Method for determining precise delivery of crowd advertisement based on knowledge graph

Abstract

The application relates to the field of artificial intelligence and discloses a crowd advertisement accurate delivery determination method based on a knowledge graph, aiming to solve the mismatch problem caused by the interest model lag of traditional advertisement delivery. The method comprises the following steps: constructing a dynamic evolution type user interest knowledge graph, fusing multi-source behavior flow and static portrait data, introducing a time decay factor and an event triggering mechanism to realize adaptive updating of node weights; through time sequence attention propagation and a cross-domain associated edge dynamic generation algorithm, short-term interest pulses are accurately captured; a double-channel graph neural network inference engine is constructed to process long-term stable and short-term burst interest paths respectively, and a delivery decision is generated through confidence weighted fusion. The application improves the timeliness and accuracy of advertisement matching, reduces the mismatch rate, and enhances user experience and self-calibration capability.

Description

Knowledge Graph-Based Methods for Determining Precise Audience Targeting in Advertising Technical Field

[0001] This invention belongs to the field of artificial intelligence, specifically relating to a method for accurately determining the target audience for advertising based on knowledge graphs. Background Technology

[0002] With the rapid development of the digital advertising industry, precision targeting based on user profiles and interest modeling has become a core means to improve advertising conversion rates. Knowledge graphs, as an effective tool for structured representation of user behavior and entity relationships, are widely used to build user interest models. They characterize the semantic relationships between users and entities such as products, brands, and topics through nodes and edges, thereby supporting the intelligent matching decisions of advertising systems. However, existing knowledge graph-driven advertising systems mostly rely on static or infrequently updated interest tags. Their modeling process ignores the essential characteristic of the dynamic evolution of user interests over time, leading to sluggish system responses to sudden, short-term interest shifts and making it difficult to achieve truly real-time and precise targeting.

[0003] Among these methods, audience-based advertising typically uses historical behavioral data as input to construct a long-term, stable user-entity relationship network, and generates interest tags for ad matching. While this method can capture stable user preferences relatively well, it cannot effectively model the phenomenon of interest drift—that is, sudden changes in user interests within a specific time period due to external events, social trends, or changes in consumption scenarios. Due to the lack of explicit modeling of behavioral temporal characteristics, the system often assigns equal weight to browsing records from weeks ago and clicks from yesterday, causing ad recommendations to lag behind the user's true intent, resulting in wasted budget and user resentment.

[0004] Current technologies for constructing user interest graphs generally lack a dynamic decay mechanism over time, and also fail to establish a quantitative model for how interest intensity changes with the timeliness of behavior. This leads to the solidification of entity association strength and rigid path recommendations within the graph. Especially in highly dynamic scenarios such as e-commerce promotions, trending events, or seasonal consumption shifts, the window of user interest is extremely short. Traditional methods cannot quickly identify and strengthen emerging interest nodes, resulting in ad matching that deviates significantly from the user's current needs, creating a technical bottleneck of "outdated recommendations and ineffective reach." Therefore, there is an urgent need for a knowledge graph-enhanced ad delivery method that integrates time-series awareness and dynamic weight adjustment mechanisms to achieve real-time capture and accurate response to shifts in user interests. Summary of the Invention

[0005] This invention provides a method for precise audience targeting of advertisements based on knowledge graphs. It constructs a dynamically evolving user interest knowledge graph, integrates multi-source user behavior data and static profile data in real time, and establishes an adaptive update model for graph node weights with time decay factors and event triggering mechanisms. This achieves millisecond-level perception of short-term interest drift and stable representation of long-term interest structures. At the graph structure level, a temporal attention propagation mechanism and a dynamic generation algorithm for cross-domain interest association edges are introduced. This allows the semantic association strength between nodes in the graph to dynamically adjust with the user's behavior sequence, thereby accurately capturing sudden interest pulses while preserving the core framework of user interests. In the advertisement matching stage, a dual-channel graph neural network inference engine is constructed to handle long-term stable interest paths and short-term sudden interest paths respectively. A path confidence weighted fusion strategy is used to generate the final delivery decision, ensuring that the advertisement content not only conforms to the user's historical preference baseline but also responds to their latest behavioral intentions, fundamentally solving the advertisement mismatch problem caused by the lag in interest models.

