Ad serving policy self-iterative closed loop system and method based on CID attribution data

Abstract

The application discloses an advertisement putting strategy self-iteration closed loop system and method based on CID attribution data. An attribution data standardization module receives original data of each platform, completes data preprocessing and core feature extraction through a preset format specification, and outputs a standardized data set. A user group feature library receives the standardized data set, classifies user groups, and constructs output user group feature label data. A putting strategy self-iteration engine internally embeds a reinforcement learning model, receives the data set and user group feature label data, and outputs a preliminary advertisement putting strategy adjustment instruction. A cross-platform collaborative scheduling module receives the preliminary adjustment instruction and channel conversion efficiency data of each platform, completes budget allocation and resource position scheduling, outputs a cross-platform advertisement putting strategy, and issues the cross-platform advertisement putting strategy to each media putting platform. The application belongs to the technical field of advertisements, and realizes intelligentization, automation, cross-platformization and personalized optimization of advertisement putting strategies.

Description

Technical Field

[0001] This invention relates to the field of online advertising technology. Background Technology

[0002] In the current digital advertising landscape, user behavior attribution data based on CID (ClickID) is a key basis for optimizing performance.

[0003] However, existing technical solutions have significant drawbacks: First, the attribution data formats and standards of various media platforms are different, making it difficult to directly integrate and analyze cross-platform data, thus forming "data silos".

[0004] Secondly, existing optimization methods are mostly manual or semi-automatic, relying on operators to periodically analyze data and manually adjust strategies, resulting in a serious lag in response and an inability to meet the needs of the real-time bidding (RTB) environment.

[0005] Furthermore, even when using automated tools, they are often limited to adjusting rules within a single platform or simply allocating budgets, lacking the intelligent collaborative decision-making capabilities that take a holistic view and comprehensively consider the competitive relationships between multiple platforms, user behavior paths across platforms, and real-time fluctuations in performance.

[0006] Finally, the lack of in-depth analysis of the user relationship networks and dynamic community structures hidden in CID customer identification data leads to rigid user profiles, making it difficult to achieve truly real-time personalized and precise outreach.

[0007] Therefore, there is an urgent need to develop a self-iterative closed-loop system and method for advertising delivery strategies based on CID attribution data to solve the problems in existing technologies. Summary of the Invention

[0008] The purpose of this invention is to provide a self-iterative closed-loop system and method for advertising placement strategies based on CID attribution data, which can realize intelligent, automated, cross-platform, and personalized optimization of advertising placement strategies. It is simple in structure and easy to use, thereby solving the problems of lag, inconsistent standards, and limited accuracy and conversion efficiency mentioned in the background technology.

[0009] To achieve the above objectives, the present invention provides the following technical solution: The advertising delivery strategy self-iteration closed-loop system based on CID attribution data includes an attribution data standardization module, a user group feature library, a delivery strategy self-iteration engine, a cross-platform collaborative scheduling module, and a closed-loop monitoring module. The attribution data standardization module is used to receive the original CID attribution data from various media platforms, perform data preprocessing and core feature extraction, construct a CID association event graph to extract graph features, and output a standardized CID attribution core feature dataset. The user group feature library receives the standardized dataset, and based on the event graph structure, uses a dynamic community discovery algorithm to classify user groups in real time and construct feature labels, and outputs user group feature label data. The ad delivery strategy self-iterative engine has a built-in reinforcement learning model and receives the standardized dataset and user group feature tag data. The state space of the reinforcement learning model is defined based on the joint feature definition of CID feature time-series vector and community tags with time window sliding, which is used to evaluate the value of actions and optimize the strategy. The engine outputs preliminary ad delivery strategy adjustment instructions including ad bid, audience targeting, and creative delivery weight. The cross-platform collaborative scheduling module receives the preliminary adjustment instruction and conversion efficiency data of each platform channel, performs budget allocation and resource allocation based on a multi-agent game theory model, outputs cross-platform advertising placement strategy and distributes it to each media placement platform. The closed-loop monitoring module monitors the end-to-end latency and policy execution consistency from policy issuance to the return of new CID attribution data in real time. When a single closed-loop cycle exceeds a preset threshold or the policy execution deviation exceeds a preset tolerance, the system is triggered to perform a self-check and switch to stable attitude quantity mode.

