Advertisement creation failure reason diagnosis method, system, equipment and medium

Abstract

The invention discloses an advertisement creation failure reason diagnosis method, system and device and a medium. The method specifically comprises the steps of collecting multi-source data in an advertisement creation process; constructing a knowledge graph of the advertisement creation process based on the multi-source data; node embedding and relation reasoning are carried out on the knowledge graph through a graph neural network, and associated features in the advertisement creation process are extracted; based on the association features, performing cross-modal alignment and fusion on the text description, the system log and the user operation video by using a Transform model, and generating a fine-grained reason diagnosis result of the failed task; and constructing a success rate prediction model according to historical data and real-time features based on the associated features and the fine-grained reason diagnosis result, and dynamically adjusting advertisement budget allocation based on a prediction result of the success rate prediction model. According to the method, automatic and intelligent diagnosis of the advertisement creation failure reason is realized, the efficiency and accuracy of advertisement creation success rate analysis and optimization are improved, and advertisement putting decision scientization is facilitated.

Description

Technical Field

[0001] This invention relates to the field of Internet technology, and in particular to a method, system, device, and medium for diagnosing the reasons for failure in ad creation. Background Technology

[0002] In today's booming advertising industry, the success rate of ad creation and the analysis of failure reasons play a crucial role in ad performance and business decisions. However, current success rate and failure reason analysis in the ad creation process mainly rely on manual statistics and analysis methods. This approach has many significant shortcomings in terms of efficiency, accuracy, data dimensions, and the provision of optimization suggestions, and is no longer suitable for the increasingly complex ad creation tasks with ever-growing data volumes.

[0003] First, manually calculating ad creation success rates and analyzing failure reasons is extremely tedious and time-consuming. Ad creation involves multiple stages and a vast amount of data; collecting, organizing, and analyzing data from each stage requires a significant investment of manpower and time. For example, for large-scale advertising campaigns, manually calculating the creation success rates for different ad accounts, ad plans, ad groups, and analyzing the detailed reasons for each failed task often takes days or even weeks. This inefficiency makes it difficult for business departments to obtain key information in a timely manner, hindering their ability to quickly respond to market changes and business needs, thus missing the best advertising opportunities and impacting advertising effectiveness and business revenue.

[0004] Secondly, manual statistical analysis is prone to oversights and errors. Due to the massive volume of data and the involvement of multiple data sources, manual operations may overlook some important data or result in errors during data recording and calculation. For example, when calculating the success rate of ad creation, inconsistent understanding of success and failure criteria or incorrect data recording may lead to inaccurate results. Such inaccurate data can seriously affect the reliability of decision-making, making the formulation of advertising strategies lack a scientific basis, thereby reducing the effectiveness and return on investment of advertising.

[0005] Finally, existing statistical methods struggle to comprehensively cover all levels and dimensions of the ad creation process. Ad creation is a complex process involving multiple entities, including ad accounts, ad campaigns, ad groups, ad creatives, creative elements, review rules, and delivery channels, as well as the various relationships and interactions between them. Manual statistics and analysis often only focus on some key metrics, such as the overall success rate and the main types of failures, but cannot delve into the relationships between various entities or the creation situation under different dimensions. For example, it cannot provide detailed analysis of the success rate of each ad campaign under different delivery channels, or the impact of different creative elements on the approval rate of ad creatives. This single data dimension cannot provide in-depth analytical results and fails to meet the needs of refined ad placement.

[0006] Furthermore, existing systems lack intelligent diagnosis and optimization suggestions for the reasons for failure. Manual analysis of failure causes typically only provides general conclusions, lacking targeted optimization measures. For example, when an ad campaign fails to create, it may only indicate that the review process was unsuccessful, but it cannot analyze the specific reasons for the failure, such as whether the creative materials did not meet the rules, the copy contained inappropriate content, or other reasons. Moreover, it cannot provide specific optimization suggestions based on the reasons for failure, such as how to modify the creative materials or adjust the copy. This leaves ad creators often at a loss when faced with failed tasks, making it difficult to effectively improve the success rate of ad creation. Summary of the Invention

[0007] The purpose of this invention is to provide a method, system, device, and medium for diagnosing the reasons for ad creation failures, thereby achieving automated and intelligent diagnosis of the reasons for ad creation failures, improving the efficiency and accuracy of ad creation success rate analysis and optimization, and assisting in the scientific decision-making of ad placement, so as to solve at least one of the aforementioned problems in the prior art.

[0008] In a first aspect, the present invention provides a method for diagnosing the reasons for failure in ad creation, the method specifically comprising: Collect multi-source data during the ad creation process, including task metadata, system logs, text descriptions, and user operation recordings; Based on multi-source data, a knowledge graph of the advertising creation process is constructed, wherein the nodes of the knowledge graph are used to represent advertising entities and the edges are used to represent the relationships between entities; By using graph neural networks to embed nodes and reason about relationships in the knowledge graph, we can extract the associated features in the ad creation process. Based on correlation features, the Transformer model is used to perform cross-modal alignment and fusion of text descriptions, system logs and user operation recordings to generate fine-grained cause diagnosis results for failed tasks. Based on correlation features and fine-grained cause diagnosis results, a success rate prediction model is constructed according to historical data and real-time features, and the advertising budget allocation is dynamically adjusted based on the prediction results of the success rate prediction model. The results of fine-grained cause diagnosis and success rate prediction models are transformed into reinforcement learning actions. The strategy is iterated based on user feedback, and the analysis results and optimization suggestions are displayed through a visual interface.

[0009] Secondly, the present invention provides a system for diagnosing the reasons for failure in ad creation, the system specifically comprising: The data acquisition module is used to collect multi-source data during the ad creation process. The multi-source data includes task metadata, system logs, text descriptions, and user operation recordings. The knowledge graph module is used to construct a knowledge graph of the advertising creation process based on multi-source data. The nodes of the knowledge graph are used to represent advertising entities, and the edges are used to represent the relationships between entities. The feature extraction module is used to extract relevant features in the ad creation process by embedding nodes and reasoning about relationships in the knowledge graph through graph neural networks. The failure diagnosis module is used to perform cross-modal alignment and fusion of text descriptions, system logs and user operation recordings based on correlation features and using the Transformer model to generate fine-grained cause diagnosis results for failed tasks. The budget allocation module is used to build a success rate prediction model based on correlation features and fine-grained cause diagnosis results, and dynamically adjust the advertising budget allocation based on the prediction results of the success rate prediction model. The reinforcement learning module is used to transform the results of fine-grained cause diagnosis and the prediction results of the success rate prediction model into reinforcement learning actions, iterate the strategy based on user feedback, and display the analysis results and optimization suggestions through a visual interface.

[0010] Thirdly, the present invention provides a computer device, comprising: a memory and a processor, and a computer program stored in the memory, wherein when the computer program is executed on the processor, it implements the method for diagnosing the failure reasons of ad creation as described in any of the above methods.

[0011] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein the computer program, when executed by a processor, implements the method for diagnosing the failure reasons of ad creation as described in any of the above methods.

[0012] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention enables automated and intelligent diagnosis of the reasons for ad creation failure, improves the efficiency and accuracy of ad creation success rate analysis and optimization, and helps to make ad placement decisions more scientific.

[0013] 2. This invention effectively mines the complex semantic and structural association information in the advertising creation process, providing key feature support for fine-grained cause diagnosis and success rate prediction.

[0014] 3. This invention accurately integrates multi-source heterogeneous data to generate fine-grained failure cause diagnosis results, improving the depth and accuracy of the analysis of the causes of failed tasks.

[0015] 4. This invention accurately predicts the success rate of advertising tasks, realizes the dynamic and reasonable allocation of advertising budget, and improves the efficiency of advertising.

[0016] 5. This invention provides personalized optimization suggestions based on real-time data and continuously optimizes strategies through user feedback to continuously improve the effectiveness of ad creation.

[0017] 6. This invention provides a reasonable training framework for reinforcement learning agents, ensuring that the agents output effective actions to optimize the advertising creation process. Attached Figure Description

[0018] To more clearly illustrate the technical solutions in the embodiments of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a flowchart illustrating a method for diagnosing the reasons for failure in ad creation according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an advertisement creation failure diagnosis system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the structure of a computer device provided in an embodiment of the present invention. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0022] It should also be understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.

[0023] As used in this application specification and the appended claims, the term "if" may be interpreted, depending on the context, as "when," "once," "in response to determination," or "in response to detection." Similarly, the phrase "if determined" or "if detected [the described condition or event]" may be interpreted, depending on the context, as meaning "once determined," "in response to determination," "once detected [the described condition or event]," or "in response to detection [the described condition or event]."

[0024] Furthermore, in the description of this application and the appended claims, the terms "first," "second," "third," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.

[0025] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0026] In this application embodiment, the entity executing the process includes a terminal device. This terminal device includes, but is not limited to, devices capable of executing the methods disclosed in this application, such as servers, computers, smartphones, and tablets. Figure 1 A flowchart illustrating a method for diagnosing the reasons for failure in ad creation according to an embodiment of the present invention is shown below in detail: S101, Collect multi-source data during the ad creation process, including task metadata, system logs, text descriptions, and user operation recordings.

[0027] In this embodiment, task metadata records basic information about the ad creation task, serving as a crucial basis for understanding the background and overall situation of ad creation. First, when an ad creation task is initiated, the system automatically generates a unique task identifier. This identifier is used throughout the entire ad creation process to associate various subsequently generated data. Next, the system records the basic attributes of the ad creation task, including but not limited to the characteristics of the target audience for the ad campaign, such as age range, gender ratio, and geographic distribution; the time range for ad placement, specifying the start and end times; the budget amount for the ad campaign, which is the maximum amount of money set by the advertiser for this campaign; and the channel information for ad placement, such as whether it is placed on social media platforms, search engine platforms, or video platforms. This information can be manually entered by the advertiser or ad operations personnel through the task creation interface of the ad creation system. After receiving the input information, the system performs format validation and data integrity checks to ensure that the input information conforms to preset specifications and requirements. This information is then stored in the database as part of the task metadata. Meanwhile, during the ad creation process, if the advertiser or operations personnel modify the basic attributes of the task, the system will record the modification time, the operator who made the modification, and the specific content before and after the modification in real time, and include this modification information in the scope of task metadata, so as to trace and analyze the changes in the ad creation task in the future.

