User interest matching network marketing system fused with knowledge graph

Abstract

The application discloses a user interest matching network marketing system fused with a knowledge graph, and relates to the technical field of computers. The system is used to solve the marketing decision deviation problem caused by the lack of dynamic context perception and causal reasoning capability in existing user interest modeling. First, a dynamic knowledge graph is constructed through distributed event stream processing, and the real-time context of users is structured as a graph sequence with timestamps. Second, a time series convolution graph network is used to extract user interest evolution features, and the influence weight of the context node is quantified through a multi-head attention mechanism. Then, a causal inference model is constructed to eliminate exposure and location bias, and an counterfactual learning framework is used to extract user essential interest representation. Finally, the knowledge graph structure is dynamically optimized based on a gradient propagation algorithm, forming a closed-loop learning system from decision-making to perception. The application realizes deep understanding and accurate matching of user interest, and improves the accuracy and interpretability of the marketing system.

Description

Technical Field

[0001] This invention relates to the field of computer technology, specifically to a user interest matching online marketing system that integrates knowledge graphs. Background Technology

[0002] In today's digital environment, users are constantly overwhelmed by information, exposed to massive amounts of product and marketing information every moment. Simultaneously, users' consumption decision-making logic is shifting from passive reception to active seeking; they expect service platforms to act like personal intelligent advisors, understanding not only their explicit historical preferences but also their immediate intentions and deeper needs in different contexts. This upgrade in marketing paradigms—from "one-size-fits-all" to "personalized experiences," and then to "one-person-one-person"—poses unprecedented challenges for businesses in leveraging data intelligence to achieve precise user reach and efficient resource allocation.

[0003] However, existing mainstream user interest modeling and matching methods suffer from several substantial technical bottlenecks. First, traditional methods, such as collaborative filtering, are severely constrained by data sparsity and cold start problems. Their recommendation logic is based on the statistical correlation of user-item co-occurrence matrices. For example, they cannot understand the semantic logic behind "why a user who bought trekking poles would also need a waterproof jacket," resulting in poor interpretability of recommendation results and difficulty in covering long-tail interests. Second, while content- or tag-based models can alleviate cold start, their interest representations are often flat and isolated, simplifying user interests to a set of discrete tags. They cannot depict the complex networked relationships between interest points, such as the dynamic excitation relationship between "weekend," "sunny weather," and "outdoor sports equipment." Furthermore, most existing methods are built on static data snapshots, resulting in lagging interest model updates. They struggle to capture and respond to real-time interest drift caused by changes in user context (such as geographical location, time, and social hotspots), ultimately leading to a disconnect between marketing decisions and the user's actual state, resulting in wasted resources and a decline in user experience. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention provides a user interest matching online marketing system that integrates knowledge graphs, thus solving the problems mentioned in the background.

[0005] To achieve the above objectives, this invention provides the following technical solution: a user interest matching online marketing system integrating knowledge graphs, comprising the following modules: a context graph construction module, an interest evolution calculation module, a causal debiasing module, and an intelligent decision-making module; the context graph construction module is used to receive user behavior event streams and environmental context data in real time through distributed event stream processing, dynamically maintain the knowledge graph structure based on a streaming graph update algorithm with time windows, and activate preset product scenario association rules when a user is detected entering a specific geofence, generating a dynamic knowledge graph incremental update sequence with event timestamps; the interest evolution calculation module is used to extract multiple user interests within a sliding time window based on the dynamic knowledge graph incremental update sequence through a temporal convolutional graph network. The system employs a skip-neighbor structure change feature, utilizing a multi-head graph attention mechanism to quantify the contribution weights of different contextual nodes to the user's interest state, and outputs a time-dependent user interest embedding vector. A causal debiasing module is used to construct a causal inference model incorporating exposure and position biases based on the user interest embedding vector. It jointly trains the sample distributions of the processing and control groups using an end-to-end counterfactual learning framework, and uses a gradient inversion layer to eliminate the confusion caused by selection bias on the interest vector, outputting a debiased representation of the user's essential interest. An intelligent decision-making module performs multi-dimensional correlation analysis between the user's essential interest representation and real-time feedback data, dynamically adjusting node attributes and edge weights in the knowledge graph of the contextual graph construction module using a gradient propagation-based graph structure optimization algorithm.

[0006] Furthermore, the specific process of dynamically maintaining the knowledge graph structure through distributed event stream processing to receive user behavior event streams and environmental context data in real time, and using a streaming graph update algorithm based on time windows, is as follows: A distributed event stream processing pipeline is established to continuously receive multi-source data streams. Event time alignment processing eliminates the impact of out-of-order events caused by network transmission. A sliding time window mechanism based on event time divides the continuous data stream into fixed-duration time segments. Within each time window, user behavior events are parsed in real time, extracting entity and relationship information. The extracted entities and relationships are incrementally updated into the knowledge graph using a streaming graph update algorithm. Version management of the knowledge graph is maintained, and each update operation is timestamped. Based on newly arrived data, the relationships between nodes are adjusted in real time to achieve dynamic evolution of the graph structure, ensuring that the knowledge graph accurately reflects the user's latest state.

[0007] Furthermore, the specific process of activating preset product scenario association rules and generating a dynamic knowledge graph incremental update sequence with event timestamps when a user is detected entering a specific geofence is as follows: Real-time monitoring of user location information and identification of whether the user has entered a preset area of ​​interest through a geofence detection algorithm; when the user enters a specific geofence, triggering the scenario recognition engine and matching the corresponding business scenario type; activating the preset product scenario association rule library, retrieving relevant product categories and attribute features under the scenario, and dynamically constructing a user-scenario-product association subgraph based on the scenario association rules; generating an incremental update operation sequence with time-series markers for each association subgraph, and sorting the incremental update sequence according to the event timestamps to ensure time sequence consistency; establishing a graph smooth transition mechanism when switching scenarios, avoiding abrupt changes in association relationships through gradual updates, and maintaining the stability of the knowledge graph.

[0008] Furthermore, based on the incremental update sequence of the dynamic knowledge graph, the specific process of extracting the multi-hop neighbor structure change features of users within a sliding time window using a temporal convolutional graph network is as follows: A temporal convolutional graph network architecture is constructed and a graph convolutional layer with temporal awareness is designed. The incremental update sequence of the knowledge graph is processed by temporal slicing within the sliding time window. Graph convolution operations are applied to each temporal slice to capture local neighbor structure features. Multi-layer graph convolutional networks are used to propagate and aggregate multi-hop neighbor information, extracting deep graph structure features. A temporal dependency modeling module is designed to capture the structural evolution patterns between adjacent time slices. Dilated temporal convolutions are used to expand the receptive field to capture long-distance temporal dependencies. A comprehensive analysis of the graph structure change patterns of users within continuous time segments is conducted to form a user behavior representation containing multi-scale temporal features.