[0010] By adopting the above technical solutions, a fully automated closed-loop system was constructed, encompassing multi-platform CID attribution data access, dynamic user segmentation, intelligent strategy generation, cross-platform game scheduling, and end-to-end monitoring and assurance. Through modular design and data flow integration, it achieves unattended operation and continuous self-iteration of advertising strategies from analysis, decision-making, execution to feedback optimization.

[0011] As a further aspect of the present invention: the preprocessing operations of the attribution data standardization module include data cleaning, deduplication, and outlier removal; the core feature extraction includes extracting CID-related delivery channels, conversion cycle duration, precise user group tags, channel conversion efficiency, and creative click-through rate; the graph features include node degree centrality, average conversion path length, and community module degree.

[0012] By adopting the above technical solutions, and by constructing an event graph of users, clicks, and conversions and extracting its network topology features, the traditional attribution analysis based on single-point statistics is upgraded to a quantitative insight into the user conversion path structure and group relationships, laying a more solid structured data foundation for subsequent accurate segmentation and strategy optimization.

[0013] As a further aspect of the present invention: the feature tag dimensions constructed by the user group feature library include platform attributes, user profiles, conversion cycle features, and conversion efficiency levels; the dynamic community discovery algorithm adopts a variant of the Louvain algorithm that introduces a time decay factor, so that the community division results reflect the recent dynamic evolution of user conversion behavior.

[0014] By adopting the above technical solution and introducing an algorithm variant with a time decay factor, it is ensured that the segmentation of user groups can closely follow the evolution of their recent behavioral patterns, thereby providing dynamic user profiles that are timely and highly targeted, supporting real-time and accurate adjustments to strategies.

[0015] As a further aspect of the present invention: the reinforcement learning model adopts a hybrid architecture of DQN deep Q network combined with policy gradient method; the DQN deep Q network is used to fit a hierarchical action value function conditioned on community labels; the policy gradient method is used to optimize a unary policy network, which receives the value distribution output by DQN and outputs specific bid adjustment coefficients, directional scaling ranges and material combination weights.

[0016] By adopting the above technical solutions and through the synergy of hierarchical value assessment and meta-strategy optimization, the system can quickly identify high-value groups and fine-tune resource parameters for those groups, thereby improving the accuracy and efficiency of strategy optimization.

[0017] As a further aspect of the present invention: the reward function of the reinforcement learning model is designed as a weighted sum of two objectives as follows: in, This represents the rate of return on investment at time t. Represents the conversion rate at time t. and These are adaptive weight parameters that are dynamically adjusted based on the business stage.

[0018] By adopting the above technical solutions, we not only pursue an absolute increase in return on investment (ROI), but also incentivize the continuous growth of conversion rate (CVR). Furthermore, we can dynamically adjust and optimize the focus through weight parameters, making model decisions more business-flexible and goal-oriented.

[0019] As a further aspect of the present invention: the online inference process of the self-iterative engine of the deployment strategy adopts an integrated inference mechanism, and simultaneously loads the latest version model and the previous stable version model for parallel inference; when the difference between the output instructions of the two models exceeds a preset safety threshold, the output instruction of the previous stable version model is adopted, and the current state is marked as a state to be reinforced for training.

[0020] By adopting the above technical solution and comparing the parallel inference of the old and new models with the safety threshold, risky instructions that may be generated due to model iteration or abnormal state are effectively intercepted. While pursuing aggressive optimization of the strategy, the stability and reliability of the online system decision-making are ensured.

[0021] As a further aspect of the present invention: the multi-agent game theory model in the cross-platform collaborative scheduling module models each media platform as an agent, whose payoff function... ( )for: in, B represents the percentage of the budget allocated to platform i, and B represents the total budget. This represents the historical average conversion rate of platform i. The historical average order value of platform i As a time discount factor, The typical conversion delay days for platform i; the constraint is... and , The budget limit for platform i is defined; the model uses a distributed iterative algorithm to solve for the Nash equilibrium point as the benchmark for budget allocation, and reserves an elastic budget pool for rapid budget replenishment.

[0022] By adopting the above technical solutions, constructing a game model with the expected conversion value of time discount as the return, and solving the Nash equilibrium, the global optimal allocation of the budget across multiple platforms is achieved; the introduction of the elastic budget pool further enhances the system's ability to cope with real-time market fluctuations.