[0028] The system logs meticulously record various events and status information during the operation of the ad creation system, playing a crucial role in analyzing technical issues and anomalies during ad creation. To collect system logs, logging functionality needs to be embedded in each key module of the ad creation system. In the ad creation task submission module, when an advertiser or operations personnel submits an ad creation task, the system records the submission time, the task identifier, the user's identity information, and system resource usage at the time of submission, such as CPU utilization and memory usage. In the ad review module, when the system reviews an ad, it records the review start time, review end time, review result (pass or fail), the specific reason code for failure (if any), and the version information of the review rules called by the system during the review process. In the ad delivery module, when an ad begins delivery, the system records basic data such as the delivery start time, the delivery channel identifier, and the initial impressions and clicks. During ad delivery, the system periodically records key metrics such as real-time impressions, clicks, and conversion rates, as well as any error information that occurs during delivery, such as network connection interruptions and server response timeouts. These log messages are written to log files in real time according to a predetermined format and storage strategy. Log files can be categorized and stored by date and time for later querying and analysis. Meanwhile, to ensure the integrity and security of the log data, the system employs a backup mechanism, periodically backing up log files to independent storage devices or cloud storage to prevent data loss.

[0029] The text description includes various textual information generated during the ad creation process, such as ad copy, ad creative descriptions, and review feedback. This information is crucial for understanding the ad content and review status. For ad copy, in the ad creation system's copy editing interface, when advertisers or operations personnel input ad copy, the system captures the input text in real time and stores it in the database, simultaneously recording the input time, the user's identity information, and the ad task identifier to which the copy belongs. Regarding ad creative descriptions, when advertisers or operations personnel provide a detailed description of the ad's creative, the system provides a dedicated text input area where users can enter creative description text. The system will store the creative description text and its related information in the database using the same data collection method as for ad copy. For review feedback, when an ad fails review, reviewers will enter detailed feedback in the system, explaining the reasons for the failure and suggesting modifications. The system will store this feedback, along with the corresponding ad task identifier, review time, and other information, in the database so that advertisers or operations personnel can view and modify the ad content. In addition, to improve the quality and standardization of text descriptions, the system can provide some prompts and templates in the text input interface to guide users to input accurate and clear text content.

[0030] User operation recordings provide a clear overview of advertisers' or operations personnel's actions during ad creation, aiding in the analysis of user habits and potential problems. To collect user operation recordings, screen recording software or integrated screen recording functionality needs to be installed on the ad creation system's client. When a user logs into the ad creation system and begins ad creation, screen recording automatically starts, recording the user's screen display and actions, including mouse movements, clicks, and keyboard input. During recording, the system generates a unique identifier for each recording session and associates this identifier with the currently active ad creation task identifier, allowing for subsequent mapping of the recording to specific ad tasks. The system also records the start and end times of the recording, as well as the user's login information. To reduce storage space and improve retrieval efficiency, screen recording software can use compression technology to compress the recorded video, reducing file size while maintaining video quality. After recording, the video file is stored on the server according to a predetermined storage path and indexed along with the corresponding ad task identifier, recording time, and other information for easy retrieval and playback. During the collection of user operation recordings, it is necessary to ensure that user privacy is protected. For some operation interfaces involving sensitive user information, methods such as blurring or encrypted storage can be used to prevent the leakage of user information.

[0031] S102, Based on multi-source data, construct a knowledge graph of the advertising creation process, wherein the nodes of the knowledge graph are used to represent advertising entities and the edges are used to represent the relationships between entities.

[0032] In this embodiment, various entities in the ad creation process are identified from preprocessed multi-source data. In the task metadata, entities such as ad accounts, ad plans, and ad groups can be identified. For example, by analyzing information about the ad placement entity in the task metadata, different ad accounts are identified; and ad plans and ad groups are divided according to the task planning and hierarchical structure of ad placement. The system logs can identify the review rule entity; by analyzing records related to ad review in the logs, specific review rule clauses are extracted. The text descriptions can identify ad creative and material element entities; information describing the core ad creative and the materials used (such as images, videos, and copywriting style) is found in the ad copy, creative descriptions, and other text content. After keyframe analysis of user operation recordings, combined with the operation context, the placement channel entity can be identified, for example, by observing the placement platform information selected by the user in the operation interface.

[0033] After identifying various entities, entity relationships are extracted. For the relationships between advertising accounts, advertising campaigns, and advertising groups, the inclusion relationships are determined based on the hierarchical information recorded in the task metadata. For example, an advertising account may contain multiple advertising campaigns, and an advertising campaign may contain multiple advertising groups. In the knowledge graph, this hierarchical inclusion relationship is represented by corresponding edges.

[0034] There is a correlation between advertising campaigns and ad creatives. Information is obtained from task metadata and text descriptions. If an advertising campaign uses a specific ad creative, an edge is established between them in the knowledge graph. There is a composition relationship between ad creatives and creative elements. By analyzing the descriptions of the creatives used in the ad creatives in the text descriptions, as well as the operation records of uploading and using creatives in the user operation recordings, it is determined which creative elements make up a specific ad creative, and this composition relationship is represented by edges in the knowledge graph.

[0035] There is a constraint relationship between advertising campaigns and review rules. By analyzing the advertising review results and rule information recorded in the system logs, we can determine which review rules a particular advertising campaign followed during the review process. If the review fails, we can identify which specific rules were violated. Corresponding constraint edges are then created in the knowledge graph to represent this relationship. There is a delivery relationship between advertising campaigns and delivery channels. Based on user operation recordings of channel selection and delivery channel settings in the task metadata, we determine which delivery channel the advertising campaign will run on, and this delivery relationship is represented by edges in the knowledge graph.

[0036] After completing entity recognition and entity relationship extraction, the knowledge graph is constructed. A suitable knowledge graph storage structure is chosen, such as a graph database. Graph databases can efficiently store and query relationships between nodes and edges, making them suitable for handling complex relational networks like those in an ad creation process.

[0037] The identified advertising entities are created as nodes in a graph database. Each node is assigned a unique identifier, and relevant entity information (such as the advertising account name, advertising campaign budget, and ad creative description) is stored as node attributes. Then, based on the extracted entity relationships, edges are created between the corresponding nodes, and each edge is assigned a relationship type identifier (such as containment, association, or composition). Additional information about the relationships (such as the specific terms of the review rules) can also be stored as edge attributes.

[0038] During the construction process, data consistency and integrity checks must be performed. Ensure that each entity has a clear definition and attribute values, that each edge correctly connects two related entities, and that the relationship types are accurate. For example, check whether the number of ad campaigns under an ad account matches the actual records, and whether the compositional relationships between ad creatives and creative elements are complete.

[0039] After construction is complete, the knowledge graph is stored in a graph database and backed up. A combination of regular full backups and real-time incremental backups can be used to ensure the security and recoverability of the knowledge graph data. Simultaneously, to facilitate subsequent queries and use, indexes can be created for the knowledge graph. Indexes can be designed based on entity attributes and relationship types to improve query efficiency. For example, indexes can be created based on attributes such as ad account name and ad campaign budget range to quickly locate relevant ad entities and relationship information.

[0040] S103 uses graph neural networks to embed nodes and reason about relationships in the knowledge graph, extracting the associated features in the ad creation process.

[0041] In this embodiment, relevant data is extracted from a pre-constructed knowledge graph of the advertising creation process. The attribute data of entities such as advertising accounts, advertising plans, advertising groups, advertising creatives, creative elements, review rules, and delivery channels in the knowledge graph are organized according to a unified format. For example, for the advertising plan entity, its budget, delivery time range, target audience, and other attribute information are organized. Simultaneously, edge information between nodes is extracted, the edge types are identified (e.g., inclusion relationship, association relationship, composition relationship, constraint relationship, and delivery relationship), and the two nodes connected by each edge are recorded.

[0042] Based on the characteristics and requirements of the knowledge graph for the advertising creation process, a suitable graph neural network model is selected. Considering the complex relationships between nodes and edges in the knowledge graph and the objective of extracting associated features, a graph convolutional network (GCN) is chosen as the base model. GCN can effectively process graph-structured data, updating node representations by aggregating information from node neighbors, thereby capturing local and global relationships between nodes.

[0043] Initialize the selected graph neural network model. Determine the number of layers, generally based on the size and complexity of the knowledge graph. For large-scale, complex advertising knowledge graphs, the number of layers can be increased to better capture deeper relational features, but excessive layers should be avoided to prevent overfitting. Set the number of neurons in each layer. The number of neurons affects the model's expressive power and computational complexity, requiring a trade-off adjustment based on the specific situation. Initialize the model's weight parameters. Random initialization can be used, assigning each weight a small random value to ensure good exploratory behavior in the early stages of training.

[0044] The preprocessed knowledge graph data is input into the initialized graph neural network model to begin node embedding. For each node in the knowledge graph, the model first obtains the initial feature representation of that node. These initial features can be vectors encoded from the node's attribute information. For example, attribute information such as advertising plan budget and delivery time can be converted into vector form through one-hot encoding or embedding encoding, serving as the node's initial features. Then, the model aggregates the neighboring node information of each node according to the structure of the knowledge graph. For each node, its directly connected neighboring nodes are considered, and the feature information of the neighboring nodes is integrated through a specific aggregation function. Aggregation functions can be average aggregation, maximum aggregation, or attention aggregation, etc. Taking average aggregation as an example, the feature vectors of neighboring nodes are added together and the average value is taken to obtain the aggregated features of the neighboring nodes. The initial features of the node are combined with the aggregated features of the neighboring nodes, and the feature representation of the node is updated through nonlinear transformation (such as using an activation function). The above aggregation and update process is repeated. As the number of model layers increases, each node can gradually aggregate information from more distant neighboring nodes, thus obtaining a node embedding vector containing richer local and global information.

[0045] After node embedding is completed, relational reasoning is performed to extract the association features between nodes. For each edge in the knowledge graph, the strength and nature of the relationship between the nodes are inferred based on the edge type and the embedding vectors of the two connected nodes. For example, for an edge representing the constraint relationship between an advertising plan and a review rule, the degree of compliance between the advertising plan and the review rule is determined by calculating the similarity or specific relationship score between the embedding vectors of the advertising plan node and the review rule node.