[0009] Furthermore, the specific process of quantifying the contribution weights of different context nodes to the user's interest state using a multi-head graph attention mechanism and outputting a time-dependent user interest embedding vector is as follows: A multi-head graph attention network is constructed, and independent feature transformation parameters are assigned to each attention head. The attention coefficients between the user node and each context node are calculated to characterize their influence on the user's interest state. A hierarchical attention mechanism is designed to calculate attention weights at both the node and graph levels, and a time decay factor is introduced to dynamically adjust the contribution weights of historical context nodes to the current interest state. A comprehensive context influence weight distribution is obtained by splicing and fusing the multi-head attention outputs. Based on the attention weights, the features of neighboring nodes are weighted and aggregated to update the embedding representation of the user node. The temporal structure features and attention-weighted features are deeply fused to generate a time-dependent user interest embedding vector.

[0010] Furthermore, based on user interest embedding vectors, the construction logic of a causal inference model incorporating exposure bias and location bias is as follows: A causal graph structure is established based on user interest embedding vectors, incorporating exposure mechanisms and location information as intervention variables into the causal graph model, defining the causal relationship path between user characteristics, product attributes, and contextual factors; a structured causal model framework is designed, using a learnable parameter matrix to characterize the influence mechanism of different bias factors on user interests, establishing a causal transmission chain from exposure events to user feedback; a latent outcome predictor based on deep neural networks is constructed to estimate user interest states with and without exposure intervention, quantifying the confusion effect of exposure bias and location bias on interest vectors; a dynamic causal graph update mechanism is designed to adjust the causal graph structure parameters based on real-time feedback data, ensuring that the causal inference model can adapt to constantly changing user behavior patterns.

[0011] Furthermore, the specific process of jointly training the sample distributions of the treatment and control groups using an end-to-end counterfactual learning framework, and using a gradient inversion layer to eliminate the confusion effect of selection bias on the interest vector, and outputting the debiased user interest representation is as follows: An end-to-end counterfactual learning network architecture is constructed, with a dual-branch structure to process the sample data of the treatment and control groups respectively. A bias-invariant feature representation is learned through a parameter sharing mechanism. A gradient inversion layer is introduced into the feature extraction layer, inverting the gradients of bias-related features during backpropagation, forcing the network to learn user interest representations independent of bias factors. An adversarial regularization loss function is designed to eliminate the influence of selection bias by minimizing the difference in feature distributions between the treatment and control groups, while maintaining the discriminative ability of the user interest representation. A multi-task learning strategy is adopted to jointly optimize the counterfactual prediction task and the bias elimination task, gradually separating the user's essential interest from bias factors through iterative training, and outputting the user interest representation vector.

[0012] Furthermore, the intelligent decision-making module includes the following steps: deeply correlate and mine the user's essential interest representation with real-time feedback data, and identify key factors affecting changes in user interest through correlation analysis and causal inference; design a graph structure optimization algorithm based on gradient propagation, construct a differentiable graph neural network computation graph, and use user feedback signals as supervision signals to guide the optimization process of the knowledge graph; calculate the gradients of knowledge graph node attributes and edge weights through backpropagation algorithm, update the graph structure parameters using an adaptive optimization algorithm, and transmit the optimized knowledge graph parameters to the context graph construction module in real time.

[0013] The present invention has the following beneficial effects:

[0014] (1) The user interest matching online marketing system integrating knowledge graphs effectively overcomes the shortcomings of existing methods in dynamic context perception and deep semantic understanding through the synergistic effect of the context graph construction module and the interest evolution calculation module. The system can capture the user's geographical location, time and environmental state in real time and integrate them into the knowledge graph in a structured way, so that user interest modeling is no longer limited to historical behavior statistics, but has the ability to dynamically respond to real-time contexts. At the same time, the application of temporal convolutional graph network and multi-head graph attention mechanism enables the system to capture the transmission and evolution path of user interests from the evolution of knowledge graphs, realize fine-grained differentiation and fusion representation of users' short-term contextual interests and long-term stable preferences, and thus generate dynamic interest vectors that better reflect the user's true intentions.

[0015] (2) The user interest matching online marketing system integrating knowledge graphs systematically solves the problems of misjudgment of interests and model staticization caused by data bias in existing methods through the closed-loop design of the causal debiasing module and the intelligent decision-making module. The causal inference model and counterfactual learning framework can identify and quantify the confusing effects of systematic biases such as exposure bias and location bias on interest vectors, thereby extracting the essential interest representation that better reflects the user's true preferences. On this basis, the intelligent decision-making module uses a graph structure optimization algorithm based on gradient propagation to link user feedback with the essential interest representation, directly adjusting the underlying structure and parameters of the knowledge graph in reverse, forming a continuous optimization closed loop from decision-making to perception, enabling the entire system to evolve adaptively and continuously improve the accuracy and robustness of interest matching.

[0016] Of course, any product implementing this invention does not necessarily need to achieve all of the advantages described above at the same time. Attached Figure Description

[0017] Figure 1 This invention provides a flowchart of a user interest matching online marketing system that integrates knowledge graphs. Detailed Implementation

[0018] This application's embodiments, through a user interest matching online marketing system that integrates knowledge graphs, solve the problems of delayed marketing decisions, severe bias interference, and insufficient interpretability caused by the lack of dynamic context awareness and causal reasoning capabilities in existing user interest modeling methods.

[0019] The overall concept of the solution in this application embodiment is as follows:

[0020] By constructing a dynamic knowledge graph that integrates real-time contexts and designing a complete technical chain including interest evolution computation, causal debiasing, and graph closed-loop optimization, the system achieves deep understanding and accurate matching of user interests. Specifically, firstly, the system transforms streaming data into a knowledge graph sequence with time-series labels, realizing a dynamic and structured representation of user contexts; secondly, it captures the evolutionary patterns of interests through time-series graph networks and attention mechanisms to generate deep interest vectors; subsequently, it introduces a causal inference and counterfactual learning framework to remove the interference of data bias on interest modeling and extract the essential user interest representations; finally, through a graph structure optimization algorithm based on gradient propagation, user feedback is directly applied to the iterative update of the knowledge graph, forming a self-evolving intelligent decision-making closed loop, thereby continuously improving the accuracy, robustness, and interpretability of the marketing system.