[0023] As a further aspect of the present invention: the system is configured to simultaneously process CID data and scheduling tasks from at least three heterogeneous media platforms, identify cross-conversion paths of users across different media platforms, and perform collaborative attribution and joint strategy optimization based on these paths.

[0024] By adopting the above technical solutions and identifying the cross-conversion paths of users across different platforms, the system can overcome the limitations of single-platform attribution, perform collaborative attribution and joint strategy optimization from the perspective of omni-channel marketing, and maximize overall benefits.

[0025] As a further aspect of the present invention: the system includes a visual configuration management platform for setting business stage goals, adjusting model parameters, viewing monitoring reports, and performing manual policy intervention in stable quantitative mode.

[0026] By adopting the above technical solutions and using a visual configuration management platform, operators are provided with interfaces for setting goals, adjusting parameters, monitoring status, and making necessary manual interventions, thereby enhancing the system's operability, transparency, and human-machine collaboration efficiency.

[0027] This invention also discloses a self-iterative closed-loop method for advertising delivery strategies based on CID attribution data, applied to the aforementioned system, characterized in that the method includes the following steps: S1: After the service provider completes the CID attribution, it inputs the original CID attribution data into the attribution data standardization module and outputs the standardized CID attribution core feature dataset. S2: The user group feature library receives the standardized dataset, performs dynamic community discovery and feature label update, and outputs user group feature label data; S3: The self-iterative engine of the advertising strategy receives the standardized dataset and user group feature label data, performs inference through the reinforcement learning model, and outputs preliminary advertising strategy adjustment instructions. S4: The cross-platform collaborative scheduling module receives the preliminary adjustment instruction and the conversion efficiency data of each platform channel, executes budget allocation and resource scheduling based on multi-agent game, and generates and distributes cross-platform advertising placement strategies; S5: Each media platform executes the cross-platform advertising strategy, generating new user behavior data; S6: The closed-loop monitoring module monitors the entire link to ensure that the single closed-loop cycle is less than the preset threshold T; if a timeout occurs or the strategy execution is abnormal, a degradation procedure is initiated. S7: The service provider completes a new round of CID attribution on the new user behavior data generated in S5, and uses the obtained raw data as input to repeat steps S1 to S6, forming a self-iterative closed loop.

[0028] By adopting the above technical solution, and using standardized CID feature data and dynamic community tags as the connecting thread, the entire process from data attribution input to strategy execution feedback and then to the next round of iteration is clarified, and clear closed-loop performance indicators are set, transforming the system's automation, real-time and closed-loop characteristics into repeatable and verifiable operating procedures.

[0029] Compared with the prior art, the beneficial effects of the present invention are: This invention achieves fully automated iterative optimization of advertising strategies at the minute level by constructing a tight technical closed loop of data attribution, dynamic clustering, intelligent decision-making, game scheduling, and execution monitoring. This significantly shortens the traditional optimization cycle of days, greatly improves strategy response speed and operational efficiency, and greatly reduces the cost of manual intervention. This invention, based on dynamic community discovery and multi-agent game collaboration, can gain real-time insight into the evolution of user behavior patterns and dynamically allocate multi-platform budgets and resources from a global optimal perspective. This enables cross-platform collaborative optimization and personalized, precise targeting of differentiated user groups, thereby effectively improving the overall conversion rate and return on investment of advertising.

[0030] Other features and advantages of the present invention will be disclosed in detail in the following detailed description and accompanying drawings. Attached Figure Description

[0031] Figure 1 This is a schematic diagram of the self-iterative closed-loop system for advertising delivery strategies based on CID attribution data in an embodiment of the present invention. Detailed Implementation

[0032] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0033] In this embodiment of the invention, a self-iterative closed-loop system and method for advertising delivery strategies based on CID attribution data are described below. Figure 1 As shown, the system can be physically deployed in a distributed cloud computing environment. Each core module is built as an independent microservice, and data exchange and event-driven operations are conducted through a high-throughput message middleware such as Kafka. Overall, the system receives raw CID attribution data from various media advertising platforms such as Douyin, Tencent Ads, Baidu, and e-commerce platforms such as Taobao. After a series of processing steps, it outputs cross-platform executable advertising strategies, forming a closed loop of continuous optimization.