[0006] This invention provides a method for accurately targeting audiences for advertising based on knowledge graphs, which includes: collecting multi-dimensional behavioral data streams of users within a preset time window, wherein the multi-dimensional behavioral data streams include page browsing trajectory, search keyword sequence, in-application interaction events, social platform like and share records, and geographic location dwell point sequence;

[0007] Real-time semantic parsing is performed on the multi-dimensional behavioral data stream to extract atomic-level interest triggering units. The atomic-level interest triggering units include behavior type identifiers, occurrence timestamps, duration markers, object entity codes, and contextual environment tags.

[0008] The atomic-level interest triggering unit is injected into a pre-constructed static user profile knowledge graph, which includes user basic attribute nodes, historical preference entity nodes, and edges of a preset interest classification system.

[0009] In the static user profile knowledge graph, a time decay coefficient and an event sensitivity threshold are configured for each node. When the number of times the entity node associated with the newly injected atomic interest triggering unit is triggered exceeds the event sensitivity threshold within a unit of time, the weight recalculation process of the node is activated. The weight recalculation process decays the historical weight according to the exponential decay function and adds the current trigger intensity contribution value to generate the updated node weight.

[0010] In the static user profile knowledge graph, cross-entity temporary association edges are dynamically generated. The generation conditions for cross-entity temporary association edges are: two entity nodes are alternately triggered by the same user in three consecutive behavioral events, and the time interval is less than a preset threshold. The initial weight of the cross-entity temporary association edge is determined by the triggering frequency and the time density, and decays at a fixed rate with subsequent no-trigger behaviors.

[0011] A dual-channel graph neural network inference engine is constructed. The first channel processes long-term stable interest paths, and the second channel processes short-term burst interest paths. The long-term stable interest paths consist of nodes with weights higher than a first preset threshold and their stable associated edges. The short-term burst interest paths consist of nodes with weights lower than a second preset threshold but trigger frequencies higher than a third preset threshold and their temporary associated edges.

[0012] Perform a multi-hop neighbor aggregation operation on the long-term stable interest path to generate a first interest representation vector, and perform a single-hop high-weight focusing operation on the short-term burst interest path to generate a second interest representation vector.

[0013] Calculate the first similarity score between the first interest representation vector and the candidate advertisement content feature vector, and calculate the second similarity score between the second interest representation vector and the candidate advertisement content feature vector.

[0014] Based on the preset long-term interest confidence coefficient and short-term interest confidence coefficient, the first similarity score and the second similarity score are weighted and summed to obtain the comprehensive placement matching score; candidate advertisements with a comprehensive placement matching score higher than the fourth preset threshold are determined as the final placement targets, and the graph path structure and weight distribution on which this placement decision is based are recorded for subsequent model self-calibration.

[0015] Compared with the prior art, the advantages and positive effects of the present invention are as follows:

[0016] 0. By constructing a user interest knowledge graph with real-time dynamic update capabilities, the technical defect of traditional static graphs being unable to detect interest drift is solved, enabling the advertising system to capture sudden, short-lived interest pulses in user behavior; by introducing a time decay factor and an event triggering mechanism, adaptive adjustment of graph node weights is achieved, avoiding the rigidity of interest representation caused by the accumulation of historical data.

[0017] 1. By dynamically generating temporary cross-entity association edges, a semantic bridge is established between short-term behavioral events, enabling the system to identify emerging interest combinations that have not yet formed stable preferences;

[0018] 2. By using a dual-channel graph neural network inference engine, long-term stable interests and short-term sudden interests are processed separately. Then, a confidence-weighted fusion strategy is used to generate the final decision, ensuring that the advertising content not only continues the user's historical preference line, but also responds to their latest behavioral intentions, thus reducing the advertising mismatch rate.

[0019] 3. By recording the graph paths and weight distributions relied upon for each delivery decision, a data foundation is provided for the model's continuous self-calibration, enabling the system to have self-evolution capabilities and fundamentally improving the timeliness, accuracy, and user experience satisfaction of advertising delivery. Attached Figure Description

[0020] Figure 1 is a schematic diagram of the overall technical solution architecture of the knowledge graph-based method for accurately determining the target audience for advertising proposed in this invention.