[0021] Please see Figure 1 This invention provides a technical solution: a user interest matching online marketing system integrating knowledge graphs, comprising the following modules: a context graph construction module, an interest evolution calculation module, a causal debiasing module, and an intelligent decision-making module; the context graph construction module is used to receive user behavior event streams and environmental context data in real time through distributed event stream processing, dynamically maintain the knowledge graph structure based on a streaming graph update algorithm with time windows, and activate preset product scenario association rules when a user is detected entering a specific geofence, generating a dynamic knowledge graph incremental update sequence with event timestamps; the interest evolution calculation module is used to extract the user's multi-hop neighbor structure within a sliding time window through a temporal convolutional graph network based on the dynamic knowledge graph incremental update sequence. The system employs a multi-head graph attention mechanism to quantify the contribution weights of different contextual nodes to the user's interest state, outputting a time-dependent user interest embedding vector. A causal debiasing module constructs a causal inference model incorporating exposure and location biases based on the user interest embedding vector. It jointly trains the sample distributions of the processing and control groups using an end-to-end counterfactual learning framework, and uses a gradient inversion layer to eliminate the confusion caused by selection bias in the interest vector, outputting a debiased representation of the user's essential interest. An intelligent decision-making module performs multi-dimensional correlation analysis between the user's essential interest representation and real-time feedback data, dynamically adjusting node attributes and edge weights in the knowledge graph of the contextual graph construction module using a gradient propagation-based graph structure optimization algorithm.

[0022] In this implementation plan, the context graph construction module is used to perform real-time structured modeling of user behavior events and external environment data in online marketing scenarios. Its core technologies are distributed event stream processing and streaming graph update algorithms. Distributed event stream processing refers to the parallel processing of continuously input behavioral data streams using a multi-node computing architecture to support high concurrency and low latency data access. The streaming graph update algorithm is an algorithm mechanism that continuously updates nodes and edges in the knowledge graph based on time windows, ensuring that the knowledge graph can reflect the dynamic changes of users and the environment in real time. The module divides the data stream into continuous segments through a time window mechanism, and each segment updates the entities (such as users, products, and scenarios) and their relationships in the graph. When the system detects that a user has entered a preset "geofence" (i.e., a virtual boundary area defined by GPS coordinates), it automatically activates the product scenario rule base, generating an incremental sequence of knowledge graphs corresponding to the user's location, time, and scenario, enabling the system to have dynamic context awareness capabilities. The technical role of this module is to establish a knowledge graph that can evolve in real time, providing a structured time-series data foundation for subsequent interest evolution calculations. The Interest Evolution Computation Module identifies and models the evolutionary patterns of user interests on a dynamic knowledge graph. Its core technologies include temporal convolutional graph networks and multi-head graph attention mechanisms. Temporal convolutional graph networks introduce a temporal dimension to traditional graph neural networks (GNNs), capturing patterns of graph structure changes over time through convolution operations within a sliding time window, thus reflecting the dynamic evolution of user interests. The multi-head graph attention mechanism assigns different attention weights to different neighbor nodes, measuring the influence of contextual nodes (such as time, location, and product features) on the user's interest state. The module's output, a "user interest embedding vector," is a high-dimensional representation that integrates user behavior, environmental context, and temporal change features, providing input for subsequent causal debiasing and decision optimization. The technology's role is to achieve a shift from static interest analysis to dynamic, multi-contextual interest representation, enabling the system to capture the temporal dependence and evolutionary trends of interests. The Causal Debiasing Module addresses the problem of distorted interest modeling in marketing recommendations caused by systematic biases such as exposure mechanisms or location information. Its key technologies include causal inference models and counterfactual learning frameworks. The causal inference model constructs a causal graph structure, modeling exposure bias (the probability difference of a product being displayed) and location bias (the bias in how display location affects click behavior) as intervention variables to identify and quantify the impact of these biases on user interest vectors. The counterfactual learning framework infers users' true interests by simulating a control scenario where the exposure or location intervention had not occurred, thereby eliminating the confounding effect introduced by biases. Internally, the module employs a gradient inversion layer technique, reversing the gradient signals of bias-related features during backpropagation in the neural network, forcing the model to learn essential interest features unrelated to biases. The technical advantage lies in stripping away the interference of external interventions on the interest model, resulting in a more accurate and interpretable representation of user interests.The intelligent decision-making module is used to adaptively optimize marketing decisions based on the debiased representation of users' essential interests. The module uses a gradient propagation-based graph structure optimization algorithm, employing real-time user feedback (such as clicks, purchases, and dwell time) as a supervisory signal to back-optimize the node attributes and edge weights of the knowledge graph. This algorithm calculates the gradient of the knowledge graph on the decision outcome, identifies key nodes influencing changes in interest (such as a certain type of product or specific scenario), and adjusts the graph structure accordingly, enabling the graph to better reflect the causal logic of user interests. The technology's role is to form a closed-loop system from "interest modeling—recommendation decision—user feedback—graph update," enabling marketing strategies to continuously learn and evolve.

[0023] Specifically, the process of dynamically maintaining the knowledge graph structure through distributed event stream processing to receive user behavior event streams and environmental context data in real time, and using a streaming graph update algorithm based on time windows, is as follows: A distributed event stream processing pipeline is established to continuously receive multi-source data streams. Event time alignment processing eliminates the impact of out-of-order events caused by network transmission. A sliding time window mechanism based on event time divides the continuous data stream into fixed-duration time segments. Within each time window, user behavior events are parsed in real time, extracting entity and relationship information. The extracted entities and relationships are incrementally updated into the knowledge graph using a streaming graph update algorithm. Version management of the knowledge graph is maintained, and each update operation is timestamped. Based on newly arrived data, the relationships between nodes are adjusted in real time to achieve dynamic evolution of the graph structure, ensuring that the knowledge graph accurately reflects the user's latest state.