[0034] 1. Attribution Data Standardization Module This module serves as the system's unified data entry point and processing center. It receives raw CID attribution logs pushed from various platforms via a pre-configured API adapter. The raw data is typically in JSON format and includes fields such as CID, click timestamp, impression timestamp, conversion timestamp, conversion type, conversion value, user anonymity identifier, ad campaign ID, creative ID, and cost, but the field names and structures differ across platforms.

[0035] The module is processed internally according to the following pipeline: Data cleaning: First, format validation is performed to remove records that cannot be parsed. Next, deduplication is performed using a composite key composed of CID and click timestamp. Records with obvious anomalies, such as conversion timestamps exceeding 30 days, are isolated and logged; the conversion time is calculated as conversion time minus click time.

[0036] Field Mapping and Standardization: Based on a predefined mapping table, heterogeneous fields from various platforms are uniformly converted into internal standard fields. For example, the plan ID and campaign ID from various platforms are uniformly mapped to "campaign_id"; all currency amounts are uniformly converted into RMB (CNY) using the real-time exchange rate.

[0037] Core Feature Extraction: A set of standard features is extracted from each cleaned record to form a feature vector with the CID as the primary key. This includes: Basic features: Distribution channels, advertising plan ID, creative material ID.

[0038] Conversion characteristics: whether conversion occurred, conversion type, total conversion cycle duration, and conversion value.

[0039] Efficiency characteristics: estimated channel conversion rate (CVR), creative click-through rate (CTR), and cost per thousand impressions (CPM).

[0040] Graph Feature Construction and Extraction: Based on data from a time window, such as the most recent hour, a user-event time-series attribute graph is constructed. Nodes in the graph are categorized into anonymous user IDs, CIDs for ad impressions, and conversion behavior nodes. Edges represent relationships between nodes, such as user-triggered clicks or clicks leading to conversions; edge weights are determined by both the time decay factor and event importance. The topological features of this graph are calculated in real-time using graph computing engines such as SparkGraphFrames, including: Node degree centrality: Reflects the number of CIDs associated with a user node, used to measure user activity.

[0041] Average conversion path length: For a conversion path, calculate the average number of CIDs experienced from the first exposure to the conversion to measure the complexity of the conversion chain.

[0042] Community modularity: Perform fast clustering such as spectral clustering on the graph to calculate the modularity score of the current partition and make a preliminary assessment of the clustering of the user group.

[0043] Output: Finally, the module outputs an enhanced, standardized CID attribution core feature dataset. Each record in this dataset contains all the traditional features and graph features mentioned above, along with a unified high-precision timestamp, and is published via a message queue for downstream modules to subscribe to and consume.

[0044] 2. User group characteristic database This module is a real-time updated user profile service center, with dynamic community discovery at its core.

[0045] Data aggregation: Subscribe to standardized datasets, using anonymous user IDs as keys, and aggregate all user behavior records within a sliding time window, such as the past 24 hours, to generate user-level aggregated features, such as a list of active platforms, average conversion cycle, total conversion value, and distribution of contact material types.

[0046] Dynamic Community Partitioning: This invention combines user aggregation features with the user-event graph structure as input to a dynamic graph embedding algorithm, employing a temporal extension variant of Node2Vec. The algorithm generates low-dimensional user node vectors that preserve both structural similarity and temporal proximity. Based on this, a variant of the Louvain algorithm incorporating a time decay factor is used for community detection. The objective function of this algorithm is: in, Indicates community division, Punishing drastic changes in community structure between adjacent time windows. To balance the parameters, the algorithm runs every fixed period, such as 15 minutes, and outputs the latest community partitioning results.

[0047] Feature label generation: Generate a set of multi-dimensional labels for each community: Platform attributes: such as Douyin-dominated type, cross-platform balanced type.

[0048] User profile: Based on the aggregation characteristics of users within the community, such as high-value, rapidly converting young people.

[0049] Conversion cycle characteristics: such as rapid conversion within 24 hours.

[0050] Conversion efficiency level: Classified into S level, A level, etc. based on the community's historical average ROI and CVR.

[0051] Storage and Updates: All community tags and their associated user lists are stored in a high-performance graph database such as Neo4j or a key-value database such as Redis, and tagged with a version number. When a new version is generated, the old version is replaced with an atomic operation.

[0052] Output: Provides a real-time tag query interface, and pushes a complete tag data snapshot to the strategy engine and scheduling module.

