Advertisement attribute consistency monitoring method, system, equipment and medium

Abstract

The invention discloses an advertisement attribute consistency monitoring method, system and device and a medium, and the method specifically comprises the steps: collecting snapshots and metadata of advertisement attributes in each processing link of a creation process, and constructing an advertisement attribute consanguinity map; extracting time sequence data of attribute values based on the advertisement attribute blood relationship map, modeling a normal fluctuation mode of attributes by using an anomaly detection algorithm, and giving risk early warning in advance for abnormal fluctuation deviating from the normal mode; analyzing conduction paths with inconsistent attributes and positioning root cause nodes by utilizing a pre-trained cause and effect graph model and combining current inconsistent node information and the state of the advertisement attribute blood relationship map; and inputting the alarm event into an alarm strategy model based on reinforcement learning training, and dynamically deciding alarm sending, priority and notification channels. According to the invention, efficient, accurate and real-time advertisement attribute consistency monitoring is realized, the accuracy and efficiency of advertisement putting are improved, and the requirement of continuous development of advertisement services can be met.

Description

Technical Field

[0001] This invention relates to the field of advertising operation technology, and in particular to a method, system, device and medium for monitoring the consistency of advertising attributes. Background Technology

[0002] In the field of advertising, the ad creation process is a complex and crucial one. The accuracy and consistency of ad attributes at different stages (such as panel input, asynchronous processing, and media API output) play a decisive role in the precise targeting and efficient operation of ads. However, in practice, inconsistencies in ad attributes frequently occur in existing ad creation processes. This directly leads to ad placement errors or inefficiencies, causing numerous adverse effects on advertisers and advertising platforms.

[0003] Traditional methods for monitoring ad attribute consistency primarily rely on manual log checks. This monitoring approach has several limitations, including: 1. Low efficiency of manual inspection: In the ad creation process, ad attributes involve numerous stages and a large amount of data. Manual inspection requires comparing ad attribute values ​​at different stages one by one. Faced with massive amounts of data and complex processes, manual operation is not only time-consuming and labor-intensive, but also difficult to complete a comprehensive inspection in a short period of time, resulting in extremely low monitoring efficiency. For example, in a large-scale advertising campaign, tens of thousands of ad attribute data may be generated every day, making manual inspection of each one virtually impossible.

[0004] 2. Prone to Errors: Manual inspection is susceptible to subjective factors. Different inspectors may have different understandings and judgment standards regarding data. Furthermore, prolonged repetitive inspections can lead to inspector fatigue, resulting in misjudgments. For example, inspectors may overlook subtle but crucial attribute differences due to negligence, or make incorrect judgments about similar but not identical attribute values. These errors directly impact the accuracy of ad placement.

[0005] 3. Lack of real-time monitoring: Traditional monitoring methods often fail to detect inconsistencies in ad attributes in real time. Manual checks are typically conducted at specific points in time or during specific periods. This means that if inconsistencies in ad attributes occur between checks, they may not be detected in time, leading to delayed problem discovery. For example, if inconsistencies in ad attributes go undetected during critical advertising periods, the optimal advertising opportunity may be missed, resulting in financial losses for advertisers.

[0006] 4. Insufficient Scalability: As the advertising business continues to develop, the number and complexity of advertising attributes are constantly increasing. New advertising types, delivery channels, and business rules are constantly emerging, making the advertising attribute system increasingly large and complex. Traditional manual monitoring methods are difficult to adapt to these rapidly changing business needs and cannot flexibly expand to cover new advertising attributes and monitoring dimensions. For example, when introducing a new advertising platform or launching a new advertising creative format, traditional methods require redesigning inspection rules and processes, consuming significant manpower and time costs.

[0007] 5. Low Detail of Alarm Information: Existing monitoring methods often provide relatively simple alarm information, typically only indicating inconsistencies in advertising attributes, but lacking more granular information such as the specific attribute values ​​of the inconsistencies, the exact time the problem occurred, and related log information. This makes it difficult for operators to quickly locate the root cause and specific location of the problem after receiving an alarm, increasing the difficulty of troubleshooting and resolving the issue. For example, operators may need to manually search for relevant information in a large amount of log data based on simple alarm information to gradually pinpoint the problem, which undoubtedly greatly prolongs the problem-solving time. Summary of the Invention

[0008] The purpose of this invention is to provide a method, system, device, and medium for monitoring the consistency of advertising attributes, which achieves efficient, accurate, and real-time monitoring of the consistency of advertising attributes, improves the accuracy and efficiency of advertising placement, and can adapt to the ever-evolving needs of the advertising business, thereby solving at least one of the aforementioned problems in the prior art.

[0009] In a first aspect, the present invention provides a method for monitoring the consistency of advertising attributes, the method specifically comprising: Collect snapshots and metadata of ad attributes at each stage of the creation process to construct an ad attribute lineage graph for recording the entire data change history; Based on the advertising attribute lineage map, time-series data of attribute values ​​are extracted, and anomaly detection algorithms are used to model the normal fluctuation pattern of attributes, and risk warnings are issued in advance for abnormal fluctuations that deviate from the normal pattern. When an actual inconsistency in attributes is detected or a risk warning is received, a pre-trained causal graph model is used to analyze the transmission path of attribute inconsistency and locate the root cause node by combining the current inconsistency node information and the state of the advertising attribute lineage graph. The root cause node, its affected attribute nodes, and the propagation path are dynamically highlighted in the visualization interface. The alarm events are input into the alarm strategy model trained by reinforcement learning. Based on historical alarm response data and the current event context, the model dynamically decides on the issuance, priority, and notification channels of alarms. By using intelligent alarm aggregation algorithms, multiple alarm events with the same or similar root cause nodes are merged into a single aggregated alarm event.

[0010] Secondly, the present invention provides an advertising attribute consistency monitoring system, the system specifically comprising: The data recording module is used to collect snapshots and metadata of advertising attributes at each processing stage of the creation process, and to build an advertising attribute lineage map to record the history of data changes throughout the entire chain. The wind direction warning module is used to extract time-series data of attribute values ​​based on the lineage graph of advertising attributes, and to use anomaly detection algorithms to model the normal fluctuation pattern of attributes and issue risk warnings in advance for abnormal fluctuations that deviate from the normal pattern. The risk attribution module is used to analyze the transmission path of attribute inconsistency and locate the root cause node when an actual attribute inconsistency is detected or a risk warning is received. It utilizes a pre-trained causal graph model, combined with the current inconsistency node information and the state of the advertising attribute lineage graph. The dynamic display module is used to dynamically highlight the root cause node, the affected attribute nodes, and the propagation path in the visualization interface. The alarm decision module is used to input alarm events into the alarm strategy model trained by reinforcement learning, and dynamically decide the issuance, priority and notification channel of alarms based on historical alarm response data and current event context. The alarm aggregation module is used to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event using an intelligent alarm aggregation algorithm.

[0011] 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 advertising attribute consistency monitoring method as described in any of the above methods.

[0012] 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 advertising attribute consistency monitoring method as described in any of the above methods.

[0013] Compared with the prior art, the present invention has at least one of the following technical effects: 1. This invention enables efficient, accurate, and real-time monitoring of advertising attribute consistency, improving the accuracy and efficiency of advertising placement and adapting to the ever-evolving needs of the advertising business.

[0014] 2. This invention extracts time-series data of attribute values ​​based on advertising attribute lineage graphs and uses anomaly detection algorithms to model the normal fluctuation patterns of attributes. It can issue risk warnings in advance for abnormal fluctuations that deviate from the normal patterns, realize real-time monitoring, and promptly detect potential problems.

[0015] 3. When the present invention detects an actual inconsistency in attributes or receives a risk warning, it uses a pre-trained causal graph model to analyze the transmission path of the attribute inconsistency and locate the root cause node. Combined with the dynamic highlighting of the visualization interface, it helps operators quickly locate the root cause of the problem.

[0016] 4. This invention improves the intelligence level of alarm processing by inputting alarm events into an alarm strategy model trained by reinforcement learning, and dynamically deciding on the issuance, priority and notification channel of alarms.

[0017] 5. This invention utilizes an intelligent alarm aggregation algorithm to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event, reducing the redundancy of alarm information and enabling operators to understand the problem situation more clearly.

[0018] 6. This invention extracts time-series feature data based on pedigree graphs, uses unsupervised algorithms to model normal fluctuation patterns, calculates abnormal quantitative scores in real time and provides early warnings, thus achieving accurate early detection of abnormal fluctuations in advertising attributes.

[0019] 7. This invention extracts alarm feature vectors and represents them as small directed graphs, performs incremental clustering to calculate composite similarity, and merges multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event, thereby reducing alarm redundancy. Attached Figure Description

[0020] 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.

[0021] Figure 1 This is a flowchart illustrating an advertising attribute consistency monitoring method according to an embodiment of the present invention; Figure 2 This is a schematic diagram of the structure of an advertising attribute consistency monitoring 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

[0022] 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.