[0024] In this implementation, a distributed event stream processing system (Kafka Stream, Flink) is used to transform the continuously arriving behavioral and environmental data streams into a dynamically evolving graph structure, thereby forming a real-time contextual knowledge graph. A distributed event stream processing pipeline is established, serving as the data artery of the entire process. It continuously receives data streams from multiple sources (user click streams, location information, environmental sensor data, etc.) and processes them in parallel on multiple nodes, ensuring real-time responsiveness even under high concurrency. Technical logic: An event-driven architecture is used, encapsulating each behavioral event as an independently processable message stream, achieving low-latency, scalable data access. Event time alignment processing is crucial because data transmission may experience delays or out-of-order delivery—a click event might arrive later than its actual occurrence sequence. To ensure temporal consistency, the events are reordered based on their timestamps, rather than by reception time. This process ensures that graph updates strictly follow the actual order of user behavior, avoiding logical inconsistencies. A sliding time window mechanism divides the continuous data stream into fixed-length time windows (e.g., every 5 seconds or every minute), with each window acting as a micro-batch processing unit. A sliding window means that adjacent windows partially overlap, allowing the capture of short-term trend changes without losing continuity. Technical logic: Data is processed using windows as boundaries, enabling continuous system updates while maintaining temporal context. Event parsing and entity relationship extraction: Within each window, the system parses the event content, extracting entities (such as users, products, and scenarios) and relationships (such as "user browsing products" and "user's location") that should be included in the graph. This process is similar to knowledge extraction and is a key step in generating structured knowledge from the data stream. Streaming graph update algorithm and incremental update mechanism: Unlike batch graph reconstruction, incremental updates are used here: only newly added or changed nodes and edges are inserted, modified, or deleted, and timestamps are added during updates to support version tracking and time-series queries. Technical logic: Time is used as the core index to enable the graph to evolve with events. Dynamic evolution and relationship reconstruction of the graph structure: After new data arrives, the system recalculates the relationships between nodes (e.g., a user's new interest behavior is associated with a new product node).

[0025] This means that a knowledge graph is not a static network, but a time-series graph that updates over time. Technical logic: To achieve real-time evolution of the graph—user states, interests, and contexts constantly change at the graph structure level, ensuring the timeliness of semantic associations.

[0026] Specifically, the process of activating preset product scenario association rules and generating a dynamic knowledge graph incremental update sequence with event timestamps when a user is detected entering a specific geofence is as follows: Real-time monitoring of user location information and identification of whether the user has entered a preset area of ​​interest using a geofence detection algorithm; triggering a scenario recognition engine and matching the corresponding business scenario type when the user enters a specific geofence; activating the preset product scenario association rule library, retrieving relevant product categories and attribute features under the scenario, and dynamically constructing a user-scenario-product association subgraph based on the scenario association rules; generating an incremental update operation sequence with time-series markers for each association subgraph, and sorting the incremental update sequence according to event timestamps to ensure temporal consistency; establishing a smooth transition mechanism for the knowledge graph during scenario switching, avoiding abrupt changes in association relationships through gradual updates, and maintaining the stability of the knowledge graph.

[0027] In this implementation scheme, the geofencing detection and scene recognition system determines whether a user has entered a preset area of ​​interest by real-time monitoring of the user's location information stream (such as GPS coordinate stream) and employing a geofencing detection algorithm. The geofencing detection algorithm is based on the Euclidean distance between the user's current location vector and the boundary of the geofencing area. ;in: : Represents the user's current location coordinates; : Represents the coordinates of the center of the nth geofence; : Represents the Euclidean distance between the user and the center of the geofence. When When a user is identified as having entered the fenced area, This is the fence radius threshold. This threshold is automatically determined based on historical location accuracy statistics, i.e.: ;in: : Represents the average value of historical positioning errors; : Represents the standard deviation of the positioning error; : is an adjustment coefficient used to control tolerance; its empirical value range is [1.5, 2.5]. When a user is detected entering the fenced area, the system triggers the scene recognition engine to identify the commercial scene type of the area by matching geographical location with semantic tags, such as "shopping mall," "dining area," or "scenic spot." The product scene association rule activation system retrieves the corresponding scene rules from the product scene association rule library based on the identified scene type. Each rule can be formally represented as: ;in: : indicates the first A set of rules for each scenario; : Represents a product category node; : Represents a product attribute node; : indicates in the scene The association weight between product categories and attributes. Weight coefficient. The determination method is based on statistical correlation and contextual co-occurrence probability: ;in: : indicates in the scene The frequency of joint occurrence of subcategories and attributes; : These represent the individual frequency of the category and attribute in the scene, respectively; This is a scene type correction factor used to adjust the weight differences between scenes. Its value is learned from historical user interaction data. The user scene product association subgraph is built after activating the rule base, and the system determines its value based on the current user. Identity identifiers and identified scenarios and its corresponding set of rules Dynamically generate a subgraph of relationships. The structure of this subgraph can be represented as a set of ternary relations: ;in: User node; : The currently identified scene node; Product category and attribute nodes; Rule activation weight; Event timestamps are used to identify the current update's temporal position. Incremental update sequence sorting and temporal consistency maintenance: To ensure temporal consistency in the graph evolution, the system uses event timestamps... Sort all incremental update operations. If multiple update operations have the same timestamp, sort them according to event priority weight. Secondary sorting is defined as: ;in: : Indicates the real-time weight of user behavior (e.g., the timeliness level of actions such as clicking, browsing, and saving); : Indicates the urgency of the scene transition; : These are weighting coefficients, which are guaranteed through normalization. After sorting, a final time-ordered incremental update sequence is formed to guide the structural evolution of the knowledge graph. V. Smooth Transition and Gradual Update of the Knowledge Graph To avoid abrupt changes in the graph structure caused by scene switching, the system implements a smooth transition and gradual update mechanism when the user transitions from a scene... Switch to new scene In this case, a smooth transition mechanism is employed. Its core idea is to smooth the association weights between old and new scenes through time decay and interpolation, ensuring semantic continuity. The smoothing function can be expressed as: ;in: , indicating time Smooth correlation weights at different times; : These represent the association weights under the old and new scenarios, respectively; Time smoothing coefficient, defined as ,in Indicates the time of leaving the old scene. A threshold for smooth transition duration is set. Through progressive weight interpolation, the system avoids abrupt changes in interest associations during scene switching, ensuring that the knowledge graph maintains structural stability and semantic continuity during dynamic evolution.

[0028] Specifically, based on the incremental update sequence of the dynamic knowledge graph, the process of extracting the multi-hop neighbor structure change features of users within a sliding time window using a temporal convolutional graph network is as follows: A temporal convolutional graph network architecture is constructed, and a graph convolutional layer with temporal awareness is designed. The incremental update sequence of the knowledge graph is processed by temporal slicing within the sliding time window. Graph convolution operations are applied to each temporal slice to capture local neighbor structure features. Multi-layer graph convolutional networks are used to propagate and aggregate multi-hop neighbor information, extracting deep graph structure features. A temporal dependency modeling module is designed to capture the structural evolution patterns between adjacent time slices. Dilated temporal convolutions are used to expand the receptive field to capture long-distance temporal dependencies. A comprehensive analysis of the graph structure change patterns of users within continuous time segments is conducted to form a user behavior representation containing multi-scale temporal features.

