Personalized Ad Creative Generation Methods Based on User Profiles

Abstract

This invention relates to the field of personalized advertising generation technology and discloses a method for generating personalized advertising creatives based on user profiles. The method includes processing user behavior logs to extract intent seeds, and based on these seeds, performing a heuristic multi-hop traversal in a multi-dimensional associated creative material graph to locate the initial generation domain. Subsequently, within the generation domain, dynamic matching is performed between the template syntax and the intent seeds to weave a structured intermediate representation of the creative. After compliance verification, a preliminary advertising creative is generated. Finally, through a multi-round optimization pipeline consisting of a simulated display environment and a lightweight user feedback model, the creative is iteratively fine-tuned across multiple dimensions to output the final creative. This invention achieves full-process automation from deep association retrieval to simulation iterative optimization, improving the personalization level and generation efficiency of advertising creatives.

Description

Technical Field

[0001] This invention relates to the field of personalized advertising generation technology, specifically a method for generating personalized advertising creatives based on user profiles. Background Technology

[0002] Currently, personalized ad creative generation in the digital marketing field mainly relies on tagging user behavior and using keyword matching or collaborative filtering techniques to select or simply replace creative materials from a fixed material library. This approach treats creative materials as isolated units, only allowing for direct comparison of surface features, and fails to uncover deeper semantic and utility relationships between materials. As a result, the generated creative combinations lack internal logic and novelty, making it difficult to adapt to users' complex and dynamic intentions.

[0003] In the evaluation and optimization phase after creative idea generation, existing solutions generally rely on real-world A / B testing after launch to select better ideas based on actual exposure data. This approach has a long optimization cycle, high economic costs, and depends on sufficient real interaction data. At the same time, due to the lack of standardized breakdown of the creative structure, the optimization process is usually a "black box" overall replacement of the complete creative product, which cannot make targeted and refined adjustments to specific elements such as copywriting, visuals, and layout, resulting in limited optimization efficiency and accuracy.

[0004] There is a need for a generation method that can deeply understand and associate the multi-dimensional attributes of creative materials, enabling intelligent and associative retrieval and combination. Simultaneously, there is an urgent need for a mechanism that can simulate, evaluate, and finely iteratively optimize creative structures in a virtual environment, thereby fundamentally improving the personalization and effectiveness of creative generation while reducing costs. Summary of the Invention

[0005] The purpose of this invention is to provide a method for generating personalized advertising creatives based on user profiles, so as to solve the problems mentioned in the background art.

[0006] To achieve the above objectives, the present invention provides a method for generating personalized advertising creatives based on user profiles, the method comprising:

[0007] Perform multi-dimensional cleaning and normalization processing on the user's original behavior logs to obtain structured behavior sequences, and extract the intent seeds representing the current core preferences based on the structured behavior sequences;

[0008] Based on the intent seed, a multi-source heterogeneous advertising creative material library is searched in parallel. The search process constructs a multi-dimensional related creative material map based on the material's meta-information tags, visual feature descriptions, and historical interaction data.

[0009] Based on the preference information carried by the intention seed, a heuristic multi-hop traversal and evaluation is performed in the creative material graph to locate an initial creative generation domain;

[0010] Within the initial creative generation domain, element matching and adaptation calculation are performed based on the predefined creative template syntax and the intent seed to dynamically weave an intermediate creative representation containing text fragments, visual elements, and layout instructions;

[0011] The intermediate representation of the creative is subjected to rule-based compliance checks and context-based coherence filling to generate a preliminary advertising creative.

[0012] The initial ad creative is input into a multi-round optimization pipeline, which iteratively fine-tunes multiple dimensions of the initial ad creative through a simulated display environment and a lightweight user feedback model, and finally outputs an optimized personalized ad creative.

[0013] Preferably, the step of performing multi-dimensional cleaning and normalization processing on the user's original behavior logs to obtain a structured behavior sequence includes:

[0014] Collect user interaction logs from different terminals and platforms, remove invalid records and noisy data, and convert all timestamps to a standard time zone format;

[0015] Identify and categorize behavior types in logs, including page views, content dwell time, click interactions, search queries, and transaction completions, and assign standardized behavior codes to each behavior type;

[0016] For each valid log record, features from three dimensions—behavioral subject, behavioral object, and behavioral environment—are extracted and arranged in ascending order of timestamps to form the structured behavioral sequence.

[0017] The extraction of the intent seed representing the current core preference based on the structured behavior sequence includes:

[0018] A structured behavior sequence is obtained based on the processed original behavior logs, and the structured behavior sequence is input into the dynamic interest network construction module;

[0019] In the dynamic interest network construction module, different user behavior nodes are connected according to behavior type, occurrence frequency and time decay factor to form a dynamic interest network with time sequence and weight attributes, and the intent seed representing the current core preference is extracted from the structure of the dynamic interest network.

[0020] The structured behavior sequences are grouped according to user identifiers, and the structured behavior sequences belonging to the same user are pushed in real time to the computing unit of the dynamic interest network construction module maintained independently for each user.

[0021] Preferably, in the dynamic interest network construction module, different user behavior nodes are connected according to behavior type, frequency of occurrence, and time decay factor to form a dynamic interest network with temporal and weight attributes, and intent seeds representing the current core preferences are extracted from the structure of the dynamic interest network, including:

[0022] Each independent behavior in the structured behavior sequence is instantiated as a behavior node in a dynamic interest network, and the occurrence time, behavior object identifier, and standardized behavior code of each behavior node are recorded.

[0023] Based on the chronological order of the actions, directed edges are established between nodes of consecutive actions, and the initial association strength is calculated for each directed edge according to the transition probability of the action type and the time interval.

[0024] A time decay factor is introduced to periodically decay the association strength between historical behavior nodes and edges, while strengthening newly added behavior nodes and their edges to maintain the timeliness of the dynamic interest network.

[0025] After the dynamic interest network reaches a stable state, the community detection algorithm is run to identify closely connected behavioral node clusters in the network, and the cluster with the highest sum of node weights is identified as the core interest cluster.

[0026] From the core interest groups, select the most recently occurring central behavior node that is associated with multiple behavior types, extract the behavior object identifier, behavior type and associated context features carried by the central behavior node, and encapsulate them as the intent seed.

[0027] Preferably, based on the intent seed, a multi-source heterogeneous advertising creative material library is searched in parallel. The search process constructs a multi-dimensional, interconnected creative material map based on the material's meta-information tags, visual feature descriptions, and historical interaction data, including:

[0028] Connect to the text material library, image material library, and video material library respectively, and start an independent feature extraction pipeline for each material library;

[0029] In the feature extraction pipeline of the text material library, the text is parsed, and keyword entities, sentiment tendencies and sentence structure features are extracted to generate text feature description vectors.