[0021] Figure 2 is a schematic diagram of the core principle framework of the dynamically evolving user interest knowledge graph in this invention;

[0022] Figure 3 is a flowchart of the logical process of atomic-level interest triggering unit injection and node weight adaptive update in this invention.

[0023] Figure 4 is a logical flowchart of the dynamic generation and decay mechanism of cross-entity temporary associated edges in this invention.

[0024] Figure 5 is a schematic diagram of the structure and path processing framework of the dual-channel graph neural network inference engine in this invention;

[0025] Figure 6 is a flowchart of the logical process of the weighted fusion decision of long-term and short-term interest representation vectors in this invention. Detailed Implementation

[0026] This invention provides a method for precise audience targeting of advertisements based on knowledge graphs. Its core lies in constructing a dynamically evolving user interest knowledge graph. By real-time fusion of multi-source behavioral flow data and static profile data, it establishes an adaptive node weight update mechanism and a dynamic generation mechanism for cross-domain interest-related edges. Finally, a dual-channel graph neural network inference engine achieves collaborative decision-making between long-term stable interests and short-term bursty interests, thereby solving the advertising mismatch problem caused by the lag in traditional static interest models. The implementation process of this method strictly follows a pre-set S-step sequence to ensure the logical rigor and engineering feasibility of the technical solution.

[0027] As one embodiment of the present invention, the collection of multi-dimensional behavioral data streams of users within a preset time window includes page browsing trajectory, search keyword sequence, in-app interaction events, social platform likes and shares records, and geographical location dwell time sequence. It should be understood that the dynamic evolution of user interests is essentially driven by their continuous behavioral sequence in digital space. Traditional advertising systems rely on user profile data often derived from periodically collected log files or offline statistical reports, with update frequencies typically measured in days or weeks, failing to capture user interest shifts on a minute or second-level timescale. For example, after browsing digital product reviews in the morning, a user might develop an interest in a travel destination in the afternoon due to trending events on social media. If the system still pushes mobile ads based on the morning's browsing history, it completely misses the afternoon's travel consumption intention window.

[0028] Therefore, in the technical solution of this invention, a distributed data acquisition agent deployed on the client and server sides is used to capture and structurally encapsulate all user interactions within a preset time window at the millisecond level. Page browsing trajectory data includes the sequence of Uniform Resource Locators (URLs) accessed by the user, page dwell time, scroll depth, and coordinates of click hotspots; search keyword sequences record the complete query terms entered by the user in the search engine or application's search box and their correction history; in-application interaction events cover fine-grained operations such as button clicks, form submissions, video playback progress, and adding items to the shopping cart; social media likes and shares records extract the user's explicit preference for specific content entities and their propagation paths; and geographic location dwell point sequences obtain the user's movement trajectory and dwell hotspots in physical space through a Global Positioning System (GPS) module or base station triangulation technology. All data streams carry a unified timestamp and are transmitted to the real-time processing engine via a lightweight message queue to ensure the temporal integrity and low-latency injection of behavioral events.

[0029] As one embodiment of the present invention, the multi-dimensional behavioral data stream is subjected to real-time semantic parsing to extract atomic-level interest triggering units. The atomic-level interest triggering units include behavior type identifiers, occurrence timestamps, duration markers, object entity codes, and contextual environment tags.

[0030] It is understandable that there are differences in information density and semantic granularity in the raw behavioral data stream. For example, the Uniform Resource Locator (URL) string in a page browsing trajectory needs to be mapped to a specific product category or content topic, while the image and text content in social media like records needs to have core entity tags extracted. Directly injecting the raw data into a knowledge graph would lead to semantic ambiguity of nodes and broken association paths. Therefore, in the technical solution of this invention, a multi-layer semantic parsing pipeline is deployed to abstract the behavioral data stream step by step.