[0023] 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.

[0024] 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.

[0025] 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]."

[0026] 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.

[0027] 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.

[0028] 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 an embodiment of the advertising attribute consistency monitoring method disclosed in this invention is shown below: S101 collects snapshots and metadata of advertising attributes at each processing stage of the creation process, and constructs an advertising attribute lineage graph to record the history of data changes across the entire chain.

[0029] In this embodiment, data collection points are set up at various processing stages of the ad creation process. These processing stages include panel input, asynchronous processing, and media API output. At the panel input stage, when advertisers or relevant operators enter ad attribute information on the ad creation panel, the data collection points capture these input operations in real time. For example, for ad attributes such as title, description, and target audience, once the operator completes the input and submits, the data collection points record these attribute values ​​and metadata information such as the input timestamp. The input timestamp, as important metadata, accurately records the input time of ad attributes at this stage, providing a temporal dimension for subsequent lineage mapping.

[0030] In asynchronous processing, ad attributes may undergo a series of complex processing logics, such as data cleaning, format conversion, and rule validation. Data collection points are set at the beginning and end of each asynchronous processing step. Taking data cleaning as an example, when ad attribute data enters the cleaning module, the collection point records the attribute value and entry time upon entry; after the cleaning operation, the processed attribute value and completion time are collected again. Simultaneously, the identification information of this asynchronous processing step is also recorded, such as whether it is a data cleaning step, a format conversion step, etc., as well as the identifier of the specific server or processing thread processing the ad attribute. This metadata information helps to clarify the flow and processing status of ad attributes during asynchronous processing.

[0031] For the media API output stage, when ad attribute data is ready to be sent to an external media platform via the media API, data collection points are set up before and after data transmission. Before transmission, information such as the ad attribute values ​​to be sent, transmission time, and the target media platform's identifier are collected; after transmission, the transmission result is recorded, such as whether the transmission was successful and the reason for transmission failure (if any). Through these data collection points, the status and changes of ad attributes during the output stage can be fully understood.

[0032] After collecting snapshots and metadata from each processing stage, the ad attribute lineage graph is constructed. Using ad attributes as nodes and the flow relationships of attributes between different processing stages as edges, an initial graph structure is built. For example, the ad attribute in the panel input stage is taken as the starting node. Based on the processing order and flow relationships of this attribute in asynchronous processing stages, nodes corresponding to subsequent processing stages are added sequentially, and edges are connected to these nodes to represent the attribute's flow path. Simultaneously, the collected metadata information is associated with the corresponding nodes and edges. For example, the input timestamp is associated with the node in the panel input stage, the identifier and processing time of the asynchronous processing stage are associated with the corresponding intermediate node, and the sending result and target media platform identifier are associated with the node in the media API output stage.

[0033] As the ad creation process continues, new ad attribute snapshots and metadata are continuously collected, and the ad attribute lineage graph is updated in real time. When a new ad creation task is initiated, an independent flow path and node information are constructed for the ad attributes in the new task using the method described above, and integrated into the entire lineage graph. For existing ad attributes, if the attribute value changes in subsequent processing stages, the changed attribute value and change time, among other metadata, are recorded at the corresponding node in the graph, ensuring that the lineage graph accurately records the data change history throughout the entire process. In this way, a complete, accurate, and real-time updated ad attribute lineage graph is constructed, providing a solid foundation for subsequent operations such as anomaly detection and root cause analysis.

[0034] S102, extract time-series data of attribute values ​​based on the lineage graph of advertising attributes, use anomaly detection algorithm to model the normal fluctuation pattern of attributes, and issue risk warnings in advance for abnormal fluctuations that deviate from the normal pattern.

[0035] In this embodiment, time-series data of attribute values ​​are extracted from a pre-constructed ad attribute lineage graph. The ad attribute lineage graph records the entire data change history of ad attributes throughout the creation process, with each attribute node associated with attribute values ​​and related metadata at different points in time. Taking a specific ad attribute as an example, such as the click-through rate (CTR) attribute, all nodes corresponding to this attribute are located in the lineage graph. These nodes cover different stages, from the panel input stage to subsequent asynchronous processing stages and media API output stages. The CTR attribute values ​​recorded at each node are extracted sequentially according to time, and the timestamp information associated with each attribute value is obtained. These timestamps accurately record the change time of the attribute value at each processing stage. The extracted CTR attribute values ​​and corresponding timestamps are combined to form the time-series data sequence of the ad's CTR attribute. Using the same method, corresponding time-series data sequences can also be extracted for other ad attributes, such as impressions and conversion rates.

[0036] We select an appropriate anomaly detection algorithm to model the normal fluctuation pattern of the attribute. Here, we take a statistical method-based anomaly detection algorithm as an example. For the extracted time-series data sequence of the ad click-through rate attribute, we first calculate some basic statistical characteristics of the sequence, such as the mean and standard deviation. The mean reflects the average level of the ad click-through rate overall, while the standard deviation reflects the dispersion of the data. Based on these statistical characteristics, we can construct a range of normal fluctuation. For example, we assume that we consider the click-through rate to be within the range of the mean plus or minus three times the standard deviation to be within the normal fluctuation range. In this way, we establish a normal fluctuation pattern model for the ad click-through rate attribute. For other ad attributes, we also use similar statistical methods to calculate the corresponding statistical characteristics based on their time-series data sequences and construct a normal fluctuation range model.

[0037] After constructing the normal fluctuation pattern model, we continuously monitor the time-series data of new ad attributes. When a new ad creation task is generated or the ad attributes of an existing task change during subsequent processing, new attribute values ​​are continuously added to the time-series data. For each newly added attribute value, we compare it with the previously constructed normal fluctuation pattern model. Taking ad click-through rate (CTR) as an example, when a new CTR attribute value falls within the previously determined normal fluctuation range (mean plus or minus three standard deviations), it is considered to be within the normal fluctuation range and no special processing is performed. However, when a new CTR attribute value exceeds this normal fluctuation range, it is determined that the attribute value has deviated from the normal pattern and exhibits abnormal fluctuation.

[0038] Once abnormal fluctuations in attribute values ​​are detected, the system immediately issues a risk warning. Risk warning information can be conveyed to relevant personnel in various ways. For example, it can flash a prominent color (such as red) on the system's visual interface, while simultaneously displaying a pop-up message containing detailed information such as the abnormal attribute name, abnormal attribute value, and the time of occurrence. Risk warning information can also be sent to pre-defined operators or managers via SMS or email. For instance, when an abnormal fluctuation in the conversion rate of an advertisement is detected, the system highlights the conversion rate attribute node of that advertisement on the visual interface and pops up a message informing the operator that the conversion rate is abnormal, including the specific abnormal value and the time of occurrence. Simultaneously, an SMS notification is sent to relevant managers so they can take timely measures, such as checking for problems with the advertising strategy and target audience settings, thereby avoiding issues such as poor campaign performance or wasted resources due to abnormal fluctuations in advertising attributes.

[0039] S103. When an actual inconsistency in attributes is detected or a risk warning is received, a pre-trained causal graph model is used to analyze the transmission path of attribute inconsistency and locate the root cause node by combining the current inconsistent node information and the state of the advertising attribute lineage graph.

[0040] In this embodiment, after the advertising attribute consistency monitoring system detects an actual inconsistency in attributes or receives a risk warning, the system immediately obtains the node information of the inconsistency. This node information includes the name of the inconsistent advertising attribute, the specific location identifier of the attribute in the advertising attribute lineage graph (e.g., node number), and the timestamp of the inconsistency. For example, if the system detects that the advertising delivery region attribute is inconsistent between the media API output stage and the panel input stage, it will record the attribute name "delivery region," its corresponding node number in the lineage graph, and the precise time when the inconsistency occurred.

[0041] Simultaneously, the system continuously monitors the current status of the ad attribute lineage graph. This graph records the entire data change history of ad attributes throughout the creation process, reflecting the relationships between attribute nodes, data flow, and the current attribute values ​​of each node. The system extracts information about other nodes associated with inconsistent attribute nodes, including upstream nodes (the source nodes affecting the attribute value) and downstream nodes (the subsequent nodes affected by the attribute value). For example, for the "Target Region" attribute node, its upstream node might be the node that sets the target region in the panel input stage, and its downstream node might be a related node in subsequent stages such as ad targeting and cost calculation.

[0042] The acquired inconsistency node information and the current state of the ad attribute lineage graph are input into a pre-trained causal graph model. This pre-trained causal graph model is trained based on a large amount of historical ad attribute data and known causal relationships, enabling it to understand the causal logic between ad attributes. For example, it learns from historical data that changes in the geographic targeting attribute may affect the ad's targeting range, thereby impacting attributes such as ad impressions and click-through rate.