[0030] In the feature extraction pipeline of the image material library, the images are segmented, the main object, background scene and color distribution are identified, and a visual feature description vector is generated.

[0031] In the feature extraction pipeline of the video material library, key frame sequences are extracted and processed in the same way as image materials. At the same time, camera motion features are analyzed to generate video feature description vectors.

[0032] The feature description vectors of all materials are associated with their inherent meta-information tags and historical click-through rate data. With a unified material identifier as the core, nodes are established in the graph database, and the multi-dimensional creative material map is constructed with feature similarity, tag co-occurrence, and historical collaborative display relationships as edges.

[0033] Preferably, the step of performing a heuristic multi-hop traversal and evaluation in the creative material graph based on the preference information carried by the intention seed, thereby locating an initial creative generation domain, includes:

[0034] Map the intent seed to one or more virtual query nodes in the creative material graph;

[0035] Starting from the virtual query node, a breadth exploration is performed on the connected material nodes according to the type and weight of the edges.

[0036] During each hop of exploration, the comprehensive relevance score between the newly arrived material node and the virtual query node is calculated. The comprehensive relevance score is derived from the weighted values ​​of feature similarity, tag matching degree, and popularity.

[0037] Set a relevance score threshold and a maximum number of exploration jumps. Only retain material nodes whose relevance scores exceed the threshold to add them to the candidate set, and start the next hop exploration from these candidate nodes.

[0038] The exploration terminates when the maximum number of jumps is reached or no new node satisfies the threshold. At this point, all candidate material nodes obtained and the edges between them together constitute the initial creative generation domain.

[0039] Preferably, within the initial creative generation domain, element matching and fit calculation are performed based on the predefined creative template syntax and the intent seed to dynamically weave an intermediate creative representation containing text fragments, visual elements, and layout instructions, including:

[0040] A pre-built, expandable creative template library is provided. Each template defines a structural framework for an advertising creative, including text placeholders, visual element slots, and layout constraint rules.

[0041] The material nodes within the initial creative generation domain are categorized by type into text material sets, image material sets, and video material sets.

[0042] For each candidate creative template, the intent seed is matched with the element type required by the template, and the specific material with the highest adaptability is selected from the corresponding material set. The adaptability is determined by semantic relevance, style consistency and size specification conformity.

[0043] The selected textual materials are filled into the text placeholders, the selected visual materials are filled into the visual element slots, and the corresponding layout instructions are generated according to the layout constraint rules, thereby instantiating a complete creative intermediate representation.

[0044] Preferably, the step of performing rule-based compliance checks and context-based coherence filling on the intermediate representation of the creative to generate a preliminary advertising creative includes:

[0045] Load the advertising compliance rule library, which includes a list of prohibited words, image content specifications, and legal and policy provisions.

[0046] Use the list of prohibited words to scan the text materials in the creative intermediate representation and replace or highlight suspected prohibited words.

[0047] Use image content guidelines to validate visual materials in creative intermediate representations, identify and filter content that does not comply with the guidelines;

[0048] After passing the compliance verification, we analyzed the semantic gaps and visual discontinuities between different elements in the creative intermediate representation.

[0049] From a general advertising corpus and visual element library, filler content that can smoothly connect the semantic gaps and visual breaks is retrieved and inserted into the designated position in the middle of the creative, forming the preliminary advertising creative that is both logically and visually coherent.

[0050] Preferably, the initial ad creative is input into a multi-round optimization pipeline, which iteratively fine-tunes multiple dimensions of the initial ad creative through simulated display environments and lightweight user feedback models, including:

[0051] Construct a simulated advertising display environment that can simulate the advertising rendering effects under different terminal device screens and different application scenario contexts;

[0052] In the simulated advertising display environment, the initial advertising concept is rendered into various display formats;

[0053] A lightweight user feedback model is invoked. This lightweight user feedback model is trained based on historical interaction data and is used to predict the types and probabilities of user interaction behaviors that may occur under each display format.

[0054] Based on the prediction results of the lightweight user feedback model, the weak dimensions of the initial advertising creative in terms of attractiveness, information clarity, or call to action are identified.

[0055] For the identified weak dimensions, the corresponding elements or attributes are adjusted from the middle level of the creative to generate the adjusted advertising creative, and then put back into the multi-round optimization pipeline for the next round of evaluation and fine-tuning, until the preset optimization termination conditions are met.

[0056] Preferably, the optimization termination condition includes any one or more of the following combinations: reaching a preset maximum number of optimization iterations, the predicted feedback scores of all key dimensions exceeding a set threshold, and the improvement in feedback scores over multiple consecutive optimization rounds being less than a specific range.

[0057] Preferably, the construction of the lightweight user feedback model includes:

[0058] Collect historical ad display records and corresponding user interaction data, including click events, browsing duration and conversion behavior. Extract ad creative features, user profile features and display environment features from the historical ad display records. The ad creative features include copywriting sentiment and visual element complexity. The user profile features include core interest group identifiers in the dynamic interest network. The display environment features include terminal device type and application scenario context.

[0059] Using a logistic regression algorithm, a prediction model is trained based on the features of the advertising creative, user profile, and display environment. The prediction model outputs the probability distribution of user interaction with the advertising creative. Feature selection and parameter pruning are performed on the prediction model to reduce computational complexity, forming the lightweight user feedback model. The lightweight user feedback model is then integrated into a simulated display environment in a multi-round optimization pipeline.

[0060] Compared with the prior art, the beneficial effects of the present invention are:

[0061] By constructing a multi-dimensional association graph that integrates metadata tags, visual features, and historical interaction data, discrete materials are transformed into a reasonable semantic network. Heuristic multi-hop traversal and evaluation are performed on this graph based on user intent seeds, overcoming the limitations of traditional retrieval methods that rely on direct label matching. This technology can explore and locate sets of materials that are more semantically related to users' potential preferences but are not directly explicit, based on the diverse connections between nodes in the graph. This allows the system to start from an initial preference and discover cross-category and cross-style related creative elements, thus ensuring the novelty and deep relevance of creative combinations during the material retrieval stage, laying the foundation for generating truly personalized creative content.

[0062] By generating structured intermediate representations of creative ideas, the ideas are deconstructed into independently operable and evaluable copywriting, visual, and layout elements. The designed multi-round optimization pipeline incorporates a simulated demonstration environment and a lightweight user feedback model, enabling pre-launch, rapid iterative fine-tuning of creative ideas. This technology allows for parallel or sequential adjustments and effect predictions of various dimensions of the intermediate creative representation within a virtual environment. This simulation-based iterative mechanism frees creative optimization from entirely relying on expensive online A / B testing, enabling extensive, low-cost, and efficient exploration and quality verification before release, thus improving the accuracy of creative generation and the final effect. Attached Figure Description

[0063] Figure 1 This is a schematic diagram illustrating the working principle of the personalized advertising creative generation method based on user profiles described in this invention.