[0031] Behavior type identifiers differentiate between five basic behavior patterns—browsing, searching, interaction, social interaction, and location-based interaction—through a predefined enumeration value system. The occurrence timestamp uses the Coordinated Universal Time (UTC) standard timestamp format to ensure global temporal alignment of user behavior across time zones. The duration marker records the effective interaction duration of the user on the current behavior object, distinguishing between shallow browsing and deep engagement. Object entity encoding matches text, images, or location information in the raw data to standardized entity identifiers through a named entity recognition model and industry knowledge base matching. For example, an iPhone 15 review video is parsed into entity code E00123 and associated with a smartphone category node. Contextual environment tags extract auxiliary information such as device type, network environment, and application version number at the time of the behavior, used for subsequent interest intensity calibration. Each atomic-level interest trigger unit is encapsulated in a key-value pair structure, cached in an in-memory database, and then injected in batches into the knowledge graph update queue, ensuring high throughput and low memory usage during the parsing process.

[0032] In one embodiment of the present invention, the atomic-level interest triggering unit is injected into a pre-constructed static user profile knowledge graph, which includes user basic attribute nodes, historical preference entity nodes, and edges of a preset interest classification system. It should be understood that the static user profile knowledge graph, as the initial state of a dynamic evolution process, directly affects the accuracy of subsequent interest drift perception due to its structural integrity and semantic coverage.

[0033] In traditional solutions, static user profile knowledge graphs typically only contain user demographic attributes and historical purchase records, lacking pre-embedded nodes for potential interest dimensions. This leads to frequent topological reconstruction of the graph when injecting new interest entities. Therefore, in the technical solution of this invention, the pre-constructed static user profile knowledge graph adopts a three-layer architecture. The first layer consists of basic user attribute nodes, including demographic characteristics such as age, gender, occupation, and region, with each node configured with an initial weight value of 0.5. The second layer consists of historical preference entity nodes, generated based on explicit user behavior data over the past six months, covering standardized entities such as product categories, content themes, and service types. Node weights are calculated using a logarithmic function normalized based on the frequency of historical behavior. The third layer consists of edges representing a pre-defined interest classification system. Hierarchical and horizontal relationships between entity nodes are constructed according to industry standard classification methods. For example, a smartphone node is connected to a parent node (consumer electronics) via a "belongs" edge, and connected to sibling nodes such as mobile accessories and mobile applications via "association" edges. All edges are configured with a fixed weight value of 0.8 as the baseline association strength for long-term stable interest paths.

[0034] When injecting atomic-level interest-triggered units, the target node is first matched by object entity encoding. If the target node does not exist, a new node is dynamically created and the weight value is initialized to 0.1. Then, the last trigger timestamp and cumulative trigger count counter of the node are updated to provide a data basis for subsequent adaptive weight updates.

[0035] In one embodiment of the present invention, in the static user profile knowledge graph, a time decay coefficient and an event sensitivity threshold are configured for each node. When the number of times the entity node associated with the newly injected atomic-level interest triggering unit is triggered exceeds the event sensitivity threshold within a unit of time, the weight recalculation process of the node is activated. The weight recalculation process decays the historical weights according to the exponential decay function and adds the current trigger intensity contribution value to generate the updated node weight.

[0036] It is understandable that the time-sensitive nature of user interests requires knowledge graph node weights to have both dynamic decay and burst enhancement capabilities. Traditional solutions use linear decay or fixed half-life models, which cannot distinguish between stable preferences and short-term attention.

[0037] Therefore, in the technical solution of this invention, a personalized time decay coefficient and event sensitivity threshold are configured for each node. The time decay coefficient is set differently according to the node type: the decay coefficient for user basic attribute nodes is 0.01, indicating that its weight hardly decays over time; the decay coefficient for historical preference entity nodes is 0.05, reflecting its slow decay characteristics; and the decay coefficient for newly created nodes is 0.2, accelerating the elimination of ineffective interests. The event sensitivity threshold is dynamically adjusted according to the node's historical trigger frequency: the initial threshold is set to 3 triggers per minute. If the number of triggers by a node in the past hour exceeds the threshold limit of 10, the threshold is increased to 5 triggers per minute to avoid misjudgment due to high-frequency noise triggers. After the weight recalculation process is activated, the historical weight residual value is first calculated according to the exponential decay function, expressed by the formula:

[0038]

[0039] in, Historical weight values The time decay coefficient, This represents the time difference between the current moment and the last update moment. Then, the current trigger strength contribution value is calculated, expressed by the formula:

[0040]

[0041] in, The weighting coefficients for behavior types are as follows: browsing behavior is 0.3, search behavior is 0.6, interactive behavior is 0.8, social behavior is 0.9, and location behavior is 0.4. The duration is indicated in seconds; The context calibration factor is set to 1.2 for mobile devices, 1.0 for desktop devices, and 0.8 for weak network environments. The final node weight is updated to the sum of the historical weight residual and the current trigger strength contribution. If the updated weight exceeds the upper limit of 1.0, it is truncated to 1.0 to ensure numerical stability.