[0043] Upon receiving input information, the causal graph model analyzes the data according to its internally constructed causal logic. Starting from the inconsistent node, it traces upwards along the data flow in the ancestry graph, looking for the source of the inconsistency. Simultaneously, it examines other nodes affected by the inconsistent attribute to comprehensively understand the transmission path of the inconsistency. For example, in the case of an inconsistent "targeting region" attribute, the causal graph model traces upwards to the panel input stage to check if an error occurred when setting the targeting region. It also examines the ad targeting nodes to analyze whether the inconsistency in targeting region leads to ads being placed in the wrong areas and whether it further affects ad exposure and other attributes.

[0044] During the analysis of the transmission path, the causal graph model evaluates and ranks potential causal nodes based on factors such as the strength of causal relationships between nodes and their frequency of occurrence in historical data. Ultimately, it identifies the root cause node most likely to lead to attribute inconsistencies. For example, if historical data shows that operational errors in the panel input stage are the main cause of inconsistent delivery locations in similar situations, and the current causal graph model analysis reveals anomalies in the relevant data of the panel input stage, then the operational node for setting the delivery location in the panel input stage will be identified as the root cause node.

[0045] The system feeds back the root cause information obtained from the location to the monitoring personnel so that they can take timely measures to repair and adjust, ensuring the accuracy and efficiency of advertising. For example, the root cause information is displayed on the monitoring system's visual interface, and relevant operators and managers are notified via SMS, email, etc., informing them of the specific location of the root cause and the possible causes of the problem.

[0046] S104 dynamically highlights the root cause node, affected attribute nodes, and propagation paths in the visualization interface.

[0047] Furthermore, firstly, the system receives analysis results data from the causal graph model and constructs a unified visual state data model containing root cause nodes, propagation paths, and related context information based on the analysis results data. Then, the unified visual state data model is published via an event bus to drive updates to the visualization rendering engine. Subsequently, in the visualization interface, layered canvas technology is used to render the bottom layer representing the system's static topology, the middle layer representing attribute data flow, and the upper layer for dynamic highlighting. When the visualization rendering engine receives an update event, it immediately updates the corresponding topology in the upper canvas based on the root cause node information in the state data model. The node graphics are given a first visual effect to highlight the root cause; simultaneously, based on the transmission path information in the state data model, a path line with dynamic animation effects is rendered along the data flow trajectory of the intermediate layer in the upper canvas to visualize the direction of fault transmission; furthermore, interactive response capabilities are added to the highlighted root cause node and the affected nodes on the transmission path, allowing users to drill down to view detailed diagnostic information and related context of the node through click operations; finally, interactive controls that are linked with the visualization view are provided, including a timeline playback control for switching the display focus and an alarm list panel that is linked bidirectionally with canvas elements to support comprehensive diagnostic analysis.

[0048] In this embodiment, after completing the root cause analysis of inconsistencies in advertising attributes, the cause-effect graph model sends the analysis results to the visualization processing module. The visualization processing module receives this analysis results, which includes detailed information about the root cause nodes, such as their identifier, name, and location within the advertising attribute lineage graph; transmission path information, such as the order of nodes along the transmission path and the connections between nodes; and contextual information, such as the time of the attribute inconsistency and the relevant advertising campaigns. Based on this analysis results, the visualization processing module constructs a unified visualization state data model. This model integrates and structures the root cause nodes, transmission paths, and contextual information for subsequent display and interactive operations. For example, root cause node information is stored according to a specific data structure for easy and quick querying and location; transmission path information is encoded to clearly represent the direction of data flow and the nodes along the path.

[0049] The visualization processing module publishes the constructed unified visualization state data model through the event bus. The event bus, as a message passing mechanism, ensures that the information from the state data model is accurately and promptly transmitted to the visualization rendering engine. The visualization rendering engine, the key component responsible for transforming the state data model into a visual display, continuously listens for messages on the event bus. Upon receiving an event published by the unified visualization state data model, the visualization rendering engine updates its internal state based on the data in the model, preparing for subsequent interface rendering.

[0050] The visualization interface employs a layered canvas technique for rendering. The bottom canvas renders the static topology of the system, showcasing the fixed structure and relationships of each processing step in the ad attribute creation process, such as the layout and connection of panel input, asynchronous processing, and media API output. The middle canvas renders the flow of attribute data, using lines or arrows to represent the transfer and changes of ad attributes between different stages, allowing users to intuitively see the data flow. The top canvas is used for dynamic highlighting and annotation, displaying information according to updates to the visualization rendering engine.

[0051] When the visualization rendering engine receives an update event, it immediately applies a primary visual effect to the corresponding topological node graphic in the upper-layer canvas based on the root cause node information in the state data model to highlight the root cause. This primary visual effect can use a striking color, such as red, or a special shape or flashing effect to make the root cause node stand out in the interface and attract the user's attention. For example, setting the border of the graphic containing the root cause node to red and flashing allows the user to quickly locate the source of the problem.

[0052] Meanwhile, based on the transmission path information in the state data model, the visualization rendering engine renders dynamically animated path lines along the data flow trajectory of the intermediate layer in the upper-layer canvas to visually demonstrate the direction of fault transmission. Dynamic animation effects can include line flow, color changes, etc., allowing users to clearly see how a fault is transmitted from one node to another. For example, blue lines represent normal data flow; when a fault transmission occurs, the lines turn red and flow along the direction of the data flow, intuitively demonstrating the fault propagation path.

[0053] Furthermore, interactive responsiveness is added to the highlighted root cause nodes and affected nodes along the transmission path. This means users can interact with these nodes through clicks. When a user clicks a root cause node or an affected node, the system drills down to view detailed diagnostic information and related context. Detailed diagnostic information may include the node's attribute values ​​at the current moment, historical attribute changes, and relationships with other nodes; related context information may involve the overall situation of the advertising campaign, relevant operation records within the specific time frame of attribute inconsistencies, etc. For example, after a user clicks a root cause node, a pop-up window displays the node's detailed attributes and historical data, helping the user gain a deeper understanding of the root cause of the problem.

[0054] Finally, interactive controls that interact with the visualization view are provided. These include a timeline playback control for switching display focus, allowing users to drag the timeline to view the status of ad attributes and the propagation of faults at different points in time, providing a more comprehensive understanding of the problem's development. For example, users can drag the timeline to the point where attribute inconsistencies occurred to view the system status and fault propagation path at that time. Additionally, a bidirectional alarm list panel that interacts with canvas elements is provided. The alarm list panel displays all alarm events in the current system. When a user clicks on an alarm event, the visualization interface automatically locates the corresponding root cause node and propagation path; conversely, when a user clicks on a node or path in the visualization interface, the alarm list panel highlights the relevant alarm event, achieving bidirectional interaction and facilitating comprehensive diagnostic analysis.

[0055] S105 inputs the alarm event into the alarm strategy model trained by reinforcement learning, and dynamically decides the issuance, priority and notification channel of the alarm based on historical alarm response data and current event context.

[0056] In this embodiment, a reinforcement learning training environment simulating an advertising attribute consistency monitoring scenario is constructed. This environment needs to be able to simulate attribute changes, abnormal fluctuations, and alarm triggering conditions at each stage of the advertising creation process. Simultaneously, historical alarm response data is collected as training samples, including but not limited to: alarm event type (e.g., missing attribute value, value conflict, format error, etc.), alarm trigger time, advertising attribute node associated with the alarm, operator response actions to the alarm (e.g., ignoring, immediate processing, delayed processing, etc.), response time, processing results (e.g., whether the problem is resolved, whether it triggers secondary problems, etc.), and current event context information (e.g., advertising placement stage, placement channel, advertising type, etc.). The collected data is cleaned and preprocessed to remove noisy data, standardize the data format, and divide the data into training, validation, and test sets for subsequent model training and evaluation.

[0057] In the reinforcement learning framework, the state space is defined as the set of relevant features of the current alarm event, including contextual information such as alarm event type, trigger time, associated ad attribute node information, current ad delivery stage, delivery channel, and ad type, as well as historical response statistics for similar alarms (such as average response time and processing success rate). The action space is defined as the possible actions for the alarm event, including issuing an alarm and setting its priority (such as high, medium, and low), selecting a notification channel (such as email, SMS, system pop-up, mobile push, etc.), delaying the issuance of the alarm (setting a delay time), or ignoring the alarm. The reward function is designed based on the operator's response to the alarm and the business objectives. For example, if the operator handles the alarm promptly and successfully resolves the problem, a positive reward is given; if the alarm is ignored or the processing is delayed, leading to the escalation of the problem, a negative reward is given. Simultaneously, the impact of the notification channel selection on the operator's response efficiency is considered; for example, choosing the operator's currently active notification channel may yield additional rewards.

