edge in the fully connected behavioral graph structure are learned by optimizing an objective function that includes sparsity constraints and acyclicity penalties. During training, the structure learner outputs an adjacency matrix in each round and calculates the directed acyclicity of the current graph structure based on the adjacency matrix; The acyclicity penalty is dynamically increased based on the degree of violation of directed acyclicity, and edges that form loops are removed, while stable associated edges with weights exceeding a threshold are retained. When the output of the structure learner changes below the preset tolerance, the model is determined to have converged, and the final adjacency matrix structure is transformed into the user interest evolution directed acyclic graph, where the edge weights represent the evolution driving strength. 4.The knowledge graph-based user portrait construction method of claim 3, wherein, Based on the aforementioned directed acyclic graph of user interest evolution, all feasible evolutionary chains from the current interest state to historical trigger nodes are derived in reverse, including: In the directed acyclic graph of user interest evolution, the node representing the user's interest state at the most recent time point is marked as the current state node; Starting from the current state node, the reverse breadth-first search algorithm is used to traverse the directed acyclic graph of user interest evolution in the opposite direction of the directed edges; During the traversal, the system records the complete sequence of nodes and edges traversed from the current state node back to any source node representing the initial interest or external trigger. Each such backtracking path is a feasible evolution chain. For each of the feasible evolutionary chains, verify whether the timestamp order of its nodes is consistent with the order of occurrence of the original behavioral data, and filter out paths with contradictory time logic. 5.The knowledge graph-based user portrait construction method of claim 4, wherein, For each of the aforementioned feasible evolutionary chains, calculate its path pattern confidence and evolutionary plausibility score, including: For a single feasible evolutionary chain, extract the evolution driving strength weights of all directed edges on the feasible evolutionary chain; A basic mode strength value is obtained by continuously multiplying all evolution driving strength weights on the feasible evolution chain. By combining the global prior distribution probability of each node in the directed acyclic graph of user interest evolution, Bayesian inference is used to correct the basic pattern strength value to obtain the pattern confidence of the feasible evolution chain. Obtain the actual time difference of adjacent nodes on the feasible evolutionary chain in the original behavioral data; The actual time difference of this behavior is compared with the common time intervals of interest transfers of the same type obtained from the statistics of historical group behavior, and the matching degree is calculated as the evolutionary rationality score of the feasible evolutionary chain. 6.The knowledge graph-based user portrait construction method of claim 5, wherein, The integration of all feasible evolutionary chains after sorting and filtering forms a hierarchical interest tracing network, including: Merge all feasible evolutionary chains in which both the pattern confidence score and the evolutionary rationality score are higher than their respective set thresholds; During the merging process, identify and merge identical nodes and edges between different feasible evolutionary chains; Using shared high-frequency nodes as the network hub, multiple feasible evolution chains are woven into a directed acyclic mesh topology. Based on the typical time stages in which nodes appear in the evolutionary chain, the nodes in the mesh topology are automatically classified into a potential interest layer, a development and evolution layer, and a stable state layer, thereby forming the hierarchical interest tracing network with a clear hierarchy. 7.The knowledge graph-based user portrait construction method of claim 6, wherein, The hierarchical interest tracing network is subgraph segmented to decompose the complex network structure into multiple interest clusters, including: Calculate the pattern correlation degree between any two nodes in the hierarchical interest tracing network. The pattern correlation degree is determined by the edge weight on the shortest path between the nodes and the number of shared neighbors. A modularity-based community detection algorithm is adopted to divide network nodes according to the pattern correlation degree, so that after the division, the nodes within the same community are closely related, while the nodes between different communities are sparsely related. Each subgraph corresponding to a community is extracted from the hierarchical interest tracing network, and the subgraph constitutes an interest cluster. For each interest cluster, a descriptive topic tag is automatically assigned based on the entity type distribution of its internal nodes. 8.The knowledge graph-based user portrait construction method of claim 7, wherein, For each of the interest clusters, a cross-time-window influence propagation analysis is performed to quantify the contribution weight of each interest-triggered node to the final interest state, including: Within a single interest cluster, the interest node located in the stable state layer is designated as the evaluation target node; An independent cascade model is applied to simulate the propagation of causal influence. Starting from each candidate trigger node in the interest cluster located in the potential interest layer or the development and evolution layer, the process of its influence spreading along the directed edge to the evaluation target node is simulated. During the simulated diffusion, the strength of an influence is multiplied by the evolution driving strength weight of the directed edge for each directed edge it passes through to decay. Finally, all influence values ​​that successfully reach the evaluation target node are accumulated. The cumulative influence value contributed by each candidate triggering node is normalized to the proportion of the total influence value contributed by all candidate triggering nodes. This proportion is the contribution weight of the candidate triggering node. 9.The knowledge graph-based user portrait construction method of claim 8, wherein, Based on the contribution weight of each interest-triggered node and combined with the domain knowledge graph, a structured dynamic user profile report is generated, including: Filter out interest-triggered nodes whose contribution weight is greater than a set contribution weight threshold, and mark them as key interest-driving factors; In the domain knowledge graph, query the superordinate concepts, synonyms, attributes, and typical application scenarios associated with each key interest-driven factor; All key interest drivers are sorted from highest to lowest according to their contribution weight; For each key interest-driven factor, the interest cluster to which it belongs, the calculated contribution weight, the related concepts and attributes retrieved from the knowledge graph are recorded in a structured manner, and potential user intentions or needs are derived accordingly. The screened key interest driving factor and the related information are organized and filled according to a preset user portrait report template, and a final structured dynamic user portrait report is generated. 10.A knowledge graph based user portrait construction system, comprising a memory, a processor, and a computer program stored in the memory and running on the processor, and characterized in that, The processor, when executing the computer program, implements the steps of the user portrait construction method based on the knowledge graph according to any one of claims 1 to 9.

Images