[0042] As one embodiment of the present invention, in the static user profile knowledge graph, a cross-entity temporary association edge is dynamically generated. The generation condition of the cross-entity temporary association edge is that two entity nodes are alternately triggered by the same user in three consecutive behavioral events, and the time interval is less than a preset threshold. The initial weight of the cross-entity temporary association edge is determined by the triggering frequency and the time density, and decays at a fixed rate with subsequent no-trigger behavior.

[0043] It is understandable that short-term user interest drift often manifests as the coordinated triggering of multiple related entities. For example, after searching for camping equipment, a user might continuously browse pages related to outdoor cooking utensils and portable tents. Traditional solutions, lacking a dynamic association mechanism between entities, cannot identify such emerging interest combinations. Therefore, in the technical solution of this invention, a sliding window detector is deployed to monitor user behavior event sequences in real time. When two entity nodes are detected to be alternately triggered in three consecutive behavior events, and the time interval between adjacent events is less than a preset threshold of 30 seconds, a temporary associated edge generation process is triggered. The initial weight calculation formula is expressed as:

[0044]

[0045] in, The total time span of the three behavioral events, in seconds; The entity relevance prior coefficient is calculated using a pre-trained semantic similarity model, determining the path distance between two entities in a static classification system. The closer the distance, the higher the coefficient, ranging from 0.5 to 1.0. Generated temporary association edges are configured with an independent decay coefficient of 0.3. If no new alternating triggering events occur within the next 5 minutes, the weight decays linearly at a rate of 0.1 per minute until it falls below 0.1, at which point the edge is automatically deleted. This mechanism enables the knowledge graph to dynamically capture short-term interest clusters emerging from user behavior. For example, camping equipment, outdoor cooking utensils, and portable tents can be linked by temporary association edges to form a temporary interest subgraph, providing a topological foundation for subsequent extraction of sudden interest paths.

[0046] As one embodiment of the present invention, the dual-channel graph neural network inference engine is constructed. The first channel processes long-term stable interest paths, and the second channel processes short-term burst interest paths. The long-term stable interest paths are composed of nodes with weights higher than a first preset threshold and their stable associated edges. The short-term burst interest paths are composed of nodes with weights lower than a second preset threshold but trigger frequencies higher than a third preset threshold and their temporary associated edges.

[0047] It is understandable that long-term stable interests and short-term bursty interests have fundamentally different topological features and semantic densities in the graph, making it difficult for a single graph neural network to simultaneously optimize the representation capabilities of both types of paths. Therefore, in the technical solution of this invention, a dual-channel graph neural network inference engine is constructed. The first channel adopts a three-layer graph convolutional network architecture, with the input being a subgraph of long-term stable interest paths. This subgraph is constructed by selecting nodes with weights higher than a first preset threshold of 0.7 and their stable associated edges. The neighborhood aggregation function of the graph convolutional layer adopts a weighted summation mode, with edge weights as aggregation coefficients, and the output is a dense vector with a dimension of 256. The second channel adopts a single-layer graph attention network architecture, with the input being a subgraph of short-term bursty interest paths. This subgraph is constructed by selecting nodes with weights lower than a second preset threshold of 0.3 but a trigger frequency higher than a third preset threshold of 5 times per minute and their temporary associated edges. The attention coefficient of the graph attention layer is calculated by multiplying the node weights and edge weights, and the output is a sparse vector with a dimension of 128. The two channels share the underlying entity embedding layer to ensure the consistency of the semantic space, but their parameters are trained independently to adapt to the feature distribution of different paths.

[0048] In one embodiment of the present invention, a multi-hop neighbor aggregation operation is performed on the long-term stable interest path to generate a first interest representation vector, and a single-hop high-weight focusing operation is performed on the short-term burst interest path to generate a second interest representation vector. It should be understood that long-term stable interest paths typically exhibit a tree-like or network topology, requiring multi-hop aggregation to capture global semantic associations, while short-term burst interest paths are mostly star-like or chain-like structures, requiring focusing on core nodes to avoid noise propagation.