[0058] The preprocessed training set data is input into the reinforcement learning model, and a suitable reinforcement learning algorithm (such as Deep Q-Network (DQN), policy gradient algorithm, etc.) is used for model training. During training, the model selects actions based on the current state and adjusts its action selection strategy based on reward signals from the environment to maximize long-term cumulative rewards. Through continuous iterative training, the model gradually learns the ability to select the optimal action in different states, that is, to dynamically decide the issuance, priority, and notification channel of alarms based on historical alarm response data and the current event context. During training, a validation set is used to perform intermediate evaluation of the model, and the model hyperparameters or training strategies are adjusted based on the evaluation results to prevent overfitting. After training is completed, a test set is used to perform a final evaluation of the model to ensure its generalization ability on unseen data.

[0059] The trained alert strategy model is deployed to the advertising attribute consistency monitoring system and integrated with the alert event generation module. When the monitoring system detects an actual inconsistency in attributes or receives a risk warning, it generates an alert event and inputs the relevant feature information (i.e., status information) of the event into the alert strategy model. Based on the input status information, the model calculates the optimal action through forward propagation, determining whether to issue an alert, the alert priority, and the notification channel. The monitoring system executes corresponding operations based on the model's decision results, such as issuing an alert notification and marking the alert priority in the visualization interface.

[0060] After model deployment, continuously collect new alarm response data, including operator feedback on model decisions (such as whether they approve of the model's decisions, actual processing effects, etc.). Regularly use newly collected data to fine-tune or retrain the model to adapt to changes in advertising operations and evolving operator behavior patterns. For example, if a particular notification channel shows significantly improved response efficiency in a specific scenario, the reward function can be updated or the model training data adjusted to make the model more inclined to select that notification channel in that scenario. Through continuous optimization and updates, ensure the alarm strategy model maintains high decision accuracy and adaptability, providing efficient and accurate alarm support for consistent monitoring of advertising attributes.

[0061] S106 utilizes an intelligent alarm aggregation algorithm to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event.

[0062] In this embodiment, when the system detects inconsistencies in advertising attributes or triggers a risk warning, it first performs structured parsing on each alarm event, extracting key information fields and storing them in a standardized format. Specifically, it extracts the root cause node identifier (such as the abnormal node ID located in the attribute lineage graph), root cause node type (such as data source anomaly, processing logic error, interface transmission failure, etc.), affected attribute set (all related attributes on the current anomaly propagation path), alarm trigger timestamp, and associated advertising delivery task identifier (such as advertising plan ID or delivery batch number) from the alarm event. Through standardization, it ensures the comparability of alarm events from different sources, providing a unified data foundation for subsequent similarity analysis.

[0063] For standardized alarm events, a multi-dimensional similarity assessment system is constructed. First, the similarity of root cause node types is calculated: if two alarm events have identical root cause node types, the highest similarity score is directly assigned; if the types are different but belong to the same logical category, a partial similarity score is assigned based on a preset classification hierarchy. Second, the similarity of affected attribute sets is calculated: the overlap ratio of affected attributes in the two alarm events is calculated using the Jaccard similarity coefficient; the higher the overlap ratio, the higher the similarity. Finally, the proximity of the alarm trigger timestamps (e.g., time difference within a preset threshold) is used as an auxiliary judgment condition. The scores from these three dimensions are combined and weighted (weights can be adjusted according to actual business needs) to obtain the total similarity score for the two alarm events.

[0064] Based on similarity scores, a hierarchical clustering algorithm is used to dynamically group alarm events. Initially, each alarm event is treated as an independent cluster. Then, all event pairs are traversed; if their similarity scores exceed a preset threshold, the corresponding event pairs are merged into the same cluster. This process is repeated until no further merging is possible. To adapt to the dynamic nature of advertising operations, the clustering process employs an incremental update mechanism: when a new alarm event arrives, only its similarity to the existing cluster center events (the cluster center events are the representative events with the highest average similarity within the cluster) needs to be calculated. If the similarity to a particular cluster exceeds a threshold, it is assigned to that cluster; otherwise, a new cluster is created. This mechanism avoids the computational overhead of full re-clustering, ensuring real-time performance.

[0065] A single aggregated alert event is generated for each cluster, with its core information extracted from key fields of events within the cluster. Specifically, the root cause node identifier is the node ID with the highest frequency within the cluster (if frequencies are the same, the node that triggered the event earliest is used); the root cause node type is the common type of events within the cluster (if multiple types exist, they are labeled "mixed type"); the affected attribute set is the union of the affected attributes of all events within the cluster; the alert trigger time is the timestamp of the earliest event within the cluster; and the associated advertising task identifier is the set of task identifiers involved in all events within the cluster. Furthermore, to help operators quickly understand the severity of the aggregated alert, information such as the number of original alert events within the cluster and the number of different advertising tasks involved can be statistically analyzed as additional descriptive fields for the aggregated alert.

[0066] In the visualization interface, aggregated alert events are displayed as individual cards, containing key information such as the aggregated root cause node (highlighted), the range of affected attributes, the number of advertising tasks involved, and the earliest trigger time. They are also labeled "Aggregated Alert" to distinguish them from the original alerts. Operators can click on the card to expand and view a detailed list of all original alert events within the cluster, including the individual root cause node, affected attributes, and trigger time for each event. Regarding notification channels, the notification method is dynamically adjusted based on the severity of the aggregated alert (e.g., the number of events within the cluster, the importance of the tasks involved): high-priority aggregated alerts (such as those involving core advertising tasks or numerous anomalies) are simultaneously notified through multiple channels including SMS, email, and system pop-ups; low-priority aggregated alerts are only notified via in-system messages. This mechanism avoids information overload while ensuring that no critical issues are overlooked.

[0067] This embodiment achieves intelligent aggregation of massive alarm events in advertising attribute consistency monitoring, significantly reducing the number of alarms that operators need to handle. At the same time, it improves the efficiency of problem localization by aggregating contextual information, providing strong support for the stable operation of advertising.

[0068] In some embodiments, step S101 above, which involves collecting snapshots and metadata of advertising attributes at each processing stage of the creation process to construct an advertising attribute lineage graph for recording the entire data change history, specifically includes: Data collection probes are deployed at multiple processing stages of the ad creation process to simultaneously capture complete snapshots of ad attributes flowing through different processing stages, as well as process metadata including business identifiers, stage identifiers, and request chain context. Encapsulate the complete snapshot of ad attributes and process metadata into standardized events and send them asynchronously to the message middleware; By consuming events in the message middleware through the stream processing service, and based on the business identifier and request chain context in the event, a corresponding attribute version node is created for the current stage in the graph database; Based on the request chain context, the parent version node generated by the upstream link is found in the graph database, and a derivative relationship edge is created from the current attribute version node to the parent version node, forming a graph structure to describe data flow and version evolution, thus forming an advertising attribute lineage graph.

[0069] In this embodiment, data acquisition probes are deployed at multiple processing stages of the ad creation process. These processing stages cover the entire process of ad attributes from initial input to final output, including but not limited to panel input, asynchronous processing, and media API output. The data acquisition probes have the ability to capture data in real time, synchronously capturing complete snapshots of ad attributes flowing through different processing stages. Simultaneously, the probes also capture process metadata including business identifiers, stage identifiers, and request chain context. The business identifier uniquely identifies a specific ad business, the stage identifier clarifies the processing stage where the data resides, and the request chain context records the flow path information of the ad attributes throughout the process; this information is crucial for subsequently constructing a lineage graph. For example, in a large-scale ad delivery system, data acquisition probes are deployed at key locations such as the panel input stage for ad creative submission, the asynchronous processing stage for ad review, and the API output stage for ad delivery to media, ensuring comprehensive acquisition of ad attribute data at different stages.

[0070] The collected snapshots of advertising attributes and process metadata are encapsulated into standardized events. The encapsulation process follows a pre-defined unified format and specifications to ensure clear event structure, complete information, and ease of parsing. After encapsulation, these standardized events are sent to a message middleware via an asynchronous communication mechanism. Asynchronous sending avoids impacting the real-time performance of data collection due to network latency or temporary blockages in processing stages, ensuring timely and stable data transmission to the message middleware for temporary storage and subsequent processing. For example, the collected snapshots of advertising attributes and metadata are encapsulated in JSON format, including key information such as advertising attribute name, attribute value, business identifier, stage identifier, and request chain context, and then asynchronously sent to the Kafka message middleware via the HTTP protocol.

[0071] The stream processing service consumes events from the message middleware in real time. It possesses the ability to efficiently process large amounts of real-time data and quickly reads standardized events from the message middleware. During event consumption, based on the business identifier and request chain context information in the event, a corresponding attribute version node is created in the graph database for the current processing stage. The graph database, with its powerful graph structure storage and query capabilities, is well-suited for constructing advertising attribute lineage graphs. Each attribute version node represents a version state of an advertising attribute at a specific processing stage, storing the advertising attribute value and related metadata information for that stage. For example, when the stream processing service reads an event about advertising A in the review stage from the Kafka message middleware, it creates a node named "Advertising A - Review Stage Attribute Version" in the Neo4j graph database based on the business identifier "Advertising A" and request chain context information in the event, storing the advertising attribute value and metadata for that stage in the node's attributes.