[0064] Figure 2 A flowchart for structured behavior sequence acquisition and intent seed extraction;

[0065] Figure 3 A flowchart for retrieving multi-source heterogeneous material libraries and constructing creative material maps;

[0066] Figure 4 The distribution of quantity decay across multiple stages for various types of creative materials;

[0067] Figure 5 The trend chart shows the iterative optimization of the interaction probability across different terminals for advertising creatives. Detailed Implementation

[0068] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0069] Please see Figure 1This invention provides a method for generating personalized advertising creatives based on user profiles. The method includes: performing multi-dimensional cleaning and normalization processing on the user's original behavior logs to obtain a structured behavior sequence, and extracting intent seeds representing the current core preferences based on the structured behavior sequence; based on the intent seeds, parallel searching of a multi-source heterogeneous advertising creative material library, constructing a multi-dimensional related creative material graph based on the material's meta-information tags, visual feature descriptions, and historical interaction data; and based on the preference information carried by the intent seeds, performing a heuristic multi-hop traversal and evaluation in the creative material graph to locate... An initial creative generation domain is generated. Within this domain, element matching and adaptation calculations are performed based on a predefined creative template syntax and the intent seed to dynamically weave an intermediate creative representation containing text fragments, visual elements, and layout instructions. Rule-based compliance checks and context-based coherence filling are performed on the intermediate creative representation to generate a preliminary advertising creative. The preliminary advertising creative is then input into a multi-round optimization pipeline, which iteratively fine-tunes multiple dimensions of the preliminary advertising creative through a simulated display environment and a lightweight user feedback model, ultimately outputting an optimized personalized advertising creative.

[0070] Example 1: See Figure 2The system collects user interaction logs from different terminals and platforms, removes invalid records and noisy data, and converts all timestamps to a standard time zone format. It identifies and classifies behavior types in the logs, including page browsing, content dwell, click interaction, search queries, and transaction completion, and assigns standardized behavior codes to each behavior type. For each valid log record, it extracts features from three dimensions: the behavior subject, the behavior object, and the behavior environment, and arranges them in ascending order of timestamps to form a structured behavior sequence. The structured behavior sequences are grouped by user identifiers, and structured behavior sequences belonging to the same user are pushed in real-time to the computing unit of a dynamic interest network construction module maintained independently for each user. Based on the processed original behavior logs, a structured behavior sequence is obtained and input into the dynamic interest network construction module. In the dynamic interest network construction module, different user behavior nodes are connected according to behavior type, frequency of occurrence, and time decay factor to form a dynamic interest network with temporal and weight attributes. Intent seeds representing the current core preferences are extracted from the structure of the dynamic interest network. Specifically, each independent behavior in the structured behavior sequence is instantiated as a behavior node in the dynamic interest network, and the occurrence time, behavior object identifier, and standardized behavior code of each behavior node are recorded. Directed edges are established between consecutively occurring behavior nodes according to the chronological order of behavior occurrence, and the initial association strength is calculated for each directed edge based on the transition probability of behavior type and the time interval. A time decay factor is introduced to periodically decay the association strength between historical behavior nodes and edges, while newly added behavior nodes and their edges are strengthened to maintain the timeliness of the dynamic interest network. After the dynamic interest network reaches a stable state, a community detection algorithm is run to identify closely connected behavior node clusters in the network, and the cluster with the highest sum of node weights is determined as the core interest cluster. From the core interest cluster, the most recently occurring central behavior node that is associated with multiple behavior types is selected, and the behavior object identifier, behavior type, and associated contextual features carried by the central behavior node are extracted and encapsulated as the intent seed.

[0071] In practical implementation, multi-source heterogeneous user interaction logs are cleaned, normalized, and structured. Based on this, a dynamic interest network is constructed to identify users' core preferences. The method uses the behavioral data of a user within an advertising push cycle as an example scenario. The original behavioral logs contain interaction records from multiple terminals, including mobile application clients, websites, and mini-programs. The record fields include user identifier, timestamp, behavior type description, behavior object Uniform Resource Locator, and device information.

[0072] In practical implementation, the processing system collects user interaction logs from different terminals and platforms. Invalid records and noisy data in these logs are removed. Invalid records include error logs generated by network request timeouts, and noisy data includes duplicate browsing records generated by automatic page refreshes. All timestamps are uniformly converted to a standard time zone format. In practical implementation, the system identifies and classifies behavior types in the logs, including page browsing, content dwell, click interaction, search query, and transaction completion. Standardized behavior codes are assigned to these five behavior types: P_VIEW, C_STAY, CLICK, SEARCH, and PURCHASE, respectively. For each valid log record, features are extracted from three dimensions: the behavior subject, the behavior object, and the behavior environment. The behavior subject features include the user identifier and login IP region; the behavior object features include the product category number and page title keywords; and the behavior environment features include the terminal device type and network connection method. These features are then arranged in ascending order of timestamps and combined to form the structured behavior sequence.

[0073] In practice, the structured behavior sequences are grouped by user identifiers, and the structured behavior sequences belonging to the same user identifier are pushed in real time to the computing unit of the dynamic interest network construction module, which is maintained independently for each user. Based on the processed original behavior logs, structured behavior sequences are obtained and input into the dynamic interest network construction module. In the dynamic interest network construction module, different user behavior nodes are connected according to behavior type, frequency of occurrence, and time decay factor to form a dynamic interest network with temporal and weight attributes. Intent seeds representing the current core preferences are extracted from the structure of the dynamic interest network.

[0074] In some embodiments, each independent behavior instance in the structured behavior sequence is instantiated as a behavior node in the dynamic interest network, and the occurrence time, behavior object identifier, and standardized behavior code are recorded for each behavior node. In an example scenario, a user's search record for "hiking shoes" is instantiated as a behavior node, with occurrence time T1, behavior object identifier "hiking shoes," and standardized behavior code SEARCH. Directed edges are established between consecutively occurring behavior nodes according to the chronological order of behavior occurrence, and the initial association strength is calculated for each directed edge based on the transition probability of behavior type and time interval. In a specific implementation, a time decay factor is introduced to periodically decay the association strength between historical behavior nodes and edges, while strengthening newly added behavior nodes and their edges to maintain the timeliness of the dynamic interest network. It can be understood that the decay calculation of association strength can be implemented using a formula that quantifies the change in behavior node weight over time. In a specific implementation, the following formula is used to calculate the decay of behavior node weights:

[0075]

[0076] Where: symbol This represents the current weight of the behavior node at the computation time, with the symbol... The initial weights at the time of behavior node generation are indicated by the symbol. Represents the preset time decay factor coefficient, symbol Indicates the current time at which the system performs the calculation, symbol This indicates the original occurrence time of the behavior node. After the dynamic interest network reaches a stable state, a community detection algorithm is run. This algorithm identifies closely connected clusters of behavior nodes in the network and determines the cluster with the highest sum of node weights as the core interest cluster. From the core interest clusters, the most recently occurring central behavior node that is associated with multiple behavior types is selected. The behavior object identifier, behavior type, and associated contextual features carried by the central behavior node are extracted and encapsulated as an intent seed.