[0049] Therefore, in the technical solution of this invention, the multi-hop neighbor aggregation operation of the first channel performs a three-layer graph convolution, with the aggregation radius increasing in each layer. The first layer aggregates direct neighbors, the second layer aggregates two-hop neighbors, and the third layer aggregates three-hop neighbors, finally generating the first interest representation vector through global max pooling. The single-hop high-weight focusing operation of the second channel only aggregates direct neighbors, but introduces a weight gating mechanism, expressed by the formula:

[0050]

[0051] in, Let be the attention weight of node i to neighbor j. Let be the weight value of node i. Let be the weight value of edge ij. Let be the set of neighbors of node i. After aggregation, a second interest representation vector is generated through mean pooling. This design ensures that long-term interest representations contain multi-level semantic abstractions, while short-term interest representations focus on highly active core entities, avoiding interference from low-weight noisy nodes.

[0052] In one embodiment of the present invention, a first similarity score is calculated between the first interest representation vector and the candidate advertisement content feature vector, and a second similarity score is calculated between the second interest representation vector and the candidate advertisement content feature vector. It should be understood that the candidate advertisement content needs to be pre-encoded as a feature vector with the same dimension as the interest representation vector. Traditional solutions using keyword matching or collaborative filtering cannot handle the semantic gap problem. Therefore, in the technical solution of the present invention, the candidate advertisement content feature vector is generated through a pre-trained multimodal encoder, which integrates advertisement text, images, and class tag information, outputting a dense vector with a dimension of 256. The first similarity score is calculated using cosine similarity, expressed by the formula:

[0053]

[0054] in, This is the first interest representation vector. This represents the feature vector of the candidate advertisement content. The second similarity score uses weighted cosine similarity, expressed by the formula:

[0055]

[0056] in, This is the second interest representation vector. This represents the highest node weight in the short-term interest path, used to amplify the decision-making influence of sudden interests. The similarity score is calculated and then subjected to min-max normalization to ensure the value is between 0 and 1.

[0057] In one embodiment of the present invention, the first similarity score and the second similarity score are weighted and summed based on preset long-term interest confidence coefficients and short-term interest confidence coefficients to obtain a comprehensive placement matching score. It should be understood that the decision confidence levels for long-term and short-term interests need to be dynamically adjusted based on their stability and suddenness; a fixed weight allocation would lead to a rigid strategy. Therefore, in the technical solution of the present invention, the long-term interest confidence coefficient is calculated based on the average weight of nodes in a long-term stable interest path, expressed by the following formula:

[0058]

[0059] in, This represents the number of nodes in the long-term path. Node weights. The short-term interest confidence coefficient is calculated based on the trigger frequency and time decay of short-term burst interest paths, expressed by the following formula:

[0060]

[0061] in, This is the trigger frequency per unit of time. This is the highest trigger frequency in history. The sudden interest decay coefficient is 0.1. This is the time difference between the last trigger and the current moment. The formula for calculating the overall delivery matching score is expressed as:

[0062]

[0063] This design allows the system to focus on historical preferences when users' long-term interests are stable, and on real-time intentions when short-term interests surge, achieving a dynamic balance.

[0064] In one embodiment of the present invention, candidate ad content with a comprehensive matching score higher than a fourth preset threshold is determined as the final ad delivery target, and the graph path structure and weight distribution on which this delivery decision is based are recorded for subsequent model self-calibration. It should be understood that ad delivery decisions require a strict threshold filtering mechanism to avoid low-quality matching consuming budget, and the decision log needs to provide feedback signals for model iteration.

[0065] Therefore, in the technical solution of this invention, the fourth preset threshold is set to 0.65, and advertising is triggered only when the overall placement matching score exceeds this threshold. After each placement decision, the system records the long-term interest path node list, short-term interest path node list, weight values ​​of each node, weight values ​​of associated edges, and similarity score components corresponding to the selected advertisement, forming a structured decision log. This log is imported into the model training platform in batches daily to optimize hyperparameters such as time decay coefficient, event sensitivity threshold, and confidence coefficient calculation formula, achieving closed-loop self-calibration of the system. For example, if the log shows that the click-through rate of an advertisement on a certain type of short-term interest path is consistently lower than the industry benchmark, the calculation weight of its short-term interest confidence coefficient is automatically reduced; conversely, the weight is increased to ensure that the model continuously adapts to market changes.