[0072] Based on the request chain context information, the parent version node generated by the upstream stage is located in the graph database. The request chain context records the flow path of the ad attribute throughout the entire process. By analyzing this information, the upstream stage and the corresponding attribute version node of the current stage can be accurately found. After finding the parent version node, a derived relationship edge pointing from the current attribute version node to the parent version node is created in the graph database. This derived relationship edge is used to describe the flow and version evolution relationship of the ad attribute between different processing stages, clarifying the source and inheritance path of the data. By continuously creating such derived relationship edges, a graph structure for describing data flow and version evolution is gradually formed, ultimately constructing a complete ad attribute lineage graph. For example, in the example of ad A above, the attribute version node of ad A in the submission stage is found as the parent version node based on the request chain context. Then, a derived relationship edge is created in the graph database from the node "Ad A - Review Stage Attribute Version" to the node "Ad A - Submission Stage Attribute Version". This process continues, continuously improving the structure of the lineage graph as the ad attribute flows through different stages.

[0073] This embodiment can fully realize the collection of snapshots and metadata of advertising attributes at each processing stage of the creation process, and construct an advertising attribute lineage map to record the history of data changes throughout the entire chain, providing a solid foundation for subsequent monitoring and analysis of advertising attribute consistency.

[0074] In some embodiments, in step S102 above, the step of extracting time-series data of attribute values ​​based on advertising attribute lineage graphs, modeling the normal fluctuation pattern of attributes using an anomaly detection algorithm, and issuing early risk warnings for abnormal fluctuations deviating from the normal pattern specifically includes: Based on the advertising attribute lineage graph, the graph is divided according to the combination dimension of attributes and processing links. Multi-dimensional time-series feature data reflecting the evolution of attribute content and processing behavior characteristics are extracted from the historical sequence of graph nodes. For each combination of attribute and processing step, unsupervised machine learning algorithms are used to train multi-dimensional time series feature data to establish a baseline model for each combination to describe its normal fluctuation pattern. As the advertising attribute data flows through any processing stage and the lineage graph is updated, the corresponding current temporal feature vector is calculated in real time based on the newly generated graph nodes and their context information. The current time series feature vector is input into the corresponding baseline model for inference to obtain an anomaly quantification score that characterizes the degree to which it deviates from the normal pattern; When the abnormal quantitative score exceeds the preset warning threshold, a risk warning prompt is generated.

[0075] In this embodiment, based on the constructed advertising attribute lineage graph, it is meticulously divided according to the combination dimension of attributes and processing steps. For example, in the advertising creation process, there is an attribute called "ad title," and multiple processing steps such as panel input, asynchronous processing, and media API output. This results in combinations of attributes and processing steps such as "ad title - panel input," "ad title - asynchronous processing," and "ad title - media API output." From the historical sequence of graph nodes, multi-dimensional temporal feature data that comprehensively reflects the evolution of attribute content and processing behavior characteristics is extracted for each combination. This feature data covers multiple aspects, such as the frequency and magnitude of attribute value changes at different time points, the type of operation (such as modification, deletion, addition, etc.) of the attribute value in the processing steps, and the number of operations. Taking the "ad title - asynchronous processing" combination as an example, feature data such as the number of weekly modifications and the average magnitude of each modification of the ad title in the asynchronous processing step under this combination are extracted over a period of time, forming a complete multi-dimensional temporal feature dataset, providing a rich data foundation for subsequent model training.

[0076] For each combination of attribute and processing step, unsupervised machine learning algorithms are used to train on multi-dimensional time-series feature data. Unsupervised machine learning algorithms can automatically discover patterns and structures in data without pre-labeled data. By learning from a large amount of historical time-series feature data, the algorithm can identify the patterns and characteristics of normal fluctuations in the attribute under this combination. For example, for the combination of "advertising delivery channel - media API output", the algorithm may find that under normal circumstances, this attribute will have a certain frequency and range of change within a specific time period each day. Based on these learned patterns, a baseline model is established for each combination of attribute and processing step to describe its normal fluctuation pattern. This baseline model is like a standard template that can clearly define the boundaries and characteristics of normal fluctuations in the attribute under this combination, providing an accurate reference for subsequent anomaly detection.

[0077] As ad attribute data flows through any processing stage and updates the ad attribute lineage graph, the system calculates the corresponding current temporal feature vector in real time based on the newly generated graph nodes and their context information. The context information includes the attribute's previous state in the current processing stage, values ​​of other related attributes, and operational parameters of the processing stage. For example, when the ad description attribute is modified in an asynchronous processing stage and the lineage graph is updated, the system obtains information such as the ad description content before modification, the specific type of modification operation (e.g., text modification, format adjustment), and the modification time. Combining this information with the newly generated graph node data, and according to a defined feature dimension, the system calculates a temporal feature vector in real time that reflects the current attribute state and processing behavior. This vector accurately captures the dynamic changes of the attribute at the current moment, providing real-time data support for subsequent anomaly detection.

[0078] The calculated current time-series feature vector is input into the corresponding baseline model for inference. The baseline model analyzes and evaluates the input current time-series feature vector based on its learned normal fluctuation patterns. By comparing the difference between the current time-series feature vector and the normal fluctuation pattern, the model outputs an anomaly quantification score to characterize the degree of deviation from the normal pattern. This score is a numerical indicator; a higher score indicates a greater deviation of the current attribute state from the normal fluctuation pattern, and a higher probability of an anomaly.

[0079] A reasonable warning threshold is pre-set. When the abnormal quantitative score exceeds this pre-set threshold, the system will automatically generate a risk warning. The setting of the warning threshold needs to comprehensively consider factors such as the actual needs of the advertising business, historical data fluctuations, and the business's tolerance for anomalies. For example, if the business has extremely high requirements for the accuracy of ad delivery time, the warning threshold for abnormal quantitative scores related to ad delivery time attributes may be set relatively low to promptly detect anomalies that may affect delivery time. After the system generates a risk warning, it will promptly transmit relevant information to relevant operators or monitoring systems so that they can take timely measures to investigate and handle the situation, avoiding ad delivery errors or inefficiencies caused by abnormal attribute fluctuations, and ensuring the normal operation of the advertising business.

[0080] This embodiment can fully realize the extraction of time-series data of attribute values ​​based on advertising attribute lineage graphs, use anomaly detection algorithms to model the normal fluctuation pattern of attributes, and issue risk warnings in advance for abnormal fluctuations that deviate from the normal pattern, effectively improving the timeliness and accuracy of advertising attribute consistency monitoring.

[0081] In some embodiments, in step S103 above, when an actual attribute inconsistency is detected or a risk warning is received, the pre-trained causal graph model is used to analyze the transmission path of attribute inconsistency and locate the root cause node by combining the current inconsistent node information and the state of the advertising attribute lineage graph. This step further includes: Based on the data flow path reflected by the advertising attribute lineage graph, a system topology graph containing logical processing units and inter-unit dependencies is abstracted. Using the system topology as the structural framework, a causal graph model in the form of a Bayesian network is constructed. The network nodes of the causal graph model are used to represent processing units or state variables with consistent attributes. By using historical failure events and their corresponding component anomaly evidence datasets, the conditional probability parameters of each node in the causal graph model are trained and learned.

[0082] In this embodiment, based on the constructed advertising attribute lineage graph, the data flow paths reflected within it are analyzed in depth. The advertising attribute lineage graph records in detail the flow of advertising attributes in each processing stage of the creation process, including information such as which stage the data flows into, what processing it undergoes, and where it flows to. By sorting and analyzing this information, a system topology diagram containing logical processing units and the dependencies between units is abstracted. For example, in the advertising creation process, there are multiple logical processing units such as the panel input stage, the asynchronous processing stage, and the media API output stage. The panel input stage is responsible for receiving the initial advertising attribute data input by the advertiser, the asynchronous processing stage further processes and optimizes the input data, and the media API output stage outputs the processed data to the media platform for advertising placement. There are clear dependencies between these processing units; for example, the asynchronous processing stage depends on the data provided by the panel input stage, and the media API output stage depends on the data processed by the asynchronous processing stage. By presenting these processing units and their dependencies graphically, a system topology diagram is formed. This diagram can clearly show the logical architecture and data flow direction between each processing unit in the advertising creation process, providing a structural foundation for the subsequent construction of a causal graph model.