[0077] In some embodiments, in the example scenario, after multiple rounds of decay calculation and edge reinforcement, a node cluster centered on "outdoor equipment" is formed in the dynamic interest network. This cluster includes multiple high-weight behavior nodes such as searching for "hiking shoes," browsing the product details page "Brand A Hiking Shoes," and lingering on product videos. In a specific implementation, the community detection algorithm identifies this cluster as the core interest cluster and selects the most recently occurring behavior node "browsing the product details page 'Brand A Hiking Shoes'" that is associated with both search and browsing behaviors as the central behavior node. Optionally, the behavior object carried by the extracted central behavior node is identified as "Brand A Hiking Shoes," the behavior type is P_VIEW, and the associated contextual features include the terminal device being a mobile phone and the network environment being Wi-Fi. This information is encapsulated to generate an intent seed.

[0078] Example 2: See Figure 3 The system connects to text, image, and video resource libraries, respectively, and initiates an independent feature extraction pipeline for each library. In the text resource library's feature extraction pipeline, the text is parsed, extracting keyword entities, sentiment characteristics, and sentence structure features to generate text feature description vectors. In the image resource library's feature extraction pipeline, images are segmented, identifying the main subject, background scene, and color distribution, and generating visual feature description vectors. In the video resource library's feature extraction pipeline, keyframe sequences are extracted and processed in the same way as the image resources, while simultaneously analyzing camera motion features to generate video feature description vectors. All resource feature description vectors are associated with their inherent metadata tags and historical click-through rate data. Using a unified resource identifier as the core, nodes are established in the graph database, and edges are formed using feature similarity, tag co-occurrence, and historical collaborative display relationships to construct the multi-dimensional creative resource graph.

[0079] In practical implementation, independent feature extraction is performed on the text, image, and video material libraries, and the extracted features are associated with metadata, ultimately constructing a graph structure in the graph database where multiple relationships exist between the materials. For example, in a scenario containing intent seeds related to "outdoor hiking," the system needs to retrieve creative materials related to hiking shoes, windbreakers, and mountaineering equipment from the material library. This material library includes a text material library storing advertising copy, an image material library storing product and scene images, and a video material library storing advertising clips.

[0080] In implementation, the processing system connects to text, image, and video resource libraries, respectively, and initiates an independent feature extraction pipeline for each library. In the text resource library's feature extraction pipeline, the system parses the text, for example, a text resource titled "Waterproof Hiking Shoes," extracting key entities such as "hiking shoes," "waterproof," and "GORE-TEX," sentiment characteristics such as "positive," and sentence structure features such as "interrogative sentence," generating a text feature description vector. In the image resource library's feature extraction pipeline, the system segments the image, for example, a picture of "hiking shoes in a mountain scene," identifying the main object such as "hiking shoes" and "rocks," the background scene such as "mountain" and "cloudy day," and the color distribution such as "main color tone is gray and green," generating a visual feature description vector. In the video resource library's feature extraction pipeline, keyframe sequences are extracted and processed in the same way as the image resource, while camera movement features such as "push-in" and "close-up" are analyzed to generate a video feature description vector.

[0081] In some embodiments, the feature description vectors of all creative materials are associated with their inherent meta-information tags and historical click-through rate (CTR) data. Meta-information tags include manually labeled product categories, applicable seasons, and style tags. Historical CTR data refers to the CTR statistics obtained by the creative material in historical ad exposures. A unified creative identifier is used as the core, and nodes are established in a graph database. The node attributes include the creative identifier, feature description vector, meta-information tags, and historical CTR. In a specific implementation, the multi-dimensional creative material graph is constructed using feature similarity, tag co-occurrence, and historical collaborative display relationships as edges. Feature similarity edges connect creative nodes with similar visual or semantic features, tag co-occurrence edges connect creative nodes with the same or highly related tags, and historical collaborative display relationship edges connect creative nodes that have been displayed together in the same advertising campaign. It can be understood that the weight of the feature similarity edge can be determined by calculating the similarity between feature vectors. In a specific implementation, the weight of the feature similarity edge between two creative nodes is calculated using the following formula:

[0082]

[0083] Where: symbol The weight of the feature similarity edge between material node i and material node j is represented by the symbol. Represents the feature description vector of material node i, symbol This represents the feature description vector of material node j. Optionally, for text feature description vectors, the above calculation is performed using the average of word embedding vectors; for visual feature description vectors, the above calculation is performed using feature vectors extracted by a deep convolutional neural network. The weights of historical collaborative display relationship edges can be set according to the frequency of material nodes co-occurring in the same advertising campaign.

[0084] In some embodiments, in the constructed multidimensional related creative material graph, an image material node representing "hiking shoes" is connected to another image material node with similar color distribution, "mountain backpack," via feature similarity edges; it is connected to a text material node, "waterproof jacket," which also carries the "waterproof" tag, via tag co-occurrence edges; and it is connected to a "campsite" video material node, which was previously deployed together, via historical collaborative display relationship edges. This graph structure allows material retrieval to move beyond single-modality or keyword matching, enabling the discovery of potentially related materials across modalities and categories through graph traversal, providing a rich pool of material candidates for subsequent creative combinations.

[0085] Example 3: Map the intent seed to one or more virtual query nodes in the creative material graph; starting from the virtual query node, conduct a breadth exploration to the material nodes connected to it based on the type and weight of the edges; during each hop of exploration, calculate the comprehensive relevance score between the newly arrived material node and the virtual query node, the comprehensive relevance score being derived from the weighted values ​​of feature similarity, tag matching degree, and popularity; set a relevance score threshold and a maximum number of exploration hops, and only retain material nodes with relevance scores exceeding the threshold to add to the candidate set, and start the next hop exploration from these candidate nodes; terminate when the exploration reaches the maximum number of hops or no new node satisfies the threshold, at which point all candidate material nodes obtained and the edges between them together constitute the initial creative generation domain.

[0086] In practice, the intent seed is mapped to the query starting point in the creative material graph. Through multiple rounds of breadth exploration and comprehensive relevance assessment, a subgraph region containing highly relevant material nodes is gradually defined. In the specific implementation, the constructed multi-dimensional related creative material graph is continued, and an intent seed carrying the behavioral object identifier "Brand A Hiking Shoes" is used as an input example. This intent seed is mapped to a virtual query node in the creative material graph.