[0046] An attention mechanism is employed to enhance the effectiveness of relational reasoning. This mechanism dynamically assigns different weights based on the importance of the relationships between nodes. During reasoning, for each node, the attention weights between it and all its neighboring nodes are calculated; the magnitude of the weight reflects the importance of that neighboring node to the current node. The embedding vectors of neighboring nodes are then weighted and aggregated according to these attention weights to obtain node representations that more accurately reflect the actual relationships. This approach allows for more accurate capture of complex relationship patterns between nodes and the extraction of representative association features.

[0047] After node embedding and relational reasoning, the relevant features in the ad creation process are extracted from the graph neural network model. These features can be the final embedding vector of each node, or specific feature combinations obtained through further processing. For example, the embedding vectors of key nodes such as ad accounts, ad campaigns, and ad groups can be concatenated or merged to obtain a comprehensive feature vector that reflects the entire hierarchical structure of ad creation.

[0048] S104, based on correlation features, uses the Transformer model to perform cross-modal alignment and fusion of text descriptions, system logs and user operation recordings to generate fine-grained cause diagnosis results for failed tasks.

[0049] In this embodiment, the associated features extracted from the knowledge graph are mapped to feature vectors from text descriptions, system logs, and user operation recordings. Based on the relationships between entities in the ad creation process, the correspondence between the associated features and each modal feature is determined. For example, if an associated feature relates to the review rules of an ad campaign, and the text description contains the ad campaign's copy and review feedback, the system log records operation information related to the ad campaign's review, and the user operation recording shows the creator's review modifications to the ad campaign, then the associated feature is associated with the feature related to the ad campaign's review among these three modalities. In this way, a fused feature matrix containing associated features and features from the three modalities is constructed.

[0050] An encoder-decoder architecture using the Transformer model is employed. In the encoder section, multiple identical encoder layers are used, each containing a multi-head self-attention mechanism and a feedforward neural network. The multi-head self-attention mechanism enables the model to simultaneously focus on information from different positions in the input sequence, capturing the dependencies within the sequence. The feedforward neural network is used to perform a non-linear transformation on the output of the self-attention mechanism, enhancing the model's expressive power. Similarly, in the decoder section, multiple identical decoder layers are used. Each decoder layer, in addition to containing the multi-head self-attention mechanism and the feedforward neural network, also introduces an encoder-decoder attention mechanism to capture the correlation information between the encoder output and the decoder input.

[0051] The fused feature matrix is ​​input into the encoder of the Transformer model. In the encoder's multi-head self-attention mechanism, by adjusting the allocation of attention weights, the model can automatically learn the importance and correlation between features of different modalities. For example, the model assigns higher attention weights to features related to failed advertising campaign reviews in text descriptions and system logs, thus achieving initial alignment of features of different modalities. Simultaneously, the encoder-decoder attention mechanism further enhances this alignment effect, enabling the decoder to more accurately focus on modal features related to failed tasks based on the encoder's output.

[0052] In the decoder section, through multiple rounds of self-attention and encoder-decoder attention mechanisms, feature information from different modalities is fully interacted and fused. The feedforward neural network further performs nonlinear transformations and feature extraction on the fused features, enabling the model to learn higher-level, more abstract cross-modal feature representations. In this way, the Transformer model performs deep cross-modal fusion of associated features with initial features from text descriptions, system logs, and user operation recordings, generating a fused feature vector rich in information.

[0053] Based on the fused feature vectors, a classifier is constructed to generate fine-grained cause diagnosis results for failed tasks. Traditional machine learning classifiers such as Support Vector Machines (SVM) and Random Forests can be chosen, or a fully connected neural network from deep learning can be used. Taking a fully connected neural network as an example, multiple hidden layers are set, each containing a certain number of neurons, and an activation function (such as ReLU) is used to introduce a non-linear factor. The output layer has a corresponding number of neurons based on the predefined number of failure cause categories, and a softmax function is used to convert the output into a probability distribution representing the likelihood of each failure cause category.

[0054] The classifier is trained using a dataset of failed tasks created from labeled advertisements. The fused feature vectors are used as input, and the corresponding failure reason category labels are used as output. The classifier parameters are adjusted using a backpropagation algorithm to minimize the error between the model's predictions and the true labels.

[0055] The trained and optimized classifier can predict the fused feature vectors of new ad creation failure tasks, generating fine-grained cause diagnosis results. For example, the diagnosis results may include various specific failure reason categories such as creative materials not complying with review rules, copy containing prohibited content, and incorrect placement channel settings. Each category corresponds to a probability value, indicating the likelihood that the reason will cause ad creation failure. Based on the probability values, the most probable failure reason can be determined, providing ad creators with accurate failure reason diagnosis information to help them optimize and improve ad creation in a targeted manner.

[0056] S105, based on correlation features and fine-grained cause diagnosis results, constructs a success rate prediction model according to historical data and real-time features, and dynamically adjusts the advertising budget allocation based on the prediction results of the success rate prediction model.

[0057] In this embodiment, a large amount of historical data on ad creation tasks is collected from the database of the ad delivery system. This data covers all aspects of ad creation, including ad account information, ad plan settings, ad group configuration, ad creative content, material element characteristics, review rules, and ad delivery channel selection.

[0058] During ad creation, feature information relevant to the current task is collected in real time. These real-time features include, but are not limited to, market environment characteristics of the current ad campaign (such as competitors' ad campaigns, market trends, etc., which can be obtained through third-party market research data or real-time monitoring tools), user behavior characteristics (such as the target audience's real-time search behavior, browsing behavior, etc., which can be obtained from user behavior analysis systems), and real-time status characteristics of the ad campaign system (such as server load, network latency, etc.). These real-time features are then correlated and integrated with historical data to form a comprehensive dataset containing both historical and real-time information.

[0059] This approach integrates relational features extracted from the knowledge graph of the ad creation process with fine-grained cause-based diagnostic results generated using the Transformer model, along with real-time features and historical data. Relational features reflect the relationships and interactions between entities during ad creation, while fine-grained cause-based diagnostic results clarify the specific reasons for past ad creation failures. By combining these with historical and real-time data, a more comprehensive and in-depth information can be provided to the success rate prediction model, enabling it to more accurately capture various factors influencing ad creation success rates.

[0060] Considering the complexity and data diversity of the ad creation success rate prediction problem, the Random Forest algorithm from ensemble learning models was chosen to construct the success rate prediction model. The Random Forest algorithm has advantages such as the ability to handle high-dimensional data, good robustness to outliers and missing values, and the ability to evaluate feature importance, making it suitable for success rate prediction in this scenario. The fused dataset was divided into training and test sets. The Random Forest model was trained using the training set. During training, the model randomly sampled multiple subsets of samples from the training set using bootstrap sampling, and built a decision tree for each subset. By continuously adjusting the parameters of the decision tree (such as tree depth, minimum number of samples for node splits, etc.), the prediction error of the model on the training set was minimized. The trained Random Forest model was evaluated and validated using the test set.

[0061] Based on the prediction results of the success rate prediction model, dynamic advertising budget allocation rules are formulated. For example, different success rate threshold ranges are set. For advertising creation tasks with a higher predicted success rate, a larger budget is allocated to expect higher advertising effectiveness and returns; for advertising creation tasks with a lower predicted success rate, a smaller budget is allocated to reduce investment risk. At the same time, the overall advertising goals and business needs are considered, such as brand promotion and product sales, and the budget allocation is adjusted appropriately according to different goals.

[0062] During ad campaign execution, the system monitors in real-time the predicted success rate of ad creation tasks and actual ad performance metrics (such as click-through rate, conversion rate, and ROI). Based on the real-time monitoring data, the ad budget is dynamically adjusted according to preset budget allocation rules.

[0063] Establish a feedback mechanism to relay the advertising performance after dynamic budget adjustments to the success rate prediction model and budget allocation strategy formulation module. Based on the feedback, continuously optimize and update the success rate prediction model to improve its prediction accuracy; simultaneously, evaluate and adjust the budget allocation strategy to continuously improve the dynamic budget allocation rules, adapting to the ever-changing market environment and business needs, achieving optimal allocation of the advertising budget, and improving the overall effectiveness of advertising and business returns.

[0064] S106 transforms the fine-grained cause diagnosis results and the prediction results of the success rate prediction model into reinforcement learning actions, iterates the strategy based on user feedback, and displays the analysis results and optimization suggestions through a visual interface.

[0065] In this embodiment, based on the business scenario and actual needs of ad creation, a set of actions that reinforcement learning can take is clearly defined. The action space covers multiple dimensions, such as actions to adjust ad creatives, including changing images, modifying video length, and adjusting copywriting style; actions for ad placement strategies, including changing placement time periods, adjusting placement regions, and adding or removing placement channels; and actions involving ad budget allocation, such as increasing or decreasing the budget for a specific ad campaign. By comprehensively reviewing the operable steps in the ad creation and placement process, a complete and detailed action space is constructed to ensure that all factors that may affect ad performance are covered.

[0066] By deeply analyzing the results of fine-grained cause diagnosis and the prediction results of the success rate prediction model, we can uncover the potential correlations between these results and the various actions in the action space. For example, if the fine-grained cause diagnosis indicates that insufficient image clarity in the advertising materials led to the failure of the review, then in the mapping relationship, this cause is closely linked to the action of replacing the image. If the success rate prediction model predicts a low success rate for an advertising campaign on a specific channel, and the fine-grained cause diagnosis does not point to any obvious problems, then actions such as reducing the budget for that channel or changing the channel can be mapped to the prediction result. In this way, the most appropriate action is found for each possible cause diagnosis and prediction result, forming a systematic and reasonable mapping rule.

[0067] In practical applications, after obtaining the fine-grained cause diagnosis results and the prediction results of the success rate prediction model, the system automatically transforms them into specific reinforcement learning actions based on the established mapping relationship. For example, if the system detects that an advertising creative fails the review due to prohibited words in the copy, and the success rate prediction model predicts that the advertisement will have a low success rate in subsequent deployments, the system will automatically generate an action instruction to modify the copy based on the mapping relationship. This instruction will then be used as an action output of reinforcement learning, providing a clear direction for subsequent advertisement optimization.

[0068] To collect user feedback on the effectiveness of reinforcement learning actions and the overall ad creation analysis results, diverse feedback channels have been established. A dedicated feedback entry point is set up in the visual interface, allowing users to express their opinions on the improvement in ad performance after action execution through text input, rating, and selection. For example, users can rate the increase in click-through rate after modifying ad creatives. Simultaneously, system logs record users' actual operational behaviors and ad performance data during ad delivery, such as whether users adjusted their delivery strategies according to system suggestions and the changes in ad conversion rates after adjustments. This data serves as indirect feedback information. Furthermore, regular user surveys, using questionnaires or interviews, can be conducted to obtain more comprehensive and in-depth user feedback.