[0083] Using the abstracted system topology as the structural framework, a Bayesian network-based causal graph model is constructed. During construction, the network nodes of the causal graph model are carefully designed to represent state variables of processing units or attribute consistency. Processing unit nodes correspond to various logical processing stages in the system topology, such as panel input processing units, asynchronous processing units, and media API output processing units. These nodes reflect the operation and changes of advertising attributes at different processing stages. Attribute consistency state variable nodes represent the consistency state of advertising attributes at each processing stage, such as whether the advertising title attribute is consistent in the panel input stage or in the asynchronous processing stage. By connecting the processing unit nodes and attribute consistency state variable nodes according to the dependencies in the system topology, a Bayesian network causal graph model with a directed acyclic graph structure is formed. In this model, directed edges between nodes represent causal relationships, meaning that a change in the state of one node may affect other nodes. For example, if an anomaly occurs in the panel input processing unit, it may cause inconsistencies in advertising attributes during the input stage, thus affecting the consistency state of advertising attributes in subsequent asynchronous processing units and media API output units. This Bayesian network-based causal graph model can accurately describe the causal relationships between various processing stages of the ad creation process, providing a reasonable model framework for subsequent parameter training and root cause localization.

[0084] To ensure the constructed causal graph model accurately reflects reality, it's necessary to train the conditional probability parameters of each node in the model using historical failure events and their corresponding component anomaly evidence datasets. The historical failure event dataset contains detailed information about various attribute inconsistencies that occurred during the past ad creation process, such as the time of the failure, the involved processing steps, and the affected ad attributes. The component anomaly evidence dataset records the specific behavior of each processing unit and attribute consistency state variable when a failure occurred, such as the operation log of a processing unit and records of abnormal attribute value changes. During training, this historical data is input into the causal graph model, and the conditional probability parameters of each node are adjusted based on the causal relationships reflected in the data and the actual situation. For example, if historical data shows a higher probability of inconsistency in the ad title attribute during the input stage when a data format error occurs in the panel input processing unit, the conditional probability parameters between the panel input processing unit node and the ad title attribute consistency state variable node will be adjusted accordingly during training to better reflect reality. Through continuous iterative training, the conditional probability parameters in the causal graph model gradually converge to their optimal values, enabling the model to accurately predict the state changes and causal relationships of each node under different conditions. This provides strong support for quickly and accurately analyzing the transmission path of attribute inconsistencies and locating root cause nodes when actual attribute inconsistencies are detected or risk warnings are received.

[0085] This embodiment completes the key preparatory work for constructing a pre-trained causal graph model based on advertising attribute lineage graphs. This pre-trained causal graph model can accurately reflect the causal relationships and attribute consistency change patterns among processing units in the advertising creation process, providing a reliable foundation for subsequent root cause localization of attribute inconsistencies. This helps improve the accuracy and efficiency of advertising attribute consistency monitoring, ensuring the precision and efficiency of advertising delivery.

[0086] In some embodiments, in step S103 above, when an actual attribute inconsistency is detected or a risk warning is received, the pre-trained causal graph model is used to analyze the transmission path of attribute inconsistency and locate the root cause node by combining the current inconsistent node information and the state of the advertising attribute lineage graph. This specifically includes: When an attribute inconsistency alarm or risk warning is received, extract evidence of the currently observed node status from the alarm and real-time monitoring data; Inject the current observational evidence into the trained causal graph model, perform probabilistic inference calculations, and obtain the posterior probability that all nodes with unclear states are identified as anomalous under the current evidence; Based on the posterior probability, the nodes are sorted and causal analysis is performed. At least one node with the highest posterior probability and no abnormal parent node is selected as the root cause node causing the current attribute inconsistency, and the propagation path from the root cause node to the alarm node is determined.

[0087] In this embodiment, when the system receives an attribute inconsistency alarm or risk warning, it immediately initiates the evidence extraction process. At this time, the system comprehensively collects currently observed node status evidence from two key data sources: alarm information and real-time monitoring data. The alarm information contains basic information about the ad attribute inconsistency, such as which stage or attribute experienced the inconsistency. Although this information may be relatively brief, it provides initial guidance for evidence extraction. Real-time monitoring data is more detailed and comprehensive, recording the real-time status information of ad attributes at each processing stage of the creation process. For example, in the panel input stage, real-time monitoring data records the input ad attribute value, input time, input device, etc.; in the asynchronous processing stage, it records the intermediate states and processing time; in the media API output stage, it records the attribute values ​​output to the media platform and the output time, etc. The system performs in-depth mining and analysis on this data to extract node status evidence related to the current attribute inconsistency or risk warning. This evidence may include abnormal operation records of a certain processing stage, abnormal changes in attribute values, abnormal fluctuations in processing time, etc., providing a solid foundation for subsequent probabilistic reasoning calculations.

[0088] The extracted current observational evidence is injected into a pre-trained causal graph model. This model is constructed based on the data flow paths and causal relationships reflected in the advertising attribute lineage graph, and has been trained on historical failure events and their corresponding component anomaly evidence datasets. It accurately reflects the causal relationships and attribute consistency changes among various processing stages in the advertising attribute creation process. After the observational evidence is injected, the model performs probabilistic inference calculations. In this calculation process, the model, based on the known observational evidence and the causal relationships and conditional probability parameters between nodes, quantifies the probability that all nodes with unclear states will be identified as anomalies under the current evidence, thereby obtaining the posterior probability of these nodes being identified as anomalies. For example, if the observational evidence shows an abnormal change in a certain attribute value in the panel input stage, the model will calculate the posterior probability that each relevant node in the asynchronous processing stage and the media API output stage will be identified as anomaly based on this evidence and the causal relationship between the panel input stage and other stages. Through this scientific probabilistic inference calculation, various possible causal relationships and influencing factors can be comprehensively considered, providing an accurate basis for subsequent root cause node selection.

[0089] Based on the calculated posterior probabilities, the system sorts and performs causal analysis on all nodes with unclear states. The sorting is based on the posterior probability of a node being identified as an anomaly; the higher the posterior probability, the greater the likelihood that the node is the root cause of the attribute inconsistency. During causal analysis, the system further examines the state of each node's parent nodes. If a node has a high posterior probability and none of its parent nodes show anomalies, then this node is very likely an independent source of anomaly, i.e., the root cause of the attribute inconsistency. To ensure accurate location, the system selects at least one node with the highest posterior probability and no anomaly parent nodes as the root cause node. After determining the root cause node, the system determines the propagation path from the root cause node to the alarm node based on the causal relationships and connection paths between nodes in the causal graph model. This propagation path clearly demonstrates how attribute inconsistencies propagate from the root cause to the node that triggers the alert. For example, if the root cause is an incorrect attribute value input in the panel input stage, the propagation path might show how this error affects the processing results in the asynchronous processing stage, leading to inconsistencies in ad attributes in the media API output stage and triggering an alert. Through this precise root cause location and propagation path determination, operators can quickly understand the cause and propagation process of the problem, enabling them to take targeted measures for repair and optimization, thus improving the efficiency and quality of ad attribute consistency maintenance.

[0090] This embodiment can quickly and accurately locate the root cause of the problem, providing a strong guarantee for the accuracy and efficiency of advertising, and helping to improve the overall operational effectiveness of advertisers and advertising platforms.

[0091] In some embodiments, step S105 above, which involves inputting the alarm event into an alarm strategy model trained based on reinforcement learning, and dynamically deciding on the issuance, priority, and notification channel of the alarm based on historical alarm response data and the current event context, specifically includes: Based on historical alarm response data, the alarm decision-making process is formalized as a Markov decision process. The state space is set as a comprehensive vector describing the characteristics of alarm events, system context and receiver state, and the action space is set as a set of combined actions including whether to suppress alarms, set priorities and select notification channels. Construct a multi-objective reward function, which is used to calculate the numerical reward or penalty obtained for each decision based on whether the alarm is subsequently confirmed as valid, the efficiency of the response, and the cost-effectiveness of the notification channel. The reinforcement learning model is trained based on the state space, action space, and multi-objective reward function to obtain the alarm policy model; When a new alarm event occurs, the corresponding current state vector is constructed in real time and input into the alarm strategy model for strategy reasoning to obtain the recommended decision action for the alarm event. Execute recommended decision actions, control the issuance of alarms, priority allocation and notification channel routing, and fully record the status, actions and time information of this decision.

[0092] In this embodiment, based on historical alarm response data, the alarm decision-making process is formalized and transformed into a Markov decision process. A Markov decision process is a mathematical model used to describe decision-making in dynamic environments; it effectively simulates the relationship between states, actions, and rewards during alarm decision-making. When setting up the state space, multiple factors are considered, defining it as a comprehensive vector describing alarm event characteristics, system context, and receiver state. Alarm event characteristics include alarm type, specific advertising attributes involved, and degree of inconsistency; system context covers the current advertising scenario and system load; receiver state involves the work status of the personnel receiving the alarm and the efficiency of historical alarm processing. This comprehensive vector comprehensively and accurately describes the environment in which the alarm decision-making takes place. The action space is set as a set of combined actions including whether to suppress the alarm, setting priority, and selecting a notification channel. Suppressing the alarm determines whether the current alarm needs to be issued immediately; setting priority categorizes alarms into different priority levels based on their urgency and importance; selecting a notification channel chooses an appropriate notification method, such as email, SMS, or in-system message, based on the receiver's characteristics and actual situation. This combination of actions allows for flexible responses to various alarm situations.