[0066] Through the above implementation methods, the dynamic evolutionary user interest knowledge graph constructed by this invention can perceive interest drift in real time, the dual-channel graph neural network inference engine can accurately distinguish between long-term stable preferences and short-term sudden intentions, and the weighted fusion decision-making mechanism can realize the spatiotemporal adaptation of advertising content, ultimately achieving a dual improvement in the accuracy and timeliness of advertising delivery.

[0067] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0068] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A method for accurately determining audience targeting in advertising based on knowledge graphs, characterized in that, include: Collect multi-dimensional behavioral data streams of users within a preset time window; Real-time semantic parsing is performed on the multi-dimensional behavioral data stream to extract atomic-level interest triggering units; the atomic-level interest triggering units are injected into a pre-built static user profile knowledge graph; in the static user profile knowledge graph, a time decay coefficient and an event sensitivity threshold are configured for each node; when the number of times the entity node associated with the newly injected atomic-level interest triggering unit is triggered exceeds the event sensitivity threshold within a unit of time, the weight recalculation process of the node is activated; the weight recalculation process decays the historical weights according to the exponential decay function and adds the current trigger intensity contribution value to generate updated node weights; In the static user profile knowledge graph, cross-entity temporary association edges are dynamically generated. The generation conditions for these cross-entity temporary association edges are: two entity nodes are alternately triggered by the same user in three consecutive behavioral events, and the time interval is less than a preset threshold. The initial weight of the cross-entity temporary association edge is determined by both the triggering frequency and the temporal density, and decays at a fixed rate with subsequent no-trigger behaviors. A dual-channel graph neural network inference engine is constructed. The first channel processes long-term stable interest paths, and the second channel processes short-term burst interest paths. The long-term stable interest paths consist of nodes with weights higher than a first preset threshold and their stable association edges, while the short-term burst interest paths consist of nodes with lower weights. The system consists of nodes that have a trigger frequency higher than a third preset threshold but a second preset threshold, along with their temporary associated edges. A multi-hop neighbor aggregation operation is performed on the long-term stable interest path to generate a first interest representation vector. A single-hop high-weight focusing operation is performed on the short-term bursty interest path to generate a second interest representation vector. A first similarity score is calculated between the first interest representation vector and the candidate ad content feature vector, and a second similarity score is calculated between the second interest representation vector and the candidate ad content feature vector. Based on preset long-term interest confidence coefficients and short-term interest confidence coefficients, a weighted sum of the first similarity score and the second similarity score is obtained to obtain a comprehensive ad placement matching score. Candidate ad content with a comprehensive placement matching score higher than the fourth preset threshold is identified as the final placement target, and the graph path structure and weight distribution on which this placement decision is based are recorded for subsequent model self-calibration.

2. The method for accurately determining audience targeting based on knowledge graphs according to claim 1, characterized in that, Real-time semantic parsing is performed on the multi-dimensional behavioral data stream to extract atomic-level interest triggering units, including: distinguishing five basic behavioral patterns (browsing, searching, interaction, social, and location) through a behavior type enumeration system and generating behavior type identifiers; marking the time of behavior occurrence using the Coordinated Universal Time (UTC) standard timestamp format and generating an occurrence timestamp; recording the effective interaction duration of the user on the current behavior object and generating a duration marker; matching the raw data with an industry knowledge base using a named entity recognition model and generating an object entity code; extracting the device type, network environment, and application version number information at the time of behavior occurrence and generating a context environment label; and encapsulating the behavior type identifier, occurrence timestamp, duration marker, object entity code, and context environment label into an atomic-level interest triggering unit.