[0069] The collected user feedback data is systematically analyzed. For direct user ratings and text feedback, natural language processing techniques are used to perform sentiment analysis and keyword extraction on the text content to understand users' positive and negative evaluations of the action's effectiveness and the key issues they are concerned about. For indirect feedback data recorded in system logs, data analysis methods are used to uncover patterns and trends. For example, analyzing the changes in advertising performance indicators after different actions are performed helps identify which actions significantly improve advertising effectiveness and which actions have little or no effect or even a negative impact. In-depth analysis of the feedback data provides a strong basis for strategy iteration.

[0070] Based on the feedback data analysis, the reinforcement learning strategy is iteratively optimized. If a certain action yields poor user feedback after multiple executions and the advertising performance metrics do not improve, the probability of selecting that action is reduced during strategy iterations, or it may even be removed from the action space or optimized further. Conversely, for actions that generate positive user feedback and significantly improve advertising performance, their selection weight in the strategy is increased. Simultaneously, the action space is continuously updated and expanded, and mapping relationships are optimized, taking into account new market environments and business needs, enabling the reinforcement learning strategy to better adapt to the dynamic changes in ad creation and delivery, and continuously improve advertising effectiveness and business revenue.

[0071] The visualization interface presents fine-grained cause diagnosis results and success rate prediction model results in an intuitive and easy-to-understand manner. For fine-grained cause diagnosis results, a tree diagram or fishbone diagram is used to clearly present the various levels of reasons for ad creation failure, such as breaking down the overall ad failure into specific reasons like creatives, copy, and placement settings, allowing users to quickly pinpoint the root cause. For success rate prediction results, line charts or bar charts are used to display the predicted success rates for different ad campaigns, ad groups, or ad creatives, along with comparisons to historical success rates. The confidence intervals for the predicted success rates are also clearly marked, giving users a clear understanding of the reliability of the prediction results. Based on the conversion results of reinforcement learning actions, specific and actionable optimization suggestions are provided to users.

[0072] In some embodiments, in step S102 above, the construction of a knowledge graph of the advertising creation process based on multi-source data, wherein the nodes of the knowledge graph represent advertising entities and the edges represent relationships between entities, specifically includes: Based on multi-source data, natural language processing technology is used to identify advertising entities and their attributes. The advertising entities include advertising accounts, advertising plans, advertising groups, advertising creatives, creative elements, review rules, and delivery channels. Identify the relationship types between advertising entities, including membership, constraint, triggering, and dependency relationships; Using advertising entities as nodes and relationship types as edges, a knowledge graph of the advertising creation process is constructed, and attribute features are assigned to nodes and edges, including creation success rate, review time, material size and rule strength. Based on user interface operation recordings, multimodal coding technology is used to embed the interface operation sequence and interface state change information as vectors into a knowledge graph and associate them with corresponding nodes.

[0073] In this embodiment, textual data from the collected multi-source data is preprocessed. These data sources include task metadata, text descriptions, etc. Named entity recognition technology from natural language processing, combined with a predefined advertising entity dictionary, is used to identify advertising entities from the preprocessed text. The advertising entity dictionary contains common expressions for various entities such as advertising accounts, advertising plans, ad groups, ad creatives, creative elements, review rules, and delivery channels. For each identified advertising entity, its attribute information is further extracted. For advertising accounts, attributes such as account type (e.g., personal account, business account), registration time, and account balance are extracted. For advertising campaigns, attributes such as campaign name, campaign objective (e.g., increasing brand awareness, increasing product sales), budget amount, and campaign duration are extracted. For ad groups, attributes such as the number of ads within the group and targeting settings (e.g., geographic targeting, audience targeting) are extracted. For ad creatives, attributes such as creative type (e.g., image creatives, video creatives) and creative content description are extracted. For creative elements, attributes such as creative type (e.g., image, video, text), creative size, and creative format are extracted. For review rules, attributes such as rule name, rule content, and scope of application are extracted. For delivery channels, attributes such as channel name, channel type (e.g., social media channel, search engine channel), and channel traffic characteristics are extracted. In this way, a detailed attribute information database is built for each advertising entity.

[0074] Based on the business logic of the ad creation process, clarify the types of relationships between ad entities, mainly including hierarchical relationships, constraint relationships, triggering relationships, and dependency relationships. A hierarchical relationship indicates that one entity belongs to another entity; for example, an ad group belongs to an ad campaign, and an ad campaign belongs to an ad account. A constraint relationship indicates that one entity imposes restrictions on the behavior or attributes of another entity; for example, review rules constrain the review standards for ad creatives. A triggering relationship indicates that a change in the state or behavior of one entity will trigger a corresponding action in another entity; for example, reaching the budget limit of an ad campaign will trigger a stoppage of ad delivery. A dependency relationship reflects that the existence or operation of one entity depends on another entity; for example, the display of ad creatives depends on the support of the delivery channel.

[0075] A combined rule-based and machine learning approach is employed to identify relationships between advertising entities. The rule-based approach searches text data for sentences that match specific relationship patterns based on predefined grammatical and business rules, thus determining the relationships between entities. The machine learning approach trains a classification model using a corpus with pre-labeled relationship types. Entity pairs from the text are input into the model, and the model outputs the relationship types between them. By combining these two methods, the accuracy and comprehensiveness of relationship identification are improved.

[0076] The identified advertising entity relationships are verified and corrected. On one hand, manual sampling is used to check whether the relationship identification results match the actual business situation, and incorrectly identified relationships are corrected. On the other hand, automatic verification is performed using the logical consistency in the ad creation process. For example, if an ad group is identified as belonging to multiple ad campaigns, which is inconsistent with business logic, the system automatically marks it and prompts for correction. Through continuous verification and correction, the accuracy of advertising entity relationship identification is ensured.

[0077] Using identified advertising entities as nodes and defined relationship types as edges, a knowledge graph of the advertising creation process is constructed. During construction, a graph database is used to store and manage the knowledge graph, ensuring efficient data storage and rapid retrieval. For example, entities such as advertising accounts, advertising plans, and advertising groups are used as nodes, connected by edges according to their hierarchical relationships, forming a hierarchical structure that intuitively displays the hierarchical relationships between entities in the advertising creation process. Attribute features are assigned to the nodes and edges in the knowledge graph to enrich its information content. For nodes, based on the previously extracted advertising entity attributes, corresponding attribute information is added to each node, such as account type and registration time for advertising account nodes; and placement goals and budget amounts for advertising plan nodes. Simultaneously, considering the characteristics of the advertising creation process, some key business-related attributes are added, such as creation success rate, used to record the success status of each entity's creation; and review time, reflecting the time spent by the advertising entity during the review process. For edges, attributes such as relationship strength and rule strength are assigned. Relationship strength indicates the closeness of the relationship between two entities, and rule strength describes the strictness of the rules in the constraint relationship. By assigning rich attribute features to nodes and edges, knowledge graphs can more comprehensively and accurately describe the ad creation process.

[0078] The captured user interface operation videos are analyzed to extract the interface operation sequence and interface state change information. Using image recognition technology and computer vision algorithms, various user operations on the interface, such as clicking, dragging, and input, are identified, and the time sequence and object of the operations are recorded. Simultaneously, interface state changes are monitored, such as changes in the displayed content of various fields on the ad creation page and updates to the ad list, and these interface state changes are correlated with the corresponding operation time points.

[0079] Multimodal coding technology is used to convert interface operation sequences and interface state change information into vector representations. For interface operation sequences, encoding is performed according to the type and order of the operations; for example, click operations are encoded as specific vector elements, drag operations as different vector elements, and these are combined into an operation sequence vector according to the operation order. For interface state change information, key state features are extracted, such as changes in the attribute values ​​of displayed advertising entities and changes in interface layout, and these features are encoded into state change vectors. Through multimodal coding, information from different modalities is uniformly converted into vector form, facilitating subsequent processing and analysis.

[0080] The encoded vectors are embedded into the knowledge graph and associated with corresponding nodes. Based on the objects and advertising entities involved in the operation recording, the corresponding nodes in the knowledge graph are determined. For example, if a user modifies an advertising creative in the operation recording, the operation sequence vector and state change vector are embedded into the corresponding advertising creative node in the knowledge graph. In this way, user interface operation information is closely linked to advertising entities in the knowledge graph, enabling the knowledge graph to reflect the user's actual operational behavior and interface interaction during the advertising creation process, providing more comprehensive data support for subsequent advertising creation analysis and optimization.

[0081] In some embodiments, step S103 above, which involves using a graph neural network to perform node embedding and relational reasoning on the knowledge graph to extract the associated features during the advertisement creation process, specifically includes: Graph structure data is obtained from the knowledge graph, which includes a set of nodes, a set of edges, node attributes, and edge attributes, and the nodes and edges are initialized with features. Based on graph structure data, node embedding representation learning is performed through a graph neural network model, neighbor node information is aggregated using a graph attention network, and different weights are assigned to different neighbors through an attention mechanism. Through multi-layer message passing and node state updates, high-dimensional embedding vectors of all nodes are output. These high-dimensional embedding vectors are used to represent the semantic and structural information of the advertising entity and its context. Based on high-dimensional embedding vectors, global average pooling or attention-weighted hierarchical readout operations are used to extract full-graph representations, which are used to describe the overall state and key features of the ad creation process. Based on full-image features, we extract the associated features in the ad creation process.

[0082] In this embodiment, graph structure data is obtained from a pre-constructed knowledge graph of the ad creation process. The knowledge graph stores various information related to the ad creation process in the form of a graph. The node set includes ad entities such as ad accounts, ad plans, ad groups, ad creatives, creative elements, review rules, and delivery channels; the edge set represents the membership, constraint, triggering, and dependency relationships between these entities. Each node and edge also has corresponding attributes. For example, node attributes include creation success rate, review time, creative size, and rule strength, while edge attributes include relationship strength and rule strength. These node sets, edge sets, node attributes, and edge attributes are extracted completely as graph structure data for subsequent processing.