[0093] To effectively guide the training of reinforcement learning models, a multi-objective reward function is constructed. This function comprehensively considers multiple factors to calculate the numerical reward or penalty for each decision. Specifically, the accuracy of the decision is evaluated based on whether the alarm is subsequently confirmed as valid. If the alarm is confirmed as valid, the decision is correct, and a positive reward is given; conversely, if the alarm is a false alarm, a corresponding negative penalty is given. Simultaneously, response efficiency is considered; if the receiver responds quickly and resolves the issue, a higher reward is given; if the response time is too long, a penalty is given. Furthermore, the cost-effectiveness of the notification channel is considered; a low-cost and effective notification channel is rewarded, while a less cost-effective channel is penalized. This multi-objective reward function comprehensively evaluates the merits of each decision, enabling the model to learn the optimal alarm strategy.

[0094] Based on the defined state space, action space, and constructed multi-objective reward function, a reinforcement learning model is trained to obtain an alarm policy model. During training, the model continuously interacts with the environment, selecting actions based on the current state and receiving corresponding rewards or penalties according to the reward function. Through continuous learning and adjustment, the model gradually optimizes its decision-making strategy to maximize cumulative rewards. After extensive training data and iterative training, the model learns the ability to select the optimal action under different states, thereby initializing the alarm policy. This initial alarm policy is learned based on historical data and experience, providing basic guidance for subsequent alarm decisions.

[0095] When a new alarm event occurs, the system constructs a corresponding current state vector in real time. This state vector is built based on a defined state space, combined with information such as the specific circumstances of the current alarm event, system context, and receiver state. For example, if the current alarm event involves an important inconsistency in advertising attributes, and the system is under high load while the receiver is handling other urgent tasks, the constructed state vector will accurately reflect this information. The constructed current state vector is then input into the alarm strategy model for strategy inference. Based on the current state vector and previously learned strategies, the alarm strategy model calculates a recommended decision action for the alarm event. This recommended decision action includes specific operations such as whether to suppress the alarm, set priority, and select notification channels.

[0096] Based on the recommended action, the system executes corresponding operations, controlling the issuance, priority allocation, and notification channel routing of alarms. If the recommended action is to issue an alarm, the system will send the alarm information to the recipient according to the set priority and notification channel; if the recommended action is to suppress an alarm, the system will not issue the alarm. Simultaneously, the system will fully record the status, action, and time information of this decision. This recorded information is crucial for subsequent model optimization and troubleshooting. By analyzing this information, we can understand the model's decision-making performance under different conditions, identify potential problems, and make timely adjustments and optimizations. For example, if a certain type of alarm is frequently incorrectly suppressed, we can analyze the reasons and adjust the model's decision-making strategy to improve the accuracy of alarm handling.

[0097] This embodiment can adjust the alarm strategy in real time according to the actual situation, improve the efficiency and accuracy of alarm processing, and provide strong support for the monitoring of advertising attribute consistency.

[0098] In some embodiments, step S106 above, which involves using an intelligent alarm aggregation algorithm to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event, specifically includes: For each input alarm event, an alarm feature vector containing root cause node identifier, affected attribute path graph structure, alarm semantics and spatiotemporal information is extracted, and the alarm feature vector is represented as a small directed graph centered on the root cause node. Based on a small directed graph, incremental clustering is performed on multiple alarm events within a dynamic sliding time window. By calculating the composite similarity between new alarm events and existing clusters, it is determined whether to merge new alarm events or create new clusters. The composite similarity includes the exact matching degree of root cause node identifiers, the similarity of affected attribute path graphs, the semantic similarity of alarm texts, and the temporal proximity. When a cluster meets the preset output triggering conditions, an aggregated alarm event is generated based on all alarm events within that cluster.

[0099] In this embodiment, for each input alarm event, the system performs comprehensive feature extraction. First, it extracts the root cause node identifier, which is crucial information for determining the core cause of the alarm event and clearly identifies the source of the problem. Next, it extracts the affected attribute path graph structure, which clearly shows the attribute nodes affected by the inconsistency in advertising attributes and their propagation paths, helping to understand the scope of the problem's impact. Simultaneously, it extracts alarm semantic information, which includes the specific meaning and description conveyed by the alarm event, enabling a semantic understanding of the alarm's essence. Furthermore, it extracts spatiotemporal information, recording the time and location of the alarm event, which is significant for analyzing the patterns and trends of problem occurrence.

[0100] The extracted information is integrated to construct an alarm feature vector containing root cause node identifiers, affected attribute path graph structure, alarm semantics, and spatiotemporal information. To more intuitively represent the relationship between the alarm feature vectors, they are represented as a small directed graph centered on the root cause node. In this small directed graph, the root cause node serves as the central node, and affected attribute nodes are connected to the root cause node through directed edges. The direction of the edges indicates the direction of propagation of attribute inconsistencies. Simultaneously, alarm semantics and spatiotemporal information are labeled as attributes of the nodes. In this way, the characteristics and structure of each alarm event can be clearly presented, providing a foundation for subsequent clustering processing.

[0101] After extracting alarm feature vectors and constructing a small directed graph, the system performs incremental clustering on multiple alarm events within a dynamic sliding time window. The dynamic sliding time window can flexibly adjust the time range according to actual needs to adapt to alarm processing requirements in different scenarios. For example, during peak advertising periods, the time window can be appropriately shortened to process alarm events more promptly; while during relatively stable business periods, the time window can be widened to reduce processing frequency.

[0102] During clustering, when a new alarm event is input, the system calculates the composite similarity between the new alarm event and existing clusters. The composite similarity is calculated based on multiple factors, including the exact match of root cause node identifiers, the similarity of affected attribute path graphs, the semantic similarity of alarm texts, and temporal proximity. The exact match of root cause node identifiers determines whether the new alarm event and existing clusters share the same root cause; the similarity of affected attribute path graphs compares the similarity of the two along inconsistent attribute propagation paths; the semantic similarity of alarm texts analyzes the similarity of alarm events from a semantic perspective; and temporal proximity considers the proximity of the occurrence times of the new alarm event and alarm events in existing clusters.

[0103] Based on the calculated composite similarity, the system determines whether to merge new alarm events into existing clusters or create new clusters. If the composite similarity is high, it indicates that the new alarm event has a high similarity to an existing cluster, and merging it into an existing cluster can further enrich the cluster's information. If the composite similarity is low, a new cluster is created to classify and manage different types of alarm events. This incremental clustering approach enables real-time dynamic clustering of alarm events, improving the accuracy and efficiency of clustering.

[0104] When an alarm cluster meets preset output triggering conditions, the system generates an aggregated alarm event based on all alarm events within that cluster. The output triggering conditions can be flexibly set according to actual business needs, such as the number of alarm events in the cluster reaching a certain threshold, or the time span of alarm events in the cluster exceeding a certain range. When the output triggering conditions are met, the system comprehensively analyzes all alarm events within the cluster, extracts key information, and generates an aggregated alarm event containing comprehensive information.

[0105] Aggregated alert events integrate key information from all alert events within the aggregation cluster, including root cause node identifiers, affected attribute path graph structures, alert semantics, and spatiotemporal information. This information is then appropriately organized and optimized, removing redundant data and highlighting key details to make the generated aggregated alert events more concise, clear, and easier to understand. For example, when describing the affected attribute path graph structure, more intuitive charts or textual explanations are used to help operators quickly understand the scope of the problem. By generating aggregated alert events, the number of redundant alerts is reduced, alert processing efficiency is improved, and operators can focus more on resolving core issues, thus enhancing the overall effectiveness of ad attribute consistency monitoring.

[0106] This embodiment can effectively integrate alarm information, improve the efficiency and accuracy of alarm processing, and provide strong support for monitoring the consistency of advertising attributes.

[0107] Reference Figure 2 An embodiment of the present invention provides an advertising attribute consistency monitoring system 2, the system 2 specifically comprising: The data recording module 201 is used to collect snapshots and metadata of advertising attributes at each processing stage of the creation process, and to construct an advertising attribute lineage map to record the history of data changes throughout the entire chain. The wind direction warning module 202 is used to extract time series data of attribute values ​​based on the advertising attribute lineage map, use anomaly detection algorithm to model the normal fluctuation pattern of attributes, and issue risk warnings in advance for abnormal fluctuations that deviate from the normal pattern. The risk attribution module 203 is used to analyze the transmission path of attribute inconsistency and locate the root cause node when an actual attribute inconsistency is detected or a risk warning is received. It utilizes a pre-trained causal graph model, combined with the current inconsistency node information and the state of the advertising attribute lineage graph. The dynamic display module 204 is used to dynamically highlight the root cause node, the affected attribute nodes, and the propagation path in the visualization interface. The alarm decision module 205 is used to input alarm events into the alarm strategy model trained by reinforcement learning, and dynamically decide the issuance, priority and notification channel of alarms based on historical alarm response data and current event context. The alarm aggregation module 206 is used to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event using an intelligent alarm aggregation algorithm.