[0087] In practice, intent seeds are mapped to one or more virtual query nodes in the creative material graph. The mapping process matches the intent seed with the feature description vectors or meta-information tags of existing material nodes in the graph, based on the behavioral object identifier, behavioral type, and contextual features carried by the intent seed. If a highly matching existing node exists, it is directly used as one of the virtual query nodes; otherwise, a new node carrying intent seed information is created as the virtual query node. In some embodiments, if the intent seed "Brand A Hiking Shoes" highly matches an image material node in the graph describing "Brand A Hiking Shoes Product Image" in terms of features, then that image material node is selected as a virtual query node. Starting from the virtual query node, a breadth exploration is performed on the connected material nodes based on the type and weight of the connecting edges. The types of connecting edges include feature similarity edges, tag co-occurrence edges, and historical collaborative display relationship edges; the weight of the edge affects the priority of the exploration.

[0088] During each hop of the exploration process, a comprehensive relevance score is calculated between the newly arrived material node and the virtual query node. The comprehensive relevance score is derived from a weighted average of feature similarity, tag matching, and popularity. Feature similarity measures the cosine similarity between the feature vectors of the new node and the virtual query node; tag matching measures the degree of overlap between the new node's metadata tags and the contextual feature tags in the intent seed; and the popularity weighted average is calculated based on the historical click-through rate data of the new node after normalization. In essence, the comprehensive relevance score is calculated using a weighted formula. In practice, the following formula is used to calculate the comprehensive relevance score between a candidate material node and a virtual query node:

[0089]

[0090] Where: symbol The final composite relevance score is represented by the symbol. Represents the normalized feature similarity value, with the sign... Represents the normalized label matching score, with the symbol... This represents the normalized weighted heat value, with the sign... , , These represent the preset weight coefficients for the corresponding dimensions, and satisfy the following conditions: Set a relevance score threshold and a maximum number of exploration hops. Only add material nodes with relevance scores exceeding the threshold to the candidate set, and start the next hop exploration from these candidate nodes. In the example, the relevance score threshold is set to 0.6, and the maximum number of exploration hops is 3. In the first hop exploration, starting from the virtual query node "Brand A Hiking Shoes Product Image", connect to the "Waterproof Technology Copywriting" node, the "Outdoor Mountain Background Image" node, and the "Brand B Hiking Shoes Product Image" node. The calculated comprehensive relevance scores are 0.85, 0.72, and 0.58, respectively. Therefore, the "Brand B Hiking Shoes Product Image" node is not added to the candidate set because its score is below 0.6.

[0091] In some embodiments, the second hop exploration begins with the candidate nodes "Waterproof Technology Copywriting" and "Outdoor Mountain Background Image" retained from the first hop. The "Waterproof Technology Copywriting" node connects to the "GORE-TEX Technology Explanation Video" node, and the "Outdoor Mountain Background Image" node connects to the "Hiking Pole Accessory Image" node. The comprehensive relevance score of these new nodes is calculated and compared with a threshold for filtering. Optionally, the exploration process continues until the maximum number of hops is reached or no new node satisfying the threshold is found starting from any of the current candidate nodes. At this point, all candidate material nodes obtained during the exploration process and the edges between them collectively constitute the initial creative generation domain. It can be understood that the initial creative generation domain is essentially a connected subgraph of the original creative material map, where the nodes have a strong direct or indirect association with the original intent seed in terms of features, tags, or historical display relationships. Optionally, in the example scenario, the initial creative generation domain formed after three rounds of navigation includes nodes such as "Brand A Hiking Shoes Product Image", "Waterproof Technology Copywriting", "Outdoor Mountain Background Image", and "GORE-TEX Technology Explanation Video" as well as various edges connecting these nodes. This sub-domain defines the range of materials for subsequent creative element matching and weaving.

[0092] Example 4: A pre-built, expandable creative template library is provided. Each template defines a structural framework for an advertising creative, including text placeholders, visual element slots, and layout constraint rules. The material nodes within the initial creative generation domain are categorized into text material sets, image material sets, and video material sets according to their types. For each candidate creative template, the intent seed is matched with the element types required by the template, and the specific material with the highest fit is selected from the corresponding material set. The fit is determined by semantic relevance, style consistency, and size specification compliance. The selected text material is filled into the text placeholders, the selected visual material is filled into the visual element slots, and corresponding layout instructions are generated according to the layout constraint rules, thereby instantiating a complete intermediate representation of the creative. The process involves loading an advertising compliance rule library, which includes a list of prohibited words, image content guidelines, and relevant laws and regulations. The prohibited word list is used to scan the text materials in the creative intermediate representation, replacing or highlighting suspected prohibited words. Image content guidelines are used to verify the visual materials in the creative intermediate representation, identifying and filtering non-compliant content. After passing the compliance verification, the semantic gaps and visual breaks between different elements in the creative intermediate representation are analyzed. Content that smoothly connects these semantic gaps and visual breaks is retrieved from a general advertising corpus and visual element library and inserted into a designated position in the creative intermediate representation, forming a preliminary advertising creative that is both logically and visually coherent.

[0093] In practice, material elements are structurally assembled using predefined templates, and then processed through content review and logic enhancement to form usable creative drafts. The resulting initial creative generation domain contains material nodes such as "Brand A Hiking Shoes Product Image," "Waterproof Technology Copy," "Outdoor Mountain Background Image," and "GORE-TEX Technology Explained Video." The intent seed carries the core preference information of "Brand A Hiking Shoes" and "Outdoor Hiking."

[0094] In implementation, the system pre-configures an expandable creative template library. Each template in the library defines a structural framework for an advertising creative, including text placeholders, visual element slots, and layout constraints. Material nodes within the initial creative generation domain are categorized into text material sets, image material sets, and video material sets. For each candidate creative template, the intent seed is matched with the element types required by the template, and the specific material with the highest fit is selected from the corresponding material set. Fit is determined by semantic relevance, style consistency, and size specification compliance. In implementation, semantic relevance is calculated by comparing the cosine similarity between the material's keywords and the preference information carried by the intent seed. Style consistency is assessed by judging whether the material's visual tone and text style match the contextual features associated with the intent seed. Size specification compliance checks whether the aspect ratio and resolution of the image or video material meet the specifications defined by the template slot. The final overall fit is a weighted calculation of these three dimensions. In implementation, the following formula is used to calculate the overall fit of a specific material for a specific template slot:

[0095]

[0096] Where: symbol The final overall fitness score is represented by the symbol. Represents the normalized semantic relevance score, symbol The normalized style consistency score is represented by the symbol. Boolean value conversion score representing dimensional specification conformity, symbol , , These represent the preset weight coefficients for the corresponding dimensions. Refer to Table 1, which illustrates part of the process for element matching and fit calculation of the two candidate templates:

[0097] Table 1: Creative Template Element Matching and Adaptability Calculation Table

[0098] ;

[0099] In some embodiments, based on the adaptation calculation results shown in Table 1, "Brand A Hiking Shoes Product Image" is selected as the main visual image for Template A, and "Waterproof Technology Copy" is selected as the advertising title copy; for Template B, "GORE-TEX Technology Explained Video" is selected as the main visual video, and "Waterproof Technology Copy" is selected as the advertising title copy. The selected copy materials are filled into the copy placeholders, the selected visual materials are filled into the visual element slots, and the corresponding layout instructions are generated according to the layout constraint rules, thereby instantiating a complete creative intermediate representation.