[0083] The acquired nodes and edges are initialized with features. For nodes, each attribute value is standardized according to its attribute information to fall within a suitable numerical range. The standardized attribute values ​​are then combined into a vector, which serves as the initial feature vector for that node. Edges undergo similar processing based on their attribute information, converting the edge's attribute values ​​into vector form as the edge's initial features. In this way, both nodes and edges in the knowledge graph are transformed into vector forms that can be processed by a computer, laying the foundation for subsequent graph neural network processing.

[0084] Graph Attention Network (GAT) was chosen as the graph neural network model for learning node embedding representations. GAT can automatically learn attention weights between nodes, thus better aggregating information from neighboring nodes, making it suitable for processing graph-structured data with complex relationships, such as knowledge graphs. In GAT, for each node, its set of neighboring nodes is first determined. Then, the importance between the node and each neighboring node is measured by calculating the attention coefficient. The calculation of the attention coefficient comprehensively considers the node's initial feature vector and the feature vectors of the edges between them. For example, if an ad group node has a strong membership relationship with an ad campaign node (reflected by edge attributes), and the ad group has a significant impact on the execution of the ad campaign in business logic, then this neighboring node (ad campaign node) will be given a higher weight when calculating the attention coefficient. In this way, differentiated weights are assigned to different neighboring nodes, and then the feature vectors of the neighboring nodes are weighted and aggregated according to these weights to obtain the aggregated feature vector of the node. The initial feature vector of each node is fused with the aggregated feature vectors of the neighboring nodes to update the node's feature representation. The fusion method can employ simple concatenation operations or more complex weighted summation methods, depending on the specific circumstances. By continuously iterating and updating the feature representations of nodes, each node can gradually learn information about itself and its surrounding neighbors, thereby better representing the semantic and structural information of the advertising entity in the knowledge graph.

[0085] A multi-layered graph neural network structure is employed for message passing. In each layer, nodes receive information from their neighbors and aggregate and update features based on the attention mechanism described above. As the number of layers increases, nodes can receive information from more distant neighbors, thus expanding their receptive field and better capturing global structural information within the knowledge graph. For example, in the first layer, an ad creative node might only receive information from its directly connected creative element nodes; in the second layer, it can receive information from other nodes related to the creative element, such as the source node. During message passing at each layer, the node's state is updated based on the received messages. This update considers not only information from neighbors but also the node's own historical state information to maintain information continuity and stability. Through this multi-layered message passing and node state update process, the feature representation of each node is continuously enriched and optimized. After multi-layered processing, high-dimensional embedding vectors for all nodes are output. These high-dimensional embedding vectors contain rich semantic and structural information about the ad entity and its context.

[0086] One approach to extracting the full graph representation is to use global average pooling. This involves averaging the high-dimensional embedding vectors of all nodes to obtain an average vector that represents the entire graph. This method is simple and direct, comprehensively considering the information from all nodes in the graph and reflecting the overall average state of the ad creation process knowledge graph.

[0087] Another more flexible approach is to employ a hierarchical readout operation based on attention weighting. First, an attention weight is assigned to each node based on its importance and relevance within the knowledge graph. The determination of importance and relevance can combine business knowledge and node attribute features; for example, an ad campaign node may be more important than a creative element node in the ad creation process, thus warranting a higher attention weight. Then, the high-dimensional embedding vectors of all nodes are weighted and summed according to these attention weights to obtain a representation of the entire graph. Furthermore, a hierarchical approach can be used, first aggregating and weighting nodes in local regions, and then further aggregating the results from different local regions to better capture the hierarchical structure information in the knowledge graph. The full-graph features extracted in this way can more accurately describe the overall state and key features of the ad creation process, highlighting nodes and relationships that significantly impact the ad creation process.

[0088] Based on the extracted full-graph features, and combined with the business logic and actual needs of the ad creation process, the relationships between various entities in the ad creation process are analyzed. For example, by analyzing the weight distribution and feature similarity among different ad entity nodes in the full-graph features, it is possible to discover which ad campaigns have strong correlations with specific delivery channels, and which ad creatives have close connections with review rules, etc.

[0089] Based on the results of the correlation feature analysis, key correlation features in the ad creation process are extracted. These correlation features can be represented in the form of feature vectors or described in the form of rules. These extracted correlation features can provide valuable references for optimizing the ad creation process and formulating ad placement strategies, helping advertising practitioners better understand the inherent rules and key factors in the ad creation process.

[0090] In some embodiments, in step S104 above, the step of using the Transformer model to perform cross-modal alignment and fusion of text descriptions, system logs, and user operation recordings based on correlation features to generate fine-grained cause diagnosis results for failed tasks specifically includes: Based on the association features, the Transformer model is used to map text features, log features and visual features to the same semantic space, and the multi-head self-attention mechanism is used to calculate the association weights between modalities. During feature fusion, intermodal dependencies are captured through a cross-modal attention layer, and a unified multimodal representation vector is generated with the help of the encoder of the Transformer model. Based on multimodal representation vectors, a classifier is used to diagnose failure causes and output the probability distribution of different failure cause categories. By using interpretable attention weights to locate key modal fragments and content that lead to failure categories, and combining them with probability distributions, fine-grained diagnostic results are generated.

[0091] In this embodiment, text features, log features, and visual features are extracted from different data sources involved in the ad creation process. For text descriptions, such as task instructions for ad creation and user feedback, a pre-trained language model (such as BERT) is used for feature extraction, converting the text into fixed-dimensional vector representations that capture the semantic information within the text. System logs contain various operation records and state changes during the ad creation process. By performing word segmentation and encoding on the logs, they are converted into feature vectors suitable for model processing. User operation recordings belong to visual information. Convolutional neural networks (CNNs) are used to extract features from each frame of the recording, obtaining visual feature vectors for the images. These vectors are then combined with time-series information to integrate the visual features of the entire recording.

[0092] Based on the previously extracted association features, these features reflect the inherent connections between different entities and stages in the ad creation process, providing a unified reference framework for mapping features of different modalities. By inputting textual, log, and visual features into the Transformer model, the linear transformation layer in the Transformer model performs linear transformations on these features of different modalities, mapping them to a semantic space of the same dimension. In this way, features of different modalities become comparable and fusionable, enabling analysis and processing at the same semantic level.

[0093] The multi-head self-attention mechanism of the Transformer model is used to calculate the correlation weights between modalities. This mechanism focuses on the relationships between different modal features from multiple subspace perspectives. For each head, it calculates attention scores between text features, log features, and visual features. These scores reflect the degree of correlation between different modal features. For example, when analyzing the reasons for ad creation failures, if the text description mentions a specific operational error, the system log also records the error, and the user's operation recording shows the specific scene of the erroneous operation, then the multi-head self-attention mechanism will assign higher attention scores to the corresponding parts of these three modal features, indicating a strong correlation between them. By combining the attention scores from multiple heads, the final inter-modal correlation weights are obtained, which will be used in the subsequent feature fusion process.

[0094] A cross-modal attention layer is introduced into the Transformer model. This layer is specifically designed to capture the dependencies between features from different modalities. The cross-modal attention layer receives mapped text features, log features, and visual features as input. By calculating the attention weights between different modal features, it determines the degree of influence of each modal feature on other modal features. For example, when analyzing ad creation failures, the error operation screen in the visual features might draw more attention to the related error descriptions in the text. The cross-modal attention layer can capture this dependency between the visual modality and the text modality and assign it appropriate weights.

[0095] After calculating the dependencies between features from different modalities in the cross-modal attention layer, this weighting information is applied to the feature fusion process. Specifically, based on the calculated attention weights, the features from different modalities are weighted and summed, ensuring that the information from different modalities is reasonably combined according to their importance and relevance in the fused features. For example, if a feature from a certain modality is more critical to analyzing the reasons for failure, then the weight of that feature will be larger during the fusion process, and its impact on the final fused features will be more significant.

[0096] Leveraging the encoder structure of the Transformer model, the features fused through the cross-modal attention layer are further encoded. The multi-layered self-attention mechanism and feedforward neural network in the encoder transform and optimize the fused features multiple times, gradually extracting higher-level, more abstract feature representations. After processing by the encoder, a unified multimodal representation vector is finally generated. This vector contains comprehensive features from text, log, and visual modalities, comprehensively reflecting the status and related factors of the ad creation task, providing rich information for subsequent failure diagnosis.

[0097] Choosing a suitable classifier is crucial for diagnosing failure causes from multimodal representation vectors. A fully connected neural network (NN) can be used as the classifier, taking a uniform multimodal representation vector as input and outputting the probability distribution of different failure cause categories. NNs can perform deep analysis and classification of the input multimodal features through multiple layers of nonlinear transformations. During training, a failure case dataset is created using labeled advertisements, with the labels representing different failure cause categories. The classifier's parameters are continuously adjusted using a backpropagation algorithm to accurately predict failure causes based on the multimodal representation vectors.

[0098] A unified multimodal representation vector is generated and input into a trained classifier. The classifier analyzes and processes this vector, calculating the probability value belonging to each failure reason category. For example, when ad creation failures may include multiple categories such as insufficient budget, unsuitable creative materials, and incorrect delivery channels, the classifier outputs a probability distribution vector. Each element in the vector corresponds to the probability of a failure reason category, and the sum of the probabilities of all elements is 1. This probability distribution provides an intuitive understanding of the most likely failure reasons for ad creation tasks and the probability of each reason, offering a reference for subsequent problem-solving. Leveraging the interpretability of attention weights in the Transformer model, the previously calculated cross-modal attention weights and multi-head self-attention weights are analyzed in depth. Visualization tools or specific analysis algorithms are used to identify attention weights that play a crucial role in the failure reason diagnosis process. For example, if a specific text fragment, log record, or visual image is assigned a high value when calculating its attention weight, it indicates that it has a significant impact on determining the failure reason.

[0099] Based on the analysis of interpretability attention weights, key modal segments and content leading to failure categories can be identified. For text modalities, it can be determined which keywords or sentences played a key role in diagnosing the failure; for log modalities, specific operation records or state changes related to the failure can be identified; for visual modalities, key images or operation steps in user operation recordings that led to the failure can be located.

[0100] By combining the failure reason category probability distribution output by the classifier with the located key modal fragments and content, fine-grained diagnostic results are generated. For example, it not only points out that the reason for the ad creation task failure is that the creative materials do not meet the requirements, but also details which aspects of the creative materials do not meet the requirements, such as size, format, and content, and provides specific evidence related to the cause in text descriptions, system logs, and user operation recordings. Such fine-grained diagnostic results help advertising practitioners more accurately understand the reasons for failure, and carry out targeted problem repair and optimization, thereby improving the efficiency and quality of ad creation.