[0029] In this implementation, temporal slicing is used to slice the incremental update sequence of the dynamic knowledge graph according to a sliding time window, with each slice representing the graph state within a specific time period. This structurates continuously changing data, enabling the network to sequentially understand the temporal evolution of user behavior. Graph convolution operations extract local neighbor features. Applying graph convolution operations to each time slice aggregates the features of each node with the features of its neighboring nodes, thereby capturing the association information of nodes in the local graph structure. This step reflects the direct interest relationships and neighbor behavior patterns of users within a specific time period. Multi-hop neighbor information propagation and aggregation are achieved by stacking graph convolutional layers to propagate multi-hop neighbor features, enabling nodes to perceive deeper levels of the graph structure. This step can capture the potential transmission paths and indirect associations of user interests, avoiding reliance on only local information. Temporal dependency modeling models the structural changes between adjacent time slices, analyzing the evolution of node features over time. Dilated temporal convolution can enhance the ability to capture long-term dependencies, understanding the dynamic changing trends of user interests over longer periods. Multi-scale feature synthesis analysis integrates short-term local structural features with long-term temporal evolution features to form a multi-scale representation of user behavior. This allows for the simultaneous consideration of users' immediate interest fluctuations and long-term stable preferences, making interest vectors more comprehensive and accurate.

[0030] Specifically, the process of quantifying the contribution weights of different context nodes to the user's interest state using a multi-head graph attention mechanism and outputting a time-dependent user interest embedding vector is as follows: A multi-head graph attention network is constructed, and independent feature transformation parameters are assigned to each attention head. The attention coefficients between the user node and each context node are calculated to characterize their influence on the user's interest state. A hierarchical attention mechanism is designed to calculate attention weights at both the node and graph levels, and a time decay factor is introduced to dynamically adjust the contribution weights of historical context nodes to the current interest state. A comprehensive context influence weight distribution is obtained by concatenating and fusing the multi-head attention outputs. Based on the attention weights, the features of neighboring nodes are weighted and aggregated to update the embedding representation of the user node. The temporal structure features and attention-weighted features are deeply fused to generate a time-dependent user interest embedding vector.

[0031] In this implementation scheme, a multi-head graph attention network is constructed, and independent transformation parameters are assigned to each attention head. This step establishes multiple parallel attention channels, each learning different types or scales of contextual influence patterns in the graph, providing diverse representations for subsequent weighted aggregation. A hierarchical attention mechanism (node-level and graph-level) is designed: first, the influence of each neighbor node on the target user node is calculated at the node level, and then different subgraphs or context groups are weighted at the graph level to reflect the differences in local and global contextual importance. A time decay factor is introduced to reduce the influence of distant historical contexts on the current interest state, ensuring the model's sensitivity to recent contexts. Multi-head output concatenation and weighted aggregation: the outputs of each attention head are concatenated or fused into a unified contextual influence distribution, and then these weights are used to perform a weighted summation of the neighbor node features to update the user node embedding. Deep fusion of temporal structural features and attention-weighted features: the structural temporal features obtained by temporal convolution are fused with the attention-weighted semantic features to obtain the final temporally dependent user interest embedding vector. The single-head local attention score calculates the compatibility score between a user node (initially represented as vector u at the current time step) and a neighbor node (represented as vector v) for the single attention head numbered p. Parameter explanation: : The user-side linear transformation matrix of the p-th attention head; : Linear transformation matrix of the neighboring nodes of the p-th attention head; u: Input feature vector of the current user node; v: Input feature vector of the neighboring nodes; D: Scaling constant of the feature space (consistent with the dimension of the transformed vector), used to stabilize the scale of the dot product value. Attention weights with time decay (node-level normalization): First, all neighbors within the same head are normalized using softmax to obtain the original coefficients, then multiplied by the time decay factor and re-normalized. ; ; Parameter explanation: : The untime-decayed attention coefficient of the p-th head on its neighbor v (softmax output); : The time decay factor of neighbor node v; Time decay rate parameter; : The time interval from the occurrence of the event corresponding to the neighbor node to the current moment; The final attention weights, decaying over time and normalized. (Regarding...) Determination method (example): Estimate the "half-life" h of the situational impact decaying over time using historical data, assuming... ; where h can be obtained by fitting the historical contextual influence's contribution to the target behavior over time (e.g., least squares fitting). This method enables... This directly corresponds to the observed time decay behavior. Multi-head aggregation and concatenation (intra-head weighted aggregation, inter-head concatenation): Each head first performs weighted neighbor feature aggregation, then inter-head concatenation and linear mapping to obtain the intermediate update vector. ; Parameter explanation: :No. The weighted aggregated output vector of each head; The total number of attention heads; : Dimensional concatenation operation; Linear mapping matrix between heads (used for dimensionality reduction or fusion); : User local representation updated based on multi-head attention. About The determination method can be achieved by cross-validation or searching on the validation set (e.g., grid search or Bayesian optimization) to select the optimal number of heads between performance and computational cost. Hierarchical (graph-level) attention reweighting: After completing the node-level update, the combined weights are recalculated at the graph level or scenario group level to reflect the overall importance of different scenario subgraphs. Suppose there are several scenario subgraph group indices... Calculate the graph weights for each group and apply them to the user representation: Parameter explanation: : Contextual subgraph The summary vector (e.g., obtained by averaging or pooling the node representations within the subgraph); : Query vector (learnable parameters) for graph-level attention; :No. Graph-level attention weights for each context subgraph; : in subgraph The user's local representation obtained from the above calculation; User representation after graph-level reweighting. About Method for determining: Treated as learnable parameters, they are optimized along with other parameters during training; their initial values ​​can be randomly initialized or initialized based on the principal component direction to accelerate convergence. Deep fusion (final embedding) with temporal structure features utilizes the temporal structure feature vector given by the temporal convolutional graph network. Representation obtained with graph attention Perform fusion (e.g., gated fusion or element-wise addition via nonlinear mapping): Parameter explanation: Temporal structural feature vectors from temporal convolutional graph networks; : The linear transformation matrix used for fusion; Bias vector; : Non-linear activation functions (e.g., ReLU or tanh); The final output is a time-dependent user interest embedding vector. Regarding the method for determining the fusion weights: a learnable matrix can be used. And optimize through joint training; if manual initialization is required, it can be set as follows: The initial value is slightly greater than The system prioritizes attentional information and then allows the training data to adjust automatically. (Number of attention heads) Selection via cross-validation or validation set grid search; time decay rate. Based on the half-life estimation of the impact of historical context, let... Smoothing / Normalization Matrix With query vector : As a learnable parameter, it is optimized during training; the initial value can be randomly initialized or initialized based on statistics; feature dimension Scaling constant: Usually consistent with the dimension of the transformed vector. Scaling is used to avoid excessively large dot products that could cause softmax saturation. If an "attention threshold" exists to prune low-impact neighbors, a threshold can be set. And determine by minimizing the performance loss through the validation set.