[0100] In practice, the system loads an advertising compliance rule library, which includes a list of prohibited words, image content guidelines, and relevant laws and regulations. The prohibited word list is used to scan the copywriting materials in the creative intermediate presentation, replacing or highlighting suspected violations. For example, if the word "top-tier" appears in the copywriting material and is on the prohibited word list, it is replaced with "excellent." Image content guidelines are used to verify the visual materials in the creative intermediate presentation, identifying and filtering non-compliant content, such as unauthorized trademarks or inappropriate content. After passing compliance verification, the semantic gaps and visual discontinuities between different elements in the creative intermediate presentation are analyzed. For example, there is a semantic gap between a "Brand A Hiking Shoes Product Image" and a "Waterproof Technology" tagline, where the visual connection is clear but the context is insufficient; there is a visual discontinuity between a "GORE-TEX Technology Explained Video" and the "Outdoor Hiking" theme, where the technical details and usage scenarios are not clearly connected.

[0101] In some embodiments, filler content that smoothly connects semantic gaps and visual breaks is retrieved from a general advertising corpus and visual element library. Optionally, for the semantic gap between the aforementioned "Brand A Hiking Shoes Product Image" and the "Waterproof Technology" copy, a transitional phrase "Fearless of wind and rain, walk freely" is retrieved from the general advertising corpus. For the visual break between the "GORE-TEX Technology Explained Video" and the "Outdoor Hiking" theme, dynamic illustrative material of "raindrops falling on the shoe surface and sliding down" is retrieved from the general visual element library. These retrieved filler contents are inserted into the designated positions in the middle of the creative, forming a preliminary advertising creative that is both logically and visually coherent.

[0102] See Figure 4 The study reveals the quantitative changes in text, images, and videos across the "initial creative generation domain → fit selection → compliance verification" process. Specifically, there are significant differences in the initial generation quantity for different material types: the initial creative generation domain quantity for image materials (approximately 28) is far higher than that for text materials (approximately 12) and video materials (approximately 15), reflecting the abundance of visual materials in the multi-source material library. During the selection phase, all three material types exhibit a step-wise decrease in quantity: in the fit selection stage, text materials decreased from 12 to 8, images from 28 to 19, and videos from 15 to 10; after compliance verification, the quantities further decreased to 7, 18, and 9, respectively. This decreasing trend reflects the dual convergence effect of fit calculation (semantic, style, and specification matching) and compliance rules (prohibited words, content specifications) on the material pool. The final retention quantity of image materials remains the highest, consistent with the core carrier attribute of visual elements in advertising creative.

[0103] Example 5: Construct a simulated advertising display environment capable of simulating advertising rendering effects under different terminal device screens and application scenario contexts; in this simulated environment, render the initial advertising creative into multiple display formats; invoke a lightweight user feedback model, trained based on historical interaction data, to predict the type and probability of user interaction behavior under each display format; based on the prediction results of the lightweight user feedback model, identify the weak dimensions of the initial advertising creative in terms of attractiveness, information clarity, or call to action; for the identified weak dimensions, adjust the corresponding elements or attributes from the creative's intermediate representation level to generate an adjusted advertising creative, and re-enter the multi-round optimization pipeline for the next round of evaluation and fine-tuning until a preset optimization termination condition is met. The optimization termination condition includes any one or more of the following combinations: reaching a preset maximum number of optimization iterations, all predicted feedback scores for key dimensions exceeding a set threshold, and the improvement in feedback scores over multiple consecutive optimization rounds being less than a specific range. The construction of the lightweight user feedback model includes: collecting historical ad display records and corresponding user interaction data, including click events, browsing duration, and conversion behavior; extracting ad creative features, user profile features, and display environment features from the historical ad display records, including ad creative features such as copy sentiment and visual element complexity; user profile features such as core interest group identifiers in a dynamic interest network; and display environment features such as terminal device type and application scenario context; training a prediction model using a logistic regression algorithm based on the ad creative features, user profile features, and display environment features; outputting the probability distribution of user interaction with the ad creative; performing feature selection and parameter pruning on the prediction model to reduce computational complexity; forming the lightweight user feedback model; and integrating the lightweight user feedback model into a multi-round optimization pipeline in a simulated display environment.

[0104] In practice, the implementation involves simulating the display environment, using a lightweight user feedback model, and determining preset optimization termination conditions. The initial advertising creative generated is a single-image, emphasis-based ad design targeting "Brand A Hiking Shoes," including a main visual image and advertising headline copy.

[0105] In the implementation, a simulated advertising display environment is constructed. This environment can simulate the advertising rendering effects under different terminal device screens and application scenario contexts. Terminal device screens include smartphone portrait mode, tablet landscape mode, and PC desktop browser. Application scenario contexts include social media feeds, news app article pages, and pre-roll video ads. In this simulated environment, the initial advertising creative is rendered into various display formats. For example, it is rendered as a card-style ad in a smartphone portrait mode social media feed and as a banner ad in a news article sidebar in a PC desktop browser. A lightweight user feedback model is invoked. This model, trained based on historical interaction data, is used to predict the type and probability of user interaction behavior under each display format. Interaction behavior types include clicking, prolonged browsing, ignoring, and closing.

[0106] In some embodiments, based on the prediction results of a lightweight user feedback model, weaknesses in the initial ad creative regarding attractiveness, message clarity, or call to action are identified. For example, the lightweight user feedback model predicts that the ad creative has a click probability of 0.015 in a smartphone portrait-oriented news feed scenario, but only 0.007 in a PC desktop banner scenario. It also predicts that the probability of prolonged viewing is below 0.01 in both scenarios. This prediction indicates weaknesses in the initial ad creative's attractiveness and overall message clarity in desktop scenarios. It is understood that the construction of a lightweight user feedback model involves its own data and processes. The construction of a lightweight user feedback model includes: collecting historical ad display records and corresponding user interaction data, including click events, browsing duration, and conversion behavior; extracting ad creative features, user profile features, and display environment features from historical ad display records; ad creative features including copy sentiment and visual element complexity; user profile features including core interest group identifiers in a dynamic interest network; and display environment features including terminal device type and application scenario context; using a logistic regression algorithm, a prediction model is trained based on the ad creative features, user profile features, and display environment features. The prediction model outputs the probability distribution of user interaction with the ad creative; feature selection and parameter pruning are performed on the prediction model to reduce computational complexity, forming a lightweight user feedback model, which is then integrated into a simulated display environment in a multi-round optimization pipeline. For identified weak dimensions, corresponding elements or attributes are adjusted at the creative intermediate representation level to generate adjusted ad creatives. For example, to address the issue of weak appeal on desktop, the layout instructions in the creative mid-layout were adjusted, increasing the size of the main visual image by 20%; to address the issue of insufficient information clarity, a supplementary copy fragment about "waterproof testing" retrieved from a general advertising corpus was inserted into the creative mid-layout.