[0108] It is understandable that, such as Figure 1 The content of the advertising attribute consistency monitoring method embodiments shown herein is applicable to the advertising attribute consistency monitoring system embodiments. The specific functions implemented by the advertising attribute consistency monitoring system embodiments are as follows: Figure 1 The advertising attribute consistency monitoring method shown in the example is the same, and the beneficial effects achieved are the same as those shown. Figure 1 The beneficial effects achieved by the advertising attribute consistency monitoring method embodiment shown are also the same.

[0109] 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.

[0110] 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.

[0111] 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 advertising attribute consistency monitoring method as described in any of the above methods.

[0112] 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.

[0113] 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.

[0114] 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.

[0115] This invention also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the advertising attribute consistency monitoring method as described in any of the above methods.

[0116] 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.

[0117] 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.

[0118] 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.

[0119] 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.

[0120] 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 monitoring the consistency of advertising attributes, characterized in that, The method specifically includes: Collect snapshots and metadata of ad attributes at each stage of the creation process to construct an ad attribute lineage graph for recording the entire data change history; Based on the advertising attribute lineage map, time-series data of attribute values ​​are extracted, and anomaly detection algorithms are used to model the normal fluctuation pattern of attributes, and risk warnings are issued in advance for abnormal fluctuations that deviate from the normal pattern. When an actual inconsistency in attributes is detected or a risk warning is received, a pre-trained causal graph model is used to analyze the transmission path of attribute inconsistency and locate the root cause node by combining the current inconsistency node information and the state of the advertising attribute lineage graph. The root cause node, its affected attribute nodes, and the propagation path are dynamically highlighted in the visualization interface. The alarm events are input into the alarm strategy model trained by reinforcement learning. Based on historical alarm response data and the current event context, the model dynamically decides on the issuance, priority, and notification channels of alarms. By using intelligent alarm aggregation algorithms, multiple alarm events with the same or similar root cause nodes are merged into a single aggregated alarm event.

2. The method according to claim 1, characterized in that, The collection of snapshots and metadata of advertising attributes at each processing stage of the creation process is used to construct an advertising attribute lineage graph to record the entire data change history, specifically including: Data collection probes are deployed at multiple processing stages of the ad creation process to simultaneously capture complete snapshots of ad attributes flowing through different processing stages, as well as process metadata including business identifiers, stage identifiers, and request chain context. Encapsulate the complete snapshot of ad attributes and process metadata into standardized events and send them asynchronously to the message middleware; By consuming events in the message middleware through the stream processing service, and based on the business identifier and request chain context in the event, a corresponding attribute version node is created for the current stage in the graph database; Based on the request chain context, the parent version node generated by the upstream link is found in the graph database, and a derivative relationship edge is created from the current attribute version node to the parent version node, forming a graph structure to describe data flow and version evolution, thus forming an advertising attribute lineage graph.

3. The method according to claim 1, characterized in that, The time-series data of attribute values ​​extracted based on advertising attribute lineage graphs are used to model the normal fluctuation patterns of attributes using anomaly detection algorithms. Abnormal fluctuations deviating from the normal patterns are then promptly alerted to potential risks. Specifically, this includes: Based on the advertising attribute lineage graph, the graph is divided according to the combination dimension of attributes and processing links. Multi-dimensional time-series feature data reflecting the evolution of attribute content and processing behavior characteristics are extracted from the historical sequence of graph nodes. For each combination of attribute and processing step, unsupervised machine learning algorithms are used to train multi-dimensional time series feature data to establish a baseline model for each combination to describe its normal fluctuation pattern. As the advertising attribute data flows through any processing stage and the lineage graph is updated, the corresponding current temporal feature vector is calculated in real time based on the newly generated graph nodes and their context information. The current time series feature vector is input into the corresponding baseline model for inference to obtain an anomaly quantification score that characterizes the degree to which it deviates from the normal pattern; When the abnormal quantitative score exceeds the preset warning threshold, a risk warning prompt is generated.

4. The method according to claim 1, characterized in that, When an attribute inconsistency is detected or a risk warning is received, a pre-trained causal graph model is used, combined with the current inconsistent node information and the state of the advertising attribute lineage graph, to analyze the transmission path of attribute inconsistency and locate the root cause node. This process also includes: Based on the data flow path reflected by the advertising attribute lineage graph, a system topology graph containing logical processing units and inter-unit dependencies is abstracted. Using the system topology as the structural framework, a causal graph model in the form of a Bayesian network is constructed. The network nodes of the causal graph model are used to represent processing units or state variables with consistent attributes. By using historical failure events and their corresponding component anomaly evidence datasets, the conditional probability parameters of each node in the causal graph model are trained and learned.

5. The method according to claim 4, characterized in that, When an attribute inconsistency is detected or a risk warning is received, a pre-trained causal graph model is used, combined with the current inconsistent node information and the state of the advertising attribute lineage graph, to analyze the transmission path of attribute inconsistency and locate the root cause node. Specifically, this includes: When an attribute inconsistency alarm or risk warning is received, extract evidence of the currently observed node status from the alarm and real-time monitoring data; Inject the current observational evidence into the trained causal graph model, perform probabilistic inference calculations, and obtain the posterior probability that all nodes with unclear states are identified as anomalous under the current evidence; Based on the posterior probability, the nodes are sorted and causal analysis is performed. At least one node with the highest posterior probability and no abnormal parent node is selected as the root cause node causing the current attribute inconsistency, and the propagation path from the root cause node to the alarm node is determined.

6. The method according to claim 1, characterized in that, The process of inputting alarm events into an alarm strategy model trained using reinforcement learning, and dynamically deciding on the issuance, priority, and notification channel of alarms based on historical alarm response data and the current event context, specifically includes: Based on historical alarm response data, the alarm decision-making process is formalized as a Markov decision process. The state space is set as a comprehensive vector describing the characteristics of alarm events, system context, and receiver status. The action space is set as a set of combined actions including whether to suppress alarms, set priorities, and select notification channels. Construct a multi-objective reward function, which is used to calculate the numerical reward or penalty obtained for each decision based on whether the alarm is subsequently confirmed as valid, the response efficiency, and the cost-effectiveness of the notification channel. The reinforcement learning model is trained based on the state space, action space, and multi-objective reward function to obtain the alarm policy model; When a new alarm event occurs, the corresponding current state vector is constructed in real time and input into the alarm strategy model for strategy reasoning to obtain the recommended decision action for the alarm event. Execute recommended decision actions, control the issuance of alarms, priority allocation and notification channel routing, and fully record the status, actions and time information of this decision.

7. The method according to any one of claims 1 to 6, characterized in that, The method of using an intelligent alarm aggregation algorithm to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event specifically includes: For each input alarm event, an alarm feature vector containing root cause node identifier, affected attribute path graph structure, alarm semantics and spatiotemporal information is extracted, and the alarm feature vector is represented as a small directed graph centered on the root cause node. Based on a small directed graph, incremental clustering is performed on multiple alarm events within a dynamic sliding time window. By calculating the composite similarity between new alarm events and existing clusters, it is determined whether to merge new alarm events or create new clusters. The composite similarity includes the exact matching degree of root cause node identifiers, the similarity of affected attribute path graphs, the semantic similarity of alarm texts, and the temporal proximity. When a cluster meets the preset output triggering conditions, an aggregated alarm event is generated based on all alarm events within that cluster.

8. An advertising attribute consistency monitoring system, characterized in that, The system specifically includes: The data recording module is used to collect snapshots and metadata of advertising attributes at each processing stage of the creation process, and to build an advertising attribute lineage map to record the history of data changes throughout the entire chain. The wind direction warning module is used to extract time-series data of attribute values ​​based on the lineage graph of advertising attributes, and to use anomaly detection algorithms to model the normal fluctuation pattern of attributes and issue risk warnings in advance for abnormal fluctuations that deviate from the normal pattern. The risk attribution module is used to analyze the transmission path of attribute inconsistency and locate the root cause node when an actual attribute inconsistency is detected or a risk warning is received. It utilizes a pre-trained causal graph model, combined with the current inconsistency node information and the state of the advertising attribute lineage graph. The dynamic display module is used to dynamically highlight the root cause node, the affected attribute nodes, and the propagation path in the visualization interface. The alarm decision module is used to input alarm events into the alarm strategy model trained by reinforcement learning, and dynamically decide the issuance, priority and notification channel of alarms based on historical alarm response data and current event context. The alarm aggregation module is used to merge multiple alarm events with the same or similar root cause nodes into a single aggregated alarm event using an intelligent alarm aggregation algorithm.

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 advertising attribute consistency monitoring method 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 advertising attribute consistency monitoring method as described in any one of claims 1 to 7.

Images