[0032] Specifically, based on user interest embedding vectors, the construction logic of a causal inference model incorporating exposure bias and location bias is as follows: A causal graph structure is established based on user interest embedding vectors, incorporating exposure mechanisms and location information as intervention variables into the causal graph model, defining the causal relationship path between user characteristics, product attributes, and contextual factors; a structured causal model framework is designed, using a learnable parameter matrix to characterize the influence mechanism of different bias factors on user interests, establishing a causal transmission chain from exposure events to user feedback; a latent outcome predictor based on deep neural networks is constructed to estimate user interest states with and without exposure intervention, quantifying the confusion effect of exposure bias and location bias on interest vectors; a dynamic causal graph update mechanism is designed to adjust the causal graph structure parameters based on real-time feedback data, ensuring that the causal inference model can adapt to constantly changing user behavior patterns.

[0033] In this implementation scheme, firstly, the system establishes a causal graph with the user interest embedding vector as the central node. The graph contains three main types of variables: first, intervention variables, including exposure mechanisms (i.e., the system's strategy for determining whether a user sees certain content) and location information (e.g., the product's ranking position on the page); second, mediating variables, such as the user's individual characteristics, the product's visual or semantic attributes, and the external environmental context; and third, outcome variables, i.e., the user's actual feedback behavior (clicks, dwell time, purchases, etc.). Next, the system constructs a structured causal model, the idea of ​​which is to use a parameter matrix to characterize the influence path of different bias factors (such as exposure bias and location bias) on user interest. For example, the parameter matrix can be represented as a learnable set of weights, reflecting the degree to which the bias variables modulate the representation of user interest. This parameter matrix is ​​automatically optimized through model training to fit the actually observed behavioral data. Then, a latent outcome predictor is set within the model. Its role is to simulate two world outcomes: one assuming the user sees a product (exposure), and the other assuming the user does not see it (no exposure). By comparing the differences in the predicted interest states under the two scenarios, the impact of exposure bias can be quantified. Similarly, by simulating different display locations, the effect of location bias can be evaluated. Finally, the system introduces a dynamic causal graph update mechanism. As user behavior and the environment change, the original causal path may become invalid. For example, the user's interest triggering mechanism may change during holidays. The system uses real-time feedback data to continuously update the structure and parameters of the causal graph, ensuring that the model always reflects the latest user behavior patterns, thereby achieving more stable and accurate interest estimation. The key point of this logic is that it is not just about statistical correlation, but explicitly models "causal relationships"—that is, which factors truly lead to changes in user interests. This method can effectively reduce spurious correlations caused by systematic biases such as exposure and ranking in traditional recommendation systems, thus obtaining a more realistic representation of interests.

[0034] Specifically, the process of jointly training the sample distributions of the treatment and control groups using an end-to-end counterfactual learning framework, and using a gradient inversion layer to eliminate the confusion effect of selection bias on the interest vector, outputting the debiased user interest representation, is as follows: An end-to-end counterfactual learning network architecture is constructed, with a dual-branch structure to process the sample data of the treatment and control groups respectively. A bias-invariant feature representation is learned through a parameter sharing mechanism. A gradient inversion layer is introduced into the feature extraction layer, inverting the gradients of bias-related features during backpropagation, forcing the network to learn user interest representations independent of bias factors. An adversarial regularization loss function is designed to eliminate the influence of selection bias by minimizing the difference in feature distributions between the treatment and control groups, while maintaining the discriminative ability of the user interest representation. A multi-task learning strategy is adopted to jointly optimize the counterfactual prediction task and the bias elimination task, gradually separating the user's essential interest from bias factors through iterative training, and outputting the user interest representation vector.

[0035] In this implementation, directly learning users' true interests from biased data (such as exposure or location preferences) is unreliable. Therefore, this step uses counterfactual learning to allow the model to consider both real-world samples (i.e., user behavior after actually seeing the product) and unexposed samples (i.e., situations the user could have seen but didn't), thereby separating the influence of bias and extracting a more objective representation of user interests. The logical process can be divided into the following key steps. First, an end-to-end counterfactual learning network architecture is constructed. This network adopts a two-branch structure, with one branch processing the processing group samples (samples where the user was exposed) and the other branch processing the control group samples (samples that were not exposed). The two branches share the underlying feature extraction parameters to ensure that the interest features learned by the network are consistent across different sample distributions. This parameter-sharing mechanism ensures that the model captures stable, bias-invariant features. Second, a gradient reversal layer (GRL) is introduced into the feature extraction layer. The gradient reversal layer works by not changing the input features during forward propagation, but inverting the gradient sign during backward propagation. For example, when the bias identification branch of the network attempts to identify whether a sample belongs to the "treatment group" or the "control group" based on features, the gradient reversal layer propagates this learning direction back to the backbone network, forcing the backbone network to learn representations independent of the bias variables. This allows the network to automatically eliminate interference from selection bias, thus focusing more on the structural differences in the interests themselves. Furthermore, an adversarial regularization loss function is designed. This can be formally described as: Let the distribution of interest features of the treatment group samples be... The interest feature distribution of the control group samples is as follows The objective of the adversarial regularization term is to minimize the distributional differences between them, which can be expressed as: ;in, The function representing the distribution difference can be the maximum mean difference or the distribution distance based on an adversarial discriminator. : The distribution of interest features output by the feature extractor for the processed group samples; The interest feature distribution output by the feature extractor from the control group samples; by minimizing The model forces the feature distributions of the two sample groups to converge, thereby eliminating selection bias. Next, the counterfactual prediction task and the bias elimination task are jointly optimized. Specifically, the network adopts a multi-task learning framework, combining two objective functions: one is the user interest prediction loss. First, it is used to maintain the discriminative ability of interest features; second, the aforementioned adversarial loss. This is used to suppress the influence of bias. The overall optimization objective can be written as: ;in, This is a weighting coefficient used to balance the relative importance of interest discrimination and bias elimination. This coefficient can be determined using a cross-validation strategy: by testing different... The prediction accuracy and distribution variability under different values ​​are considered, and the optimal value that simultaneously ensures prediction performance and distribution balance is selected. Finally, interest and bias factors are gradually separated through iterative training. In each round of training, the model is updated alternately between the treatment group and the control group samples, allowing the backbone network to gradually approach the bias-independent interest representation space. After several rounds of convergence, the output user interest embedding vector represents the user's essential interest, a neutral representation that removes the influence of exposure and selection bias.