[0053] 3. Iterative engine for delivery strategies This module is an online reinforcement learning inference service that uses the Meta-RL framework for rapid adaptation.

[0054] Model Architecture: The model employs an architecture combining Model-Agnostic Meta-Learning (MAML) and Proximal Policy Optimization (PPO). The inner policy network, FastNet, takes the current state as input, including community labels and temporal statistical features, and outputs actions. The outer meta-policy network, SlowNet, learns the initial policy parameters across different communities, enabling the inner network to quickly adapt to limited new data.

[0055] State and Action: State It's a concatenated vector, including time-series statistics such as average CVR and CPM across various channels over the past hour, as well as the IDs and efficiency level codes of currently active communities. (Action) It is a hybrid vector: the discrete part is the predefined audience targeting package selection; the continuous part is [bid adjustment coefficient, targeting expansion and contraction coefficient, material A weight, material B weight, ...].

[0056] Reward function: Employing a dual-objective adaptive reward: Weight , It can be configured through a visual platform to dynamically adjust at different business stages, such as new user acquisition, user activation, and user acquisition.

[0057] Online Inference and Security Mechanisms: Service Loads Two Models: Latest Version And the previous stable version Upon receiving the status request, parallel inference is performed using both models. Key action dimensions of the outputs are compared, such as the direction and magnitude of bid adjustments. If the difference exceeds a safety threshold (e.g., a bid adjustment magnitude difference > 20%), the decision is adopted. The output is then used to mark the state as "to be explored" and add it to the subsequent training set.

[0058] Output: The engine outputs structured JSON commands, such as: {"platform":"douyin","target_community":"S1","bid_adjust":+0.15,"audience_expand":0.1,"creative_weights":{"A":0.6,"B":0.4}}.

[0059] 4. Cross-platform collaborative scheduling module This module is a global optimizer for resource allocation, based on cooperative game theory.

[0060] Game Theory Modeling: Each media platform is modeled as an agent, such as i = Douyin, j = Xiaohongshu. Agent i's strategy is to strive for a certain percentage of the budget. Its benefits It not only considers its own conversion value, but also the synergistic benefits of collaborating with other platforms: Where B represents the total budget. As a time discount factor, This is a typical conversion delay on platform i. It is the sum of historical value extracted from the CID association graph, representing the conversion of a user from exposure on platform i to conversion on platform j. These are collaborative weights. The constraints are: =1, and Platform capacity limit.

[0061] Solution and Allocation: The Shapley Value is used to calculate the marginal contribution of each platform to the total revenue of the consortium, and this contribution is used as the basis for budget allocation. The basis for this is to allocate 85% of the total budget. 15% is allocated to a flexible budget pool.

[0062] Real-time adjustments: The module monitors the budget consumption rate and actual costs of each platform. If the cost of a platform is lower than expected and the conversion rate is good, the budget can be added from the elastic pool in "microseconds" via the platform API to capture instantaneous opportunities.

[0063] Strategy compilation and distribution: Combine the policy engine's instructions with the budget allocation results, compile them into specific operation instructions that can be recognized by the APIs of various platforms, and distribute them through a parallel thread pool.

[0064] 5. Closed-loop monitoring module This module is the system's immune system, ensuring health throughout the entire process.

[0065] End-to-end tracing: Generates a globally unique trace_id for each policy iteration, spanning all services through which the data flows, to achieve end-to-end tracing.

[0066] Performance and anomaly monitoring: Latency monitoring: Calculate the closed-loop cycle, which is the time difference from the generation of the policy instruction to the completion of the standardization processing of the corresponding new data. Set a threshold T=15 minutes, and trigger an alarm if the timeout occurs.

[0067] Consistency monitoring: Compare the issued commands with the actual policy status queried through the platform API. An alarm will be triggered if the deviation exceeds 5%.

[0068] Graph Neural Network Anomaly Detection: The graph autoencoder is used to reconstruct the real-time user-event graph, calculate the reconstruction error, and trigger an alert when the error is abnormally high.

[0069] Level 3 downgrade and self-healing: Level 1 downgrade, policy rollback: Pause new model instructions and roll back to the previous stable policy parameters.

[0070] Second-level downgrade, model switching: Switch to a fully stable baseline model, such as a rule-based model.

[0071] Level 3 downgrade, manual takeover: Notify operations and maintenance personnel to intervene through the visualization platform.