3. The method for accurately determining audience targeting based on knowledge graphs according to claim 2, characterized in that, Injecting the atomic-level interest triggering unit into a pre-built static user profile knowledge graph includes: the static user profile knowledge graph adopts a three-layer architecture, the first layer is user basic attribute nodes, the second layer is historical preference entity nodes, and the third layer is preset interest classification system edges; if the object entity code in the atomic-level interest triggering unit has no corresponding node in the graph, a new node is dynamically created and its weight value is initialized to 0.1; the last trigger timestamp and cumulative trigger count counter of the target node are updated; the initial weight value of the user basic attribute node is 0.5, the weight of the historical preference entity node is calculated according to the frequency of historical behavior and normalized by a logarithmic function, and the preset interest classification system edges are configured with a fixed weight value of 0.

8.

4. The method for accurately determining audience targeting based on knowledge graphs according to claim 3, characterized in that, The process of recalculating the weights of this node includes: calculating the historical weight residuals based on the exponential decay function, using the following formula: ,in Historical weight values The time decay coefficient, The time difference between the current moment and the last update moment; calculate the current trigger strength contribution value using the following formula: ,in For behavior type weighting coefficients, For duration markers, The context environment calibration factor is used; the historical weight residual value is added to the current trigger strength contribution value to obtain the updated node weight. If the result exceeds 1.0, it is truncated to 1.

0.

5. The method for accurately determining audience targeting based on knowledge graphs according to claim 4, characterized in that, Dynamically generate temporary cross-entity association edges, including: deploying a sliding window detector to monitor user behavior event sequences; triggering temporary association edge generation when two entity nodes are alternately triggered in three consecutive behavior events, and the time interval between adjacent events is less than 30 seconds; calculating initial weights: ,in The total time span of the three behavioral events. The prior coefficient for entity relevance is set; an independent decay coefficient of 0.3 is configured. If no new alternating triggering event occurs within 5 minutes, the weight will decrease linearly by 0.1 per minute until it falls below 0.1 and is then automatically deleted.

6. The method for accurately determining audience targeting based on knowledge graphs according to claim 5, characterized in that, A dual-channel graph neural network inference engine is constructed, comprising: the first channel adopts a three-layer graph convolutional network architecture, the input is a long-term stable interest path subgraph, the neighborhood aggregation function adopts a weighted summation mode, and the edge weights are used as aggregation coefficients; the second channel adopts a single-layer graph attention network architecture, the input is a short-term bursty interest path subgraph, and the attention coefficients are calculated by multiplying the node weights and edge weights; the two channels share the underlying entity embedding layer, and the parameters are trained independently to adapt to different path feature distributions.

7. The method for accurately determining audience targeting based on knowledge graphs according to claim 6, characterized in that, Perform a multi-hop neighbor aggregation operation on the long-term stable interest path to generate a first interest representation vector, including: performing a three-layer graph convolution, the first layer aggregating direct neighbors, the second layer aggregating two-hop neighbors, and the third layer aggregating three-hop neighbors; and generating a first interest representation vector with a dimension of 256 through global max pooling.

8. The method for accurately determining audience targeting based on knowledge graphs according to claim 7, characterized in that, Perform a single-hop high-weight focusing operation on the short-term bursty interest path to generate a second interest representation vector, including: aggregating only direct neighbors, introducing a weight gating mechanism, and calculating attention weights. ,in Let i be the weight of node i. Let ij be the edge weight. Let i be the set of neighbors of node i; generate a second interest representation vector with dimension 128 through mean pooling.

9. The method for accurately determining audience targeting based on knowledge graphs according to claim 8, characterized in that, The calculation of the first and second similarity scores includes: the candidate ad content feature vectors are generated through a pre-trained multimodal encoder with a dimension of 256; the first similarity score uses cosine similarity. ,in This is the first interest representation vector. The candidate ad content feature vector is used; the second similarity score uses weighted cosine similarity. ,in This is the second interest representation vector. The highest node weight in the shortest path; for and Perform min-max normalization to the 0-1 interval.

10. The method for accurately determining audience targeting based on knowledge graphs according to claim 9, characterized in that, The weighted summation of the first similarity score and the second similarity score includes: calculating the long-term interest confidence coefficient. ,in This represents the number of nodes in the long-term path. Calculate node weights; calculate short-term interest confidence coefficients: ,in This is the trigger frequency per unit of time. This is the highest trigger frequency in history. The sudden interest decay coefficient is 0.

1. The time difference between the last trigger and the current time; calculate the overall delivery matching score: 。

Images