[0036] Specifically, the intelligent decision-making module includes the following steps: deeply correlate and mine the user's essential interest representation with real-time feedback data, and identify key factors affecting changes in user interest through correlation analysis and causal inference; design a graph structure optimization algorithm based on gradient propagation, construct a differentiable graph neural network computation graph, and use user feedback signals as supervision signals to guide the optimization process of the knowledge graph; calculate the gradients of knowledge graph node attributes and edge weights through backpropagation algorithm, update the graph structure parameters using an adaptive optimization algorithm, and transmit the optimized knowledge graph parameters to the context graph construction module in real time.

[0037] In this implementation scheme, firstly, a deep correlation mining is performed between the user's essential interest representation and real-time feedback data. The system first obtains the debiased user interest vector (representing the user's true preferences) and real-time feedback data (such as clicks, dwell time, and purchase behavior), and identifies behavioral patterns highly correlated with interest changes by calculating the correlation index between the two. For example, the Pearson correlation coefficient or mutual information method can be used to quantify the statistical dependence between feedback features and interest changes, thereby screening out potential influencing factors. Next, the system introduces a causal inference mechanism, identifying the true driving factors rather than superficially related interfering variables by analyzing the causal path of "whether changes in several feedback behaviors lead to changes in interest state." This step ensures that the basis for subsequent optimization has causal validity, rather than misleading statistical correlation. Secondly, a graph structure optimization algorithm based on gradient propagation is designed. This step transforms the knowledge graph into a differentiable computational structure so that it can be optimized through the backpropagation algorithm. Specifically, the node attribute vectors and edge weights in the knowledge graph are parameterized to construct a differentiable graph neural network (GNN) model. Let the node attribute matrix be... The edge weight matrix is superscript The graph represents the first time. The state of the next iteration. For each node Its update in GNN can be represented as: ;in: :node In the Embedded representation at the next iteration; :node The set of neighboring nodes; :node and The learnable edge weight coefficients between them represent the strength of their semantic or behavioral association; : Node feature transformation matrix, which controls the aggregation method of neighbor information; : Self-loop transformation matrix, used to preserve the influence of node's own characteristics during the update; A non-linear activation function is used to enhance the model's expressive power. Next, user feedback signals are used as supervision signals to guide the optimization direction. Assume the model generates the following prediction output based on the current map structure: If the user's actual feedback is y, then the supervised loss function is defined as follows. ; represents the mean squared error between the system's prediction and the actual feedback. To enable the graph structure to adaptively align with user feedback, the gradient of the loss function with respect to the graph parameters needs to be calculated to determine the optimization direction. Subsequently, the gradients of the knowledge graph node attributes and edge weights are calculated using the backpropagation algorithm. Regarding the aforementioned loss function... Taking the partial derivative, we can obtain the gradient of the node attribute. With edge weight gradient These gradients characterize which node attributes or edge associations should be strengthened or weakened under the current decision error. Next, an adaptive optimization algorithm is used to update the graph structure parameters. The update formula can be written as... ;in: : Represents the set of graph parameters to be updated (including the node attribute matrix) With edge weight matrix ); The learning rate controls the step size for each parameter update. The moving average of the first-order gradient is used to smooth gradient changes. The second-order gradient moving average is used to adjust the adaptive step size for different parameters. This prevents small constants with zero denominators. This formula corresponds to the general form of adaptive optimization algorithms (similar to the Adam optimizer), dynamically adjusting the learning rate of each parameter to make convergence more stable and efficient. Finally, the optimized knowledge graph parameters are transmitted to the context graph construction module in real time. This step implements the system's "closed-loop" mechanism: the optimization results act inversely on the knowledge graph construction module, enabling subsequently generated context graphs to automatically absorb historical optimization experience. For example, if a certain type of geographical scene node is consistently positively correlated with increased user interest, the system will increase the initial association weight of such nodes after the update; conversely, it will decrease their weight or weaken edge connection strength.

[0038] In summary, this application has at least the following effects:

[0039] This user interest matching online marketing system integrates knowledge graphs. By introducing user-nature interest representation and causal inference mechanisms, it effectively identifies and eliminates confounding factors caused by exposure bias, location bias, and selection bias, thereby obtaining a more realistic and stable representation of user interests. Through an end-to-end counterfactual learning and graph neural network optimization framework, the system achieves collaborative training of user interest modeling and knowledge graph structure optimization, enabling the model to adaptively update graph parameters based on real-time feedback. By introducing a gradient propagation-based graph structure optimization algorithm and an adaptive parameter update mechanism, the system possesses continuous self-learning capabilities, maintaining high-precision personalized decision-making performance in dynamically changing user behaviors and contextual environments. Compared with existing static recommendation or non-causal modeling methods, this application significantly improves the robustness of interest modeling and the intelligence level of decision optimization, demonstrating strong scalability and application value.