[0072] Meanwhile, the states that trigger anomalies will be automatically added to the adversarial training sample set for subsequent model updates, thereby enhancing robustness.

[0073] 6. Visual configuration management platform This is a web application for operations personnel, providing the following functionalities: Business objective configuration: Select the current business stage, such as sales surge during major promotions or routine stability maintenance. The system will automatically recommend or you can manually set the reward function weights. , .

[0074] Monitoring dashboard: Real-time display of key indicators such as closed-loop cycle, consumption and conversion of each platform, and community evolution trend.

[0075] Manual intervention console: When the system enters a stable volume mode, it provides an interface for operators to manually adjust bids, targeting, or budgets.

[0076] A system workflow example, taking a beauty brand's advertising campaigns on Douyin, Xiaohongshu, and Kuaishou as an example: 1. Initial: System cold start, using historical data to pre-train the model, budget allocated according to experience ratio.

[0077] 2. First closed loop, 0-15 minutes: Attribution data shows that Douyin's conversion rate for female users born after 1995 reached 80% within 24 hours, Xiaohongshu's conversion value for mothers was high but the cycle was 3 days, and Kuaishou's overall conversion rate was only 20%.

[0078] The dynamic community division has been updated, and the tag library marks the three core groups mentioned above.

[0079] Strategy engine instructions: Douyin post-95 women's community bid +20%, main material A; Xiaohongshu mothers' community bid unchanged, but material switched to evaluation type; Kuaishou suspend large-scale advertising.

[0080] After game-theoretic scheduling calculations, the budget allocation is adjusted as follows: 50% for Douyin, 30% for Xiaohongshu, and most of the 20% for Kuaishou are transferred to the elastic pool.

[0081] Monitoring confirmed that the closed loop was completed within 12 minutes, with no abnormalities.

[0082] 3. Second closed loop, 15-30 minutes: New data shows that click-through rates for Xiaohongshu review materials have surged, but conversions have not occurred, while Douyin's costs have increased by 5% due to intensified competition.

[0083] The system dynamically adjusted its strategy: Xiaohongshu slightly reduced its bid by 5% to control costs; part of the Douyin budget was shifted to less competitive nighttime hours; and the system utilized its flexible budget pool to capture a suddenly active segment of traffic on Kuaishou.

[0084] 4. Continuous Iteration: This cycle continues, allowing the system to be continuously optimized without human intervention. Data after one week shows an overall ROI increase of 35%, with the conversion rate on Douyin's target community increasing to 85% and Xiaohongshu's conversion rate increasing to 78%.

[0085] The above description is merely a preferred embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0086] This invention provides a self-iterative closed-loop system and method for advertising delivery strategies based on CID attribution data, which can realize intelligent, automated, cross-platform, and personalized optimization of advertising delivery strategies with high reliability.

[0087] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the invention can be implemented in other specific forms without departing from its spirit or essential characteristics. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0088] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A self-iterative closed-loop system for advertising delivery strategies based on CID attribution data, characterized in that: This includes an attribution data standardization module, a user group feature library, a self-iterable engine for delivery strategies, a cross-platform collaborative scheduling module, and a closed-loop monitoring module; The attribution data standardization module is used to receive the original CID attribution data from various media platforms, perform data preprocessing and core feature extraction, construct a CID association event graph to extract graph features, and output a standardized CID attribution core feature dataset. The user group feature library receives the standardized dataset, and based on the event graph structure, uses a dynamic community discovery algorithm to classify user groups in real time and construct feature labels, and outputs user group feature label data. The self-iterative engine for the ad delivery strategy has a built-in reinforcement learning model and receives the standardized dataset and user group feature tag data. The state space of the reinforcement learning model is defined based on the joint feature definition of CID feature time-series vectors and community tags with time window sliding, which is used to evaluate the value of actions and optimize the strategy. The engine outputs preliminary ad delivery strategy adjustment instructions including ad bid, audience targeting, and creative delivery weights. The cross-platform collaborative scheduling module receives the preliminary adjustment instruction and conversion efficiency data of each platform channel, performs budget allocation and resource allocation based on a multi-agent game theory model, outputs cross-platform advertising placement strategy and distributes it to each media placement platform. The closed-loop monitoring module monitors the end-to-end latency and policy execution consistency from policy issuance to the return of new CID attribution data in real time. When a single closed-