[0107] In practice, the adjusted ad creatives are re-entered into a multi-round optimization pipeline for further evaluation and fine-tuning. The adjusted ad creatives are re-rendered in a simulated ad display environment, and a lightweight user feedback model predicts their interaction probability under the new configuration. The optimization process iterates until preset optimization termination conditions are met. These conditions include one or more combinations of the following: reaching the preset maximum number of optimization iterations, all predicted feedback scores for key dimensions exceeding set thresholds, and feedback score improvements in consecutive optimization rounds falling within a specific range. For example, the maximum number of optimization iterations is set to 5 rounds, the thresholds for key dimensions are 0.02 and 0.012, and the threshold for score improvements in two consecutive rounds is 0.001. The lightweight user feedback model scores the ad creative version generated in each iteration, and the scoring function integrates the predicted probabilities of multiple interactive behaviors. The lightweight user feedback model is trained based on historical ad display records and corresponding user interaction data, and can predict the types and probabilities of various user interactions under specific display formats, including click events, browsing duration, and conversion behavior. The scoring function integrates the predicted probabilities of these interactive behaviors to form a comprehensive evaluation score that reflects the overall performance of the advertising creative in the simulated environment. For example, the probability distribution of the model output is processed to generate a unified metric for quantifying the appeal, clarity of message, or appeal level of an advertising creative version. In practice, the comprehensive predicted feedback score for an advertising creative version is calculated using the following formula:

[0108]

[0109] Where: symbol This represents the probability of a user click predicted by a lightweight user feedback model, denoted by [symbol]. Represents the natural constant, symbol The parameters of the logistic regression model are represented by the symbol. This represents the specific values ​​of the input feature variables, such as ad creative features, user profile features, and display environment features. Optionally, after the third iteration, the predicted click probability of the ad creative in the desktop scenario increases to 0.019, and the long-term browsing probability increases to 0.0115. Furthermore, after the fourth iteration, the improvement in both scores is less than 0.001. At this point, the termination condition of "the improvement in feedback scores across multiple consecutive optimization rounds is less than a specific range" is met. In other words, when any preset optimization termination condition is met, the multi-round optimization pipeline stops iterating, and the ad creative version generated in the current round is used as the final optimized personalized ad creative output.

[0110] See Figure 5During the iterative optimization of advertising creatives across multiple rounds, the dynamic trends of interaction probabilities across different terminal scenarios were revealed. Specifically, the graph uses the iteration round as the horizontal axis and the interaction probability value as the vertical axis, distinguishing the change curves of four types of indicators: desktop click probability, desktop long-term browsing probability, mobile click probability, and mobile long-term browsing probability. The click probability threshold (0.02), long-term browsing threshold (0.012), and the termination iteration round (round 3) are also marked. During the iteration process, the desktop click probability gradually increased from 0.007 in the initial round to nearly 0.019 in round 3; the desktop long-term browsing probability simultaneously increased from 0.005 to approximately 0.0115; the mobile click probability maintained a consistently high growth trend, exceeding the corresponding desktop indicator in each round; and the mobile long-term browsing probability showed a steady upward trend. When the third iteration was reached, the desktop click probability was close to the threshold, the long-term browsing probability was close to the threshold, and the improvement of the indicators in subsequent rounds narrowed. This met the termination condition of "the feedback score improvement of multiple consecutive optimization rounds is less than a certain range". Therefore, the third round was determined to be the optimization termination round.

[0111] 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.

[0112] 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 generating personalized advertising creatives based on user profiles, characterized by the following steps: include: Perform multi-dimensional cleaning and normalization processing on the user's original behavior logs to obtain structured behavior sequences, and extract the intent seeds representing the current core preferences based on the structured behavior sequences; Based on the intent seed, a multi-source heterogeneous advertising creative material library is searched in parallel. The search process constructs a multi-dimensional related creative material map based on the material's meta-information tags, visual feature descriptions, and historical interaction data. Based on the preference information carried by the intention seed, a heuristic multi-hop traversal and evaluation is performed in the creative material graph to locate an initial creative generation domain; Within the initial creative generation domain, element matching and adaptation calculation are performed based on the predefined creative template syntax and the intent seed to dynamically weave an intermediate creative representation containing text fragments, visual elements, and layout instructions; The intermediate representation of the creative is subjected to rule-based compliance checks and context-based coherence filling to generate a preliminary advertising creative. The initial ad creative is input into a multi-round optimization pipeline, which iteratively fine-tunes multiple dimensions of the initial ad creative through a simulated display environment and a lightweight user feedback model, and finally outputs an optimized personalized ad creative. The step of performing a heuristic multi-hop traversal and evaluation in the creative material graph based on the preference information carried by the intention seed, thereby locating an initial creative generation domain, includes: The intent seed is mapped to one or more virtual query nodes in the creative material graph. Starting from the virtual query node, a breadth exploration is performed on the material nodes connected to it, based on the type and weight of the edges. During each hop of exploration, the comprehensive relevance score between the newly arrived material node and the virtual query node is calculated. The comprehensive relevance score is derived by weighting feature similarity, tag matching degree, and popularity. A relevance score threshold and a maximum number of exploration hops are set. Only material nodes with relevance scores exceeding the threshold are added to the candidate set, and the next hop of exploration begins from these candidate nodes. The exploration terminates when the maximum number of hops is reached or no new node satisfies the threshold. At this point, all candidate material nodes and the edges between them together constitute the initial creative generation domain.

2. The method for generating personalized advertising creatives based on user profiles according to claim 1, characterized in that, The process of performing multi-dimensional cleaning and normalization on the user's original behavior logs to obtain structured behavior sequences includes: Collect user interaction logs from different terminals and platforms, remove invalid records and noisy data, and convert all timestamps to a standard time zone format; Identify and categorize behavior types in logs, including page views, content dwell time, click interactions, search queries, and transaction completions, and assign standardized behavior codes to each behavior type; For each valid log record, features in three dimensions—behavior subject, behavior object, and behavior environment—are extracted and arranged in ascending order of timestamps to form the structured behavior sequence. The extraction of the intent seed representing the current core preference based on the structured behavior sequence includes: A structured behavior sequence is obtained based on the processed original behavior logs, and the structured behavior sequence is input into the dynamic interest network construction module; In the dynamic interest network construction module, different user behavior nodes are connected according to behavior type, occurrence frequency and time decay factor to form a dynamic interest network with time sequence and weight attributes, and the intent seed representing the current core preference is extracted from the structure of the dynamic interest network. The structured behavior sequences are grouped according to user identifiers, and the structured behavior sequences belonging to the same user are pushed in real time to the computing unit of the dynamic interest network construction module maintained independently for each user.