[0101] In some embodiments, step S105 above, which involves constructing a success rate prediction model based on historical data and real-time features according to correlation features and fine-grained cause diagnosis results, and dynamically adjusting the advertising budget allocation based on the prediction results of the success rate prediction model, specifically includes: Integrate correlation features, fine-grained failure cause diagnosis results, historical data, and real-time features to construct a unified feature set; Based on a unified feature set, the mapping relationship between features and success rate is learned using historical success rate as a supervisory signal. A time-enhanced machine learning model is then used to train and generate a success rate prediction model. The success rate prediction model is used to predict the success rate of real-time advertising tasks and obtain the predicted success rate score for each task or channel. Based on the predicted success rate score, the optimal budget allocation scheme is calculated through a dynamic budget allocation algorithm, and the advertising budget allocation is adjusted according to the optimal budget allocation scheme.

[0102] In this embodiment, associated features, fine-grained failure cause diagnosis results, historical data, and real-time features are integrated to construct a unified feature set. During the integration process, different types of data are standardized to ensure they have the same dimensions and format for subsequent model training and analysis. All processed features are combined into a complete feature vector, forming a unified feature set, providing a data foundation for training the success rate prediction model.

[0103] The success rate of historical advertising campaigns is used as a monitoring signal. The historical success rate can be calculated based on the actual performance of the advertising campaigns. For example, if an advertising campaign achieves its expected conversion goal, it is considered successful with a success rate of 1; if it fails to achieve the expected goal, it is considered a failure with a success rate of 0. For cases where the goal is partially achieved, the success rate can be calculated based on the ratio of the actual conversion rate to the expected conversion rate. The success rate of each historical advertising campaign is used as a label and associated with a feature vector from a corresponding unified feature set to form a labeled training dataset.

[0104] Given the temporal characteristics of advertising data, such as variations in user behavior and market competition across different time periods, time-enhanced machine learning models are chosen for training. For example, recurrent neural network models like Long Short-Term Memory (LSTM) or Gated Recurrent Units (GRUs) can effectively handle long-term dependencies in time-series data. These models can learn from historical feature sequences, capturing patterns of feature changes over time, thereby better predicting the success rate of advertising campaigns.

[0105] The labeled training dataset is divided into training, validation, and test sets. The temporal augmentation machine learning model is trained using the training set. During training, the model continuously adjusts its parameters based on the input feature vectors and corresponding success rate labels to minimize the error between the predicted and actual success rates. The model's performance is progressively optimized using backpropagation and optimization algorithms (such as stochastic gradient descent). Simultaneously, the model is evaluated using the validation set. Based on the evaluation results on the validation set, the model's hyperparameters, such as the learning rate and the number of hidden layer nodes, are adjusted to prevent overfitting.

[0106] After multiple iterations of training, when the model's performance on the validation set stabilizes and meets requirements, training is stopped, and the model parameters are saved, generating a success rate prediction model. The generated success rate prediction model is then evaluated using a test set, calculating metrics such as prediction accuracy, recall, and F1 score to assess the model's generalization ability and predictive performance. If the model's evaluation metrics meet expectations, it is considered suitable for actual advertising success rate prediction; otherwise, further optimization or adjustment is required.

[0107] When a new real-time advertising task requires success rate prediction, relevant features of the task, including related and real-time features, are collected using a unified feature set construction method. These features are then organized into a feature vector with the same format as the training data. For example, for a new social media advertising task, the creative features of the ad, the target audience features, and the user traffic features of the current social media platform are collected to form a feature vector. This organized real-time advertising task feature vector is then input into a trained success rate prediction model. The model predicts the success rate of the advertising task based on the learned mapping relationship between features and success rate, and outputs a predicted success rate score. This score typically ranges from 0 to 1; a higher score indicates a greater probability of success for the advertising task.

[0108] If you need to predict the success rate of multiple real-time advertising tasks or advertising tasks on different channels simultaneously, you can input the feature vectors of each task or channel into the model to obtain their respective predicted success rate scores. For example, for an advertiser who is simultaneously advertising on multiple social media platforms and search engines, you can collect the advertising features of each platform and channel, input them into the model to obtain the predicted success rate scores of the advertising tasks on each platform and channel, so as to conduct comprehensive analysis and budget allocation later.

[0109] Design a dynamic budget allocation algorithm based on predicted success rate scores. The goal of this algorithm is to allocate the budget rationally according to the predicted success rate scores of each advertising task or channel, given a fixed total budget, thereby maximizing the overall success rate of advertising campaigns. For example, a weighted allocation method can be used, calculating the total budget by weighting the predicted success rate scores and allocating a corresponding budget to each advertising task or channel. Specifically, for each advertising task or channel, the allocated budget is directly proportional to the predicted success rate score; the higher the predicted success rate score, the more budget is allocated.

[0110] The predicted success rate scores of each real-time advertising task or channel are input into a dynamic budget allocation algorithm, which calculates the optimal budget allocation scheme. During the calculation process, the algorithm considers the constraints of the total budget, ensuring that the sum of the budgets allocated to each task or channel equals the total budget. Simultaneously, the algorithm dynamically adjusts the budget allocation based on changes in the predicted success rate scores to adapt to the constantly changing advertising environment. For example, if the predicted success rate score of a certain channel suddenly increases, the algorithm automatically increases the budget allocation for that channel; conversely, if the predicted success rate score of a certain channel decreases, the algorithm will correspondingly decrease its budget allocation.

[0111] Based on the calculated optimal budget allocation plan, the advertising budget is adjusted accordingly. The adjusted budget is then distributed to various advertising tasks or channels through the advertising platform's budget management system. During the adjustment process, the accuracy and timeliness of budget allocation are ensured to avoid impacting advertising effectiveness due to untimely budget adjustments. Simultaneously, the actual effectiveness of advertising campaigns and changes in predicted success rate scores are continuously monitored, and the budget allocation plan is further optimized and adjusted based on actual conditions to maximize advertising effectiveness.

[0112] In some embodiments, step S106 above, which involves converting the fine-grained cause diagnosis results and the prediction results of the success rate prediction model into reinforcement learning actions, iterating the strategy based on user feedback, and displaying the analysis results and optimization suggestions through a visual interface, specifically includes: Based on the fine-grained cause diagnosis results and success rate prediction results, a state space, action space, and reward function are constructed. Based on the state space, action space, and reward function, a reinforcement learning agent is trained using the Actor-Critic framework to output reinforcement learning actions. The reinforcement learning actions are transformed into actionable optimization suggestions, which are transmitted to the advertising system via API interface. At the same time, they are displayed to users through a visual interface and feedback is collected to obtain user feedback data. Based on user feedback data, the policy of the reinforcement learning agent is iteratively optimized using the policy gradient update method, and the action selection probability is adjusted.

[0113] In this embodiment, the state space is used to comprehensively describe the current situation during the advertising delivery process, providing a basis for decision-making for the reinforcement learning agent. First, the fine-grained cause diagnosis results are encoded. For example, if the diagnosis indicates that the advertising is ineffective due to audience targeting bias, this can be encoded as a specific numerical or vector representation, covering the specific dimensions of the bias (such as age, region, etc.) and the degree of bias. Simultaneously, the success rate prediction results are also incorporated into the state space, directly representing the probability of future success of the advertising delivery task in numerical form. These different types of information are then integrated and transformed into a unified format feature vector through feature engineering methods, collectively constituting the state space.

[0114] The action space defines all possible actions that a reinforcement learning agent can take to optimize advertising strategies. Actions can be categorized into several types, such as budget adjustment actions (increasing or decreasing the budget for specific advertising tasks or channels, with adjustments set in different percentage levels); audience targeting actions (expanding or narrowing the target audience, modifying audience characteristics such as age, gender, and interests); and advertising creative actions (changing creative elements such as images, copy, and videos in advertisements). Specific value ranges or options are assigned to each category of actions, and these actions are combined to form the complete action space.

[0115] The reward function measures the benefit gained by a reinforcement learning agent after taking a certain action, guiding the agent to learn in a direction that optimizes advertising effectiveness. The reward function is designed considering multiple factors, with advertising success rate and economic benefits as the main indicators. When an action taken by the agent increases the success rate of advertising, a positive reward is given. The size of the reward can be set according to the magnitude of the increase in success rate; for example, a certain positive reward is given for every 5% increase in success rate. Simultaneously, the cost and revenue of advertising are considered. If an action leads to an increase in advertising conversion rate, resulting in more actual revenue, a corresponding positive reward is also given. Conversely, if an action leads to a decrease in advertising effectiveness, such as a decrease in success rate or an increase in cost leading to a decrease in revenue, a negative reward is given.

[0116] The Actor-Critic framework is a reinforcement learning framework that combines policy gradient methods and value function approximation methods, consisting of an Actor network and a Critic network. The Actor network is responsible for selecting actions based on the current state; its input is a feature vector in the state space, and its output is the probability of choosing each action in the action space. The Critic network evaluates the value of the current state; its input is also a feature vector in the state space, and its output is a scalar value representing the expected cumulative reward obtained by taking the optimal action in the current state. When building these two networks, a suitable neural network structure is chosen, such as a multilayer perceptron (MLP), and the number of layers and neurons is determined based on the complexity of the problem.

[0117] Historical data from the ad delivery process is collected as training data. This data includes state information, actions taken, and corresponding reward values ​​at different times. The collected historical data is preprocessed to ensure its format conforms to the input requirements of the Actor-Critic framework. During training, the Actor-Critic framework learns the optimal strategy by continuously interacting with the environment (i.e., the ad delivery system). Specifically, first, the current state is input into the Actor network, and an action is randomly selected based on the action selection probability output by the Actor network. Then, the next state and reward value fed back by the environment after executing the action are observed. The current state, the action taken, the next state, and the reward value are combined into an experience sample and stored in the experience replay buffer. A batch of experience samples is randomly sampled from the experience replay buffer to update the parameters of the Actor and Critic networks. The Critic network updates its parameters by minimizing the mean squared error loss function, enabling it to more accurately evaluate state value. The Actor network updates its parameters using the policy gradient method based on the state value information provided by the Critic network, improving the probability of selecting a high-quality action. This process is repeated until the agent's performance reaches a stable state, meaning it can select a relatively optimized action in different states.