[0040] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0041] This invention is described with reference to flowchart illustrations and / or block diagrams of systems, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0042] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0043] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0044] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0045] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A user interest matching online marketing system integrating knowledge graphs, characterized in that: It includes the following modules: context graph construction module, interest evolution calculation module, causal debiasing module, and intelligent decision-making module; The context graph construction module is used to receive user behavior event streams and environmental context data in real time through distributed event stream processing. A streaming graph update algorithm based on time windows dynamically maintains the knowledge graph structure. When a user is detected entering a specific geofence, preset product scenario association rules are activated, generating a dynamic knowledge graph incremental update sequence with event timestamps. The specific process is as follows: Real-time monitoring of user location information and identification of whether users have entered preset interest areas through geofence detection algorithms. When a user enters a specific geofence, the scene recognition engine is triggered and the corresponding business scene type is matched. Activate the preset product scenario association rule base, retrieve the relevant product categories and attribute features in the scenario, and dynamically construct the user-scenario-product association subgraph based on the scenario association rules; Generate an incremental update operation sequence with time-series markers for each associated subgraph, and sort the incremental update sequence according to the event timestamps to ensure time-series consistency; Establish a smooth transition mechanism for the knowledge graph when switching scenarios, and avoid abrupt changes in relationships through gradual updates to maintain the stability of the knowledge graph; The interest evolution computation module is used to incrementally update the sequence based on the dynamic knowledge graph. It extracts the multi-hop neighbor structure change features of the user within the sliding time window through a temporal convolutional graph network, quantifies the contribution weight of different contextual nodes to the user's interest state using a multi-head graph attention mechanism, and outputs a time-dependent user interest embedding vector. The specific process is as follows: A multi-head graph attention network is constructed and independent feature transformation parameters are assigned to each attention head. The attention coefficients between user nodes and context nodes are calculated to characterize their influence on the user's interest state. The design employs a hierarchical attention mechanism to calculate attention weights at both the node and graph levels, and introduces a time decay factor to dynamically adjust the contribution weight of historical context nodes to the current interest state. A comprehensive contextual influence weight distribution is obtained by splicing and fusing multi-head attention outputs. Based on the attention weights, the features of neighboring nodes are weighted and aggregated to update the embedding representation of user nodes. By deeply fusing temporal structure features with attention-weighted features, a user interest embedding vector with temporal dependence is generated. The causal debiasing module is used to construct a causal inference model that includes exposure bias and position bias based on the user interest embedding vector. It uses an end-to-end counterfactual learning framework to jointly train the sample distribution of the processing group and the control group, and uses a gradient inversion layer to eliminate the confusion effect of selection bias on the interest vector, and outputs the debiased user essential interest representation. The intelligent decision-making module is used to perform multi-dimensional correlation analysis between the user's essential interest representation and real-time feedback data. It dynamically adjusts the node attributes and edge weights in the knowledge graph of the context graph construction module through a gradient propagation-based graph structure optimization algorithm.

2. The user interest matching online marketing system integrating knowledge graphs according to claim 1, characterized in that: The specific process of dynamically maintaining the knowledge graph structure using a time-window-based streaming graph update algorithm, which receives user behavior event streams and environmental context data in real time through distributed event stream processing, is as follows: A distributed event stream processing pipeline is established to continuously receive multi-source data streams. Event time alignment processing is used to eliminate the impact of out-of-order events caused by network transmission. A sliding time window mechanism based on event time is used to divide the continuous data stream into time segments of fixed duration. User behavior events are analyzed in real time within each time window, entity and relationship information in the events are extracted, and the extracted entities and relationships are incrementally updated to the knowledge graph through a streaming graph update algorithm. Version management of the knowledge graph is maintained and each update operation is timestamped. By adjusting the relationships between nodes in real time based on newly arriving data, the knowledge graph structure can be dynamically evolved, ensuring that the knowledge graph can accurately reflect the user's latest status.

3. The user interest matching online marketing system integrating knowledge graphs according to claim 1, characterized in that: The specific process of extracting the multi-hop neighbor structure change features of a user within a sliding time window using a temporal convolutional graph network based on the incremental update sequence of the dynamic knowledge graph is as follows: A temporal convolutional graph network architecture is constructed and a graph convolutional layer with temporal awareness is designed to perform temporal slicing on the incremental update sequence of the knowledge graph within a sliding time window; Graph convolution operations are applied to each temporal slice to capture local neighbor structural features. Multi-layer graph convolutional networks are used to propagate and aggregate multi-hop neighbor information and extract deep graph structural features. The design of the temporal dependency modeling module captures the structural evolution patterns between adjacent time slices, and uses dilated temporal convolution to expand the receptive field to capture long-distance temporal dependencies. By comprehensively analyzing the changes in the user's geographic structure over continuous time segments, a user behavior representation containing multi-scale temporal features is formed.

4. The user interest matching online marketing system integrating knowledge graphs according to claim 1, characterized in that: The construction logic for building a causal inference model that includes exposure bias and location bias based on user interest embedding vectors is as follows: A causal graph structure is established based on user interest embedding vectors. Exposure mechanisms and location information are incorporated as intervention variables into the causal graph model to define the causal relationship path between user characteristics, product attributes and contextual factors. Design a structured causal model framework, characterize the influence mechanism of different bias factors on user interest through a learnable parameter matrix, and establish a causal transmission chain from exposure events to user feedback; A potential outcome predictor based on a deep neural network is constructed to estimate the user interest state with and without exposure intervention, quantify the confusion effect of exposure bias and position bias on the interest vector, design a dynamic causal graph update mechanism, and adjust the causal graph structure parameters according to real-time feedback data to ensure that the causal inference model can adapt to constantly changing user behavior patterns.

5. The user interest matching online marketing system integrating knowledge graphs according to claim 4, characterized in that: The specific process of jointly training the sample distributions of the processing group and the control group using an end-to-end counterfactual learning framework, eliminating the confusion effect of selection bias on interest vectors, and outputting the debiased user essential interest representation is as follows: An end-to-end counterfactual learning network architecture was constructed, and a dual-branch structure was designed to process the sample data of the treatment group and the control group respectively. The bias-invariant feature representation was learned through a parameter sharing mechanism. A gradient inversion layer is introduced into the feature extraction layer to invert the gradient of bias-related features during backpropagation, forcing the network to learn user interest representations that are independent of bias factors. We design an adversarial regularization loss function to eliminate the influence of selection bias by minimizing the difference in feature distribution between the treatment group and the control group, while maintaining the discriminative ability of user interest representation; A multi-task learning strategy is adopted to jointly optimize the counterfactual prediction task and the bias elimination task. Through iterative training, the essential user interests and bias factors are gradually separated, and the user interest representation vector is output.

6. The user interest matching online marketing system integrating knowledge graphs according to claim 1, characterized in that: The intelligent decision-making module includes the following steps: By deeply linking and mining user's essential interest representation with real-time feedback data, key factors influencing changes in user interest can be identified through correlation analysis and causal inference. Design a graph structure optimization algorithm based on gradient propagation, construct a differentiable graph neural network computation graph, and use user feedback signals as supervision signals to guide the optimization process of the knowledge graph; The gradients of knowledge graph node attributes and edge weights are calculated using the backpropagation algorithm, and the graph structure parameters are updated using an adaptive optimization algorithm. The optimized knowledge graph parameters are then transmitted to the context graph construction module in real time.

Images