3. The method for generating personalized advertising creatives based on user profiles according to claim 2, characterized in that, In the dynamic interest network construction module, different user behavior nodes are connected based on behavior type, frequency of occurrence, and time decay factor to form a dynamic interest network with temporal and weight attributes. Intent seeds representing current core preferences are extracted from the structure of the dynamic interest network, including: Each independent behavior in the structured behavior sequence is instantiated as a behavior node in a dynamic interest network, and the occurrence time, behavior object identifier, and standardized behavior code of each behavior node are recorded. Based on the chronological order of the actions, directed edges are established between nodes of consecutive actions, and the initial association strength is calculated for each directed edge according to the transition probability of the action type and the time interval. A time decay factor is introduced to periodically decay the association strength between historical behavior nodes and edges, while strengthening newly added behavior nodes and their edges to maintain the timeliness of the dynamic interest network. After the dynamic interest network reaches a stable state, the community detection algorithm is run to identify closely connected behavioral node clusters in the network, and the cluster with the highest sum of node weights is identified as the core interest cluster. From the core interest groups, select the most recently occurring central behavior node that is associated with multiple behavior types, extract the behavior object identifier, behavior type and associated context features carried by the central behavior node, and encapsulate them as the intent seed.

4. The method for generating personalized advertising creatives based on user profiles according to claim 3, characterized in that, Based on the intent seed, a multi-source heterogeneous advertising creative material library is retrieved in parallel. The retrieval process constructs a multi-dimensional, interconnected creative material map based on the material's meta-information tags, visual feature descriptions, and historical interaction data, including: Connect to the text material library, image material library, and video material library respectively, and start an independent feature extraction pipeline for each material library; In the feature extraction pipeline of the text material library, the text is parsed, and keyword entities, sentiment tendencies and sentence structure features are extracted to generate text feature description vectors. In the feature extraction pipeline of the image material library, the images are segmented, the main object, background scene and color distribution are identified, and a visual feature description vector is generated. In the feature extraction pipeline of the video material library, key frame sequences are extracted and processed in the same way as image materials. At the same time, camera motion features are analyzed to generate video feature description vectors. The feature description vectors of all materials are associated with their inherent meta-information tags and historical click-through rate data. With a unified material identifier as the core, nodes are established in the graph database, and the multi-dimensional creative material map is constructed with feature similarity, tag co-occurrence, and historical collaborative display relationships as edges.

5. The method for generating personalized advertising creatives based on user profiles according to claim 4, characterized in that, Within the initial creative generation domain, element matching and fit calculation are performed based on the predefined creative template syntax and the intent seed to dynamically weave an intermediate creative representation containing text fragments, visual elements, and layout instructions, including: A pre-built, expandable creative template library is provided. Each template defines a structural framework for an advertising creative, including text placeholders, visual element slots, and layout constraint rules. The material nodes within the initial creative generation domain are categorized by type into text material sets, image material sets, and video material sets. For each candidate creative template, the intent seed is matched with the element type required by the template, and the specific material with the highest adaptability is selected from the corresponding material set. The adaptability is determined by semantic relevance, style consistency and size specification conformity. The selected textual materials are filled into the text placeholders, the selected visual materials are filled into the visual element slots, and the corresponding layout instructions are generated according to the layout constraint rules, thereby instantiating a complete creative intermediate representation.

6. The method for generating personalized advertising creatives based on user profiles according to claim 5, characterized in that, The process of performing rule-based compliance checks and context-based coherence filling on the intermediate representation of the creative to generate a preliminary advertising creative includes: Load the advertising compliance rule library, which includes a list of prohibited words, image content specifications, and legal and policy provisions. Use the list of prohibited words to scan the text materials in the creative intermediate representation and replace or highlight suspected prohibited words. Use image content guidelines to validate visual materials in creative intermediate representations, identify and filter content that does not comply with the guidelines; After passing the compliance verification, we analyzed the semantic gaps and visual discontinuities between different elements in the creative intermediate representation. From a general advertising corpus and visual element library, filler content that can smoothly connect the semantic gaps and visual breaks is retrieved and inserted into the designated position in the middle of the creative, forming the preliminary advertising creative that is both logically and visually coherent.

7. The method for generating personalized advertising creatives based on user profiles according to claim 6, characterized in that, The initial ad creative is input into a multi-round optimization pipeline, which iteratively fine-tunes multiple dimensions of the initial ad creative through simulated display environments and lightweight user feedback models, including: Construct a simulated advertising display environment that can simulate the advertising rendering effects under different terminal device screens and different application scenario contexts; In the simulated advertising display environment, the initial advertising concept is rendered into various display formats; A lightweight user feedback model is invoked. This lightweight user feedback model is trained based on historical interaction data and is used to predict the types and probabilities of user interaction behaviors that may occur under each display format. Based on the prediction results of the lightweight user feedback model, the weak dimensions of the initial advertising creative in terms of attractiveness, information clarity, or call to action are identified. For the identified weak dimensions, the corresponding elements or attributes are adjusted from the middle level of the creative to generate the adjusted advertising creative, and then put back into the multi-round optimization pipeline for the next round of evaluation and fine-tuning, until the preset optimization termination conditions are met.

8. The method for generating personalized advertising creatives based on user profiles according to claim 7, characterized in that, The optimization termination conditions include any one or more of the following combinations: reaching the preset maximum number of optimization iterations, the predicted feedback scores of all key dimensions exceeding the set threshold, and the improvement in feedback scores over multiple consecutive optimization rounds being less than a specific range.

9. The method for generating personalized advertising creatives based on user profiles according to claim 8, characterized in that, The construction of the lightweight user feedback model includes: Collect historical ad display records and corresponding user interaction data, including click events, browsing duration and conversion behavior. Extract ad creative features, user profile features and display environment features from the historical ad display records. The ad creative features include copywriting sentiment and visual element complexity. The user profile features include core interest group identifiers in the dynamic interest network. The display environment features include terminal device type and application scenario context. Using a logistic regression algorithm, a prediction model is trained based on the features of the advertising creative, user profile, and display environment. The prediction model outputs the probability distribution of user interaction with the advertising creative. Feature selection and parameter pruning are performed on the prediction model to reduce computational complexity, forming the lightweight user feedback model. The lightweight user feedback model is then integrated into a simulated display environment in a multi-round optimization pipeline.

Images