[0118] After sufficient training, the Actor network in the Actor-Critic framework can output the selection probability of each action in the action space based on the current input state. The action with the highest selection probability is the current optimal action output by the reinforcement learning agent, which will be used to guide the optimization of the advertising delivery strategy.

[0119] The actions output by the reinforcement learning agent are transformed into specific optimization suggestions that the advertising delivery system can understand and execute. For example, if the action is to increase the budget of an advertising channel, it is transformed into a specific budget increase amount and execution time, generating an optimization suggestion such as "Increase the budget of [advertising channel name] by 20% in the next 24 hours." For audience targeting actions, they are transformed into specific audience characteristic modification instructions, such as "Change the age range of the target audience to 25-35 years old." For ad creative actions, the creative elements that need to be replaced and their replacement content are clearly specified, such as "Replace the ad image with [new image name]."

[0120] The generated optimization suggestions are transmitted to the advertising delivery system through a predefined API interface.

[0121] Develop a visual interface to display ad campaign analysis results and optimization suggestions to users. The interface can use various formats such as charts, reports, and dashboards to intuitively present key ad performance metrics, such as success rate, conversion rate, and budget usage. Simultaneously, the optimization suggestions generated by the reinforcement learning agent should be displayed on the interface in a clear and easy-to-understand manner, for example, by listing each suggestion in a list format with detailed explanations for each suggestion. Furthermore, the visual interface should provide historical data query functionality, allowing users to view ad campaign performance and the effectiveness of optimization suggestions across different time periods.

[0122] A feedback channel is set up on the visual interface, allowing users to evaluate and provide feedback on optimization suggestions. Users can express their satisfaction with the suggestions, whether to implement them, and the effects after implementation by clicking buttons, filling out forms, etc. The collected user feedback data will be used for subsequent strategy iteration and optimization.

[0123] The collected user feedback data is organized and analyzed to extract valuable information. For example, user satisfaction with different types of optimization suggestions is statistically analyzed, along with the reasons why users implement or do not implement these suggestions, and the changes in advertising effectiveness after implementing them. User feedback data is then correlated with corresponding state, action, and reward information to form new training samples used to update the reinforcement learning agent's policy. The policy gradient update method is used to adjust the reinforcement learning agent's policy based on the user feedback data. The core idea of ​​the policy gradient method is to update the policy parameters by calculating the policy gradient, increasing the probability of the agent taking high-quality actions. Specifically, the policy gradient is calculated based on the reward information and action selection in the user feedback data, and then the parameters of the Actor network are updated along the direction of the policy gradient.

[0124] By continuously collecting user feedback data and adjusting policy parameters using the policy gradient update method, the reinforcement learning agent's policy is iteratively optimized. During each iteration, the agent's performance is observed, such as whether the success rate of ad delivery improves or user satisfaction increases. If performance improves, iteration continues in the current manner; if performance does not improve significantly or even declines, the reasons are analyzed, the parameters of the policy gradient update method are adjusted, or the processing method of user feedback data is changed, and iteration optimization is repeated. After multiple iterations, the reinforcement learning agent's policy gradually approaches optimality, enabling it to select actions that better meet user needs and ad delivery goals based on different states.

[0125] Furthermore, the construction of the state space, action space, and reward function based on the fine-grained cause diagnosis results and success rate prediction results specifically includes: Based on the fine-grained cause diagnosis results and success rate prediction results, a state space is constructed, which includes advertising account state characteristics, creation process characteristics, and environmental impact characteristics. Based on the state space, an action space is constructed, which includes actions such as adjusting the delivery strategy, optimizing creatives, and managing accounts. Based on the state space and action space, a reward function is designed. The reward function is a multi-dimensional reward mechanism that includes a success rate reward component, an efficiency reward component, and a cost reward component.

[0126] In this embodiment, the advertising account status characteristics are key information reflecting the overall operation of the advertising account. First, basic information about the advertising account is collected, such as the account balance, which directly determines the duration and scale of continuous advertising. Sufficient balances allow for more resources for ad display and promotion. Account activity is also an important characteristic, which can be measured by statistically analyzing the account's login frequency and number of operations over a certain period. Highly active accounts typically indicate that advertisers are more attentive to ad placements and may be more proactive in adjusting their strategies. Furthermore, historical ad placement data is indispensable, including ad placement volume, click-through rate, and conversion rate over different time periods. This data reflects the account's placement habits and performance trends. The collected basic advertising account information is cleaned and preprocessed to remove outliers and noisy data. Then, it is encoded and quantified according to certain rules, such as dividing the account balance into different levels based on different intervals, and calculating a comprehensive score for activity based on login frequency and number of operations. Finally, this information is integrated into an advertising account status feature vector, which serves as part of the state space.

[0127] The creation process features primarily focus on key stages throughout the entire advertising journey, from planning to delivery. In the advertising planning stage, the ad's theme and target audience are crucial. The clarity and appeal of the ad theme influence user attention; for example, a fashion and beauty ad theme that stays on trend is more likely to attract the target audience's attention. The accuracy of the target audience definition determines the ad's effectiveness. By analyzing the target audience's age, gender, location, interests, and other characteristics, we can assess whether the ad can accurately reach potential customers. In the ad production stage, the quality and diversity of creative materials are key features, including image clarity, color schemes, video length, and content creativity. High-quality materials increase ad appeal and click-through rates. In the ad review stage, the approval time and review comments are also important information. Excessive approval time may affect timely ad delivery, while review comments can provide direction for ad optimization. By extracting and organizing these key pieces of information from the creation process—for example, categorizing and coding ad themes according to popularity, quantifying target audience characteristics, and scoring creative material quality using professional evaluation tools—a creation process feature vector is ultimately formed and incorporated into the state space.

[0128] Environmental impact characteristics encompass various factors in the external environment in which advertising placement takes place. Market trends are a crucial environmental factor; analyzing industry reports and market research data helps understand current market trends, consumer preferences, and competitive landscape. Competitor advertising activities are also key environmental characteristics, including their advertising channels, frequency, content, and promotional activities. Monitoring competitors' advertising dynamics reveals the intensity of market competition, providing a reference for one's own advertising strategy. Changes in policies and regulations also impact advertising placement. For example, certain industries may be subject to strict advertising regulations, requiring timely adjustments to advertising content to comply with policy requirements. Collecting and analyzing these environmental impact factors—such as scoring market trends across different dimensions, converting competitor advertising activities into specific indicator data, and categorizing policy and regulatory changes—can ultimately lead to the integration of these factors into an environmental impact characteristic vector, forming a complete state space.

[0129] Adjusting advertising strategies aims to optimize the way and scope of ad placement to improve ad exposure and conversion rates. Regarding channels, actions to increase or decrease placement on specific channels can be designed based on their characteristics and target user groups. Regarding timing, actions to adjust ad placement periods can be designed by analyzing user behavior and activity times. Regarding geographic targeting, actions to expand or narrow the geographic scope can be designed based on the product's sales region and target market. If the product is sold nationwide and certain regions have significant market potential, actions to expand placement in those regions can be designed; conversely, if sales in certain regions are unsatisfactory and placement costs are high, actions to reduce placement in those regions can be designed.

[0130] Creative optimization primarily targets and improves the creative elements of advertisements, such as images, videos, and copy, to enhance ad appeal and click-through rates. For image creatives, actions such as replacing images, adjusting image size, and optimizing image colors can be designed. For video creatives, actions such as editing videos, adding effects, and adjusting video length can be designed. If a video is too long and users are impatient to watch it all, actions can be designed to shorten it. For copy creatives, actions such as modifying copy content, adjusting copy style, and optimizing copy layout can be designed. For example, if the current ad copy is too stiff, actions can be designed to make the copy more vivid and engaging.

[0131] Account management primarily involves managing advertising account funds, setting permissions, and analyzing data. In terms of fund management, actions such as top-ups, withdrawals, and budget allocation adjustments can be designed. For example, when the account balance is insufficient, a top-up action can be designed to ensure normal ad execution; when the budget for a particular advertising project is depleted too quickly, a budget allocation adjustment action can be designed to allocate funds reasonably to other projects. Regarding permission settings, actions such as adding or deleting account administrators and modifying administrator permissions can be designed.

[0132] The success rate reward component is primarily used to encourage reinforcement learning agents to take actions that improve the success rate of ad delivery. Ad delivery success rate can be measured by various metrics, such as click-through rate, conversion rate, and registration rate. Depending on different advertising goals and business needs, appropriate key metrics should be selected as the standard for measuring success rate.

[0133] Efficiency rewards are designed to incentivize agents to improve ad delivery efficiency, meaning achieving better results with the same resource investment. Efficiency rewards can be considered from two aspects: ad delivery speed and resource utilization. Regarding ad delivery speed, a positive reward is given if the agent can quickly display ads to the target user, reducing ad delay time.

[0134] The cost-reward component is used to incentivize agents to reduce advertising costs while ensuring advertising effectiveness. Advertising costs include ad production costs, distribution channel fees, and personnel management costs. Positive rewards are given when an agent's actions reduce advertising costs.

[0135] By combining success rate reward components, efficiency reward components, and cost reward components according to certain weights, a multi-dimensional reward mechanism is formed to comprehensively evaluate the quality of the actions taken by the agent and guide the agent to learn in the direction of optimizing advertising effectiveness, improving efficiency, and reducing costs.

[0136] Reference Figure 2 An embodiment of the present invention provides a diagnostic system 2 for the reasons for failure in ad creation, the system 2 specifically comprising: Data acquisition module 201 is used to collect multi-source data during the advertisement creation process. The multi-source data includes task metadata, system logs, text descriptions, and user operation recordings. The knowledge graph module 202 is used to construct a knowledge graph of the advertising creation process based on multi-source data. The nodes of the knowledge graph are used to represent advertising entities, and the edges are used to represent the relationships between entities. The feature extraction module 203 is used to extract the associated features in the advertising creation process by embedding nodes and reasoning about relationships in the knowledge graph through graph neural networks; The failure diagnosis module 204 is used to perform cross-modal alignment and fusion of text descriptions, system logs and user operation recordings based on correlation features using the Transformer model to generate fine-grained cause diagnosis results for failed tasks. The budget allocation module 205 is used to build a success rate prediction model based on correlation features and fine-grained cause diagnosis results, according to historical data and real-time features, and dynamically adjust the advertising budget allocation based on the prediction results of the success rate prediction model. The reinforcement learning module 206 is used to transform the fine-grained cause diagnosis results and the prediction results of the success rate prediction model into reinforcement learning actions, iterate the strategy based on user feedback, and display the analysis results and optimization suggestions through a visual interface.

[0137] It is understandable that, such as Figure 1 The content of the illustrated ad creation failure diagnosis method embodiment is applicable to the ad creation failure diagnosis system embodiment. The specific functions implemented by the ad creation failure diagnosis system embodiment are as follows: Figure 1 The method for diagnosing the failure of ad creation shown in the example is the same, and the beneficial effects achieved are the same as those described above. Figure 1 The beneficial effects achieved by the illustrated method for diagnosing the reasons for failure in ad creation are the same.

[0138] It should be noted that the information interaction and execution process between the above systems are based on the same concept as the method embodiments of the present invention. For details on their specific functions and technical effects, please refer to the method embodiments section, which will not be repeated here.

[0139] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the above-described division of functional units and modules is merely an example. In practical applications, the above functions can be assigned to different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiments can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit. Furthermore, the specific names of the functional units and modules are only for easy differentiation and are not intended to limit the scope of protection of this application. The specific working process of the units and modules in the above system can be referred to the corresponding process in the foregoing method embodiments, and will not be repeated here.

[0140] Reference Figure 3 The present invention also provides a computer device 3, including: a memory 302 and a processor 301, and a computer program 303 stored on the memory 302. When the computer program 303 is executed on the processor 301, it implements the method for diagnosing the failure of ad creation as described in any of the above methods.

[0141] The computer device 3 may be a desktop computer, laptop, handheld computer, or cloud server, etc. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will understand that... Figure 3 The computer device 3 is merely an example and does not constitute a limitation on the computer device 3. It may include more or fewer components than shown in the figure, or combine certain components, or different components, such as input / output devices, network access devices, etc.

[0142] The processor 301 may be a Central Processing Unit (CPU), or it may be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor.

[0143] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may be an external storage device of the computer device 3, such as a plug-in hard disk, smart media card (SMC), secure digital (SD) card, flash card, etc., equipped on the computer device 3. Furthermore, the memory 302 may include both internal and external storage units of the computer device 3. The memory 302 is used to store the operating system, applications, boot loader, data, and other programs, such as the program code of the computer program. The memory 302 can also be used to temporarily store data that has been output or will be output.

[0144] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the method for diagnosing the failure of ad creation as described in any of the above methods.

[0145] In this embodiment, if the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include at least: any entity or device capable of carrying computer program code to a photographing device / terminal device, a recording medium, a computer memory, a read-only memory (ROM), a random access memory (RAM), an electrical carrier signal, a telecommunication signal, and a software distribution medium. Examples include USB flash drives, portable hard drives, magnetic disks, or optical disks. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electrical carrier signals or telecommunication signals.

[0146] In the above embodiments, the descriptions of each embodiment have different focuses. For parts that are not described in detail or recorded in a certain embodiment, please refer to the relevant descriptions of other embodiments.

[0147] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0148] In the embodiments disclosed in this application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual coupling or direct coupling or communication connection may be through some interfaces; the indirect coupling or communication connection between devices or units may be electrical, mechanical, or other forms.

[0149] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

Claims

1. A method for diagnosing the reasons for failure in ad creation, characterized in that, The method specifically includes: Collect multi-source data during the ad creation process, including task metadata, system logs, text descriptions, and user operation recordings; Based on multi-source data, a knowledge graph of the advertising creation process is constructed, wherein the nodes of the knowledge graph are used to represent advertising entities and the edges are used to represent the relationships between entities; By using graph neural networks to embed nodes and reason about relationships in the knowledge graph, we can extract the associated features in the ad creation process. Based on correlation features, the Transformer model is used to perform cross-modal alignment and fusion of text descriptions, system logs and user operation recordings to generate fine-grained cause diagnosis results for failed tasks. Based on correlation features and fine-grained cause diagnosis results, a success rate prediction model is constructed according to historical data and real-time features, and the advertising budget allocation is dynamically adjusted based on the prediction results of the success rate prediction model. The results of fine-grained cause diagnosis and success rate prediction models are transformed into reinforcement learning actions. The strategy is iterated based on user feedback, and the analysis results and optimization suggestions are displayed through a visual interface.

2. The method according to claim 1, characterized in that, The aforementioned knowledge graph, constructed based on multi-source data, represents the advertising creation process. Nodes in this knowledge graph represent advertising entities, and edges represent relationships between entities. Specifically, it includes: Based on multi-source data, natural language processing technology is used to identify advertising entities and their attributes. The advertising entities include advertising accounts, advertising plans, advertising groups, advertising creatives, creative elements, review rules, and delivery channels. Identify the relationship types between advertising entities, including membership, constraint, triggering, and dependency relationships; Using advertising entities as nodes and relationship types as edges, a knowledge graph of the advertising creation process is constructed, and attribute features are assigned to nodes and edges, including creation success rate, review time, material size and rule strength. Based on user interface operation recordings, multimodal coding technology is used to embed the interface operation sequence and interface state change information as vectors into a knowledge graph and associate them with corresponding nodes.

3. The method according to claim 1, characterized in that, The method of using graph neural networks to embed nodes and infer relationships in the knowledge graph to extract relevance features during the ad creation process specifically includes: Graph structure data is obtained from the knowledge graph, which includes a set of nodes, a set of edges, node attributes, and edge attributes, and the nodes and edges are initialized with features. Based on graph structure data, node embedding representation learning is performed through a graph neural network model, neighbor node information is aggregated using a graph attention network, and different weights are assigned to different neighbors through an attention mechanism. Through multi-layer message passing and node state updates, high-dimensional embedding vectors of all nodes are output. These high-dimensional embedding vectors are used to represent the semantic and structural information of the advertising entity and its context. Based on high-dimensional embedding vectors, global average pooling or attention-weighted hierarchical readout operations are used to extract full-graph representations, which are used to describe the overall state and key features of the ad creation process. Based on full-image features, we extract the associated features in the ad creation process.

4. The method according to claim 1, characterized in that, Based on correlation features, the Transformer model is used to perform cross-modal alignment and fusion of text descriptions, system logs, and user operation recordings to generate fine-grained cause diagnosis results for failed tasks, specifically including: Based on the association features, the Transformer model is used to map text features, log features and visual features to the same semantic space, and the multi-head self-attention mechanism is used to calculate the association weights between modalities. During feature fusion, intermodal dependencies are captured through a cross-modal attention layer, and a unified multimodal representation vector is generated with the help of the encoder of the Transformer model. Based on multimodal representation vectors, a classifier is used to diagnose failure causes and output the probability distribution of different failure cause categories. By using interpretable attention weights to locate key modal fragments and content that lead to failure categories, and combining them with probability distributions, fine-grained diagnostic results are generated.

5. The method according to claim 1, characterized in that, The process involves constructing a success rate prediction model based on correlation features and fine-grained cause diagnosis results, using historical data and real-time features, and dynamically adjusting the advertising budget allocation based on the prediction results of the success rate prediction model. Specifically, this includes: Integrate correlation features, fine-grained failure cause diagnosis results, historical data, and real-time features to construct a unified feature set; Based on a unified feature set, the mapping relationship between features and success rate is learned using historical success rate as a supervisory signal. A time-enhanced machine learning model is then used to train and generate a success rate prediction model. The success rate prediction model is used to predict the success rate of real-time advertising tasks and obtain the predicted success rate score for each task or channel. Based on the predicted success rate score, the optimal budget allocation scheme is calculated through a dynamic budget allocation algorithm, and the advertising budget allocation is adjusted according to the optimal budget allocation scheme.

6. The method according to claim 1, characterized in that, The process of converting fine-grained cause diagnosis results and success rate prediction model prediction results into reinforcement learning actions, iterating strategies based on user feedback, and displaying analysis results and optimization suggestions through a visual interface specifically includes: Based on the fine-grained cause diagnosis results and success rate prediction results, a state space, action space, and reward function are constructed. Based on the state space, action space, and reward function, a reinforcement learning agent is trained using the Actor-Critic framework to output reinforcement learning actions. The reinforcement learning actions are transformed into actionable optimization suggestions, which are transmitted to the advertising system via API interface. At the same time, they are displayed to users through a visual interface and feedback is collected to obtain user feedback data. Based on user feedback data, the policy of the reinforcement learning agent is iteratively optimized using the policy gradient update method, and the action selection probability is adjusted.

7. The method according to claim 6, characterized in that, The process of constructing a state space, action space, and reward function based on fine-grained cause diagnosis results and success rate prediction results specifically includes: Based on the fine-grained cause diagnosis results and success rate prediction results, a state space is constructed, which includes advertising account state characteristics, creation process characteristics, and environmental impact characteristics. Based on the state space, an action space is constructed, which includes actions such as adjusting the delivery strategy, optimizing creatives, and managing accounts. Based on the state space and action space, a reward function is designed. The reward function is a multi-dimensional reward mechanism that includes a success rate reward component, an efficiency reward component, and a cost reward component.

8. A system for diagnosing reasons for advertising creation failures, characterized in that, The system specifically includes: The data acquisition module is used to collect multi-source data during the ad creation process. The multi-source data includes task metadata, system logs, text descriptions, and user operation recordings. The knowledge graph module is used to construct a knowledge graph of the advertising creation process based on multi-source data. The nodes of the knowledge graph are used to represent advertising entities, and the edges are used to represent the relationships between entities. The feature extraction module is used to extract relevant features in the ad creation process by embedding nodes and reasoning about relationships in the knowledge graph through graph neural networks. The failure diagnosis module is used to perform cross-modal alignment and fusion of text descriptions, system logs and user operation recordings based on correlation features using the Transformer model, and generate fine-grained cause diagnosis results for failed tasks. The budget allocation module is used to build a success rate prediction model based on correlation features and fine-grained cause diagnosis results, and dynamically adjust the advertising budget allocation based on the prediction results of the success rate prediction model. The reinforcement learning module is used to transform the results of fine-grained cause diagnosis and the prediction results of the success rate prediction model into reinforcement learning actions, iterate the strategy based on user feedback, and display the analysis results and optimization suggestions through a visual interface.

9. A computer device, characterized in that, include: The memory and processor, and the computer program stored in the memory, when the computer program is executed on the processor, implement the method for diagnosing the reasons for failure in ad creation as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, It stores a computer program, which, when executed by a processor, implements the method for diagnosing the reasons for failure in ad creation as described in any one of claims 1 to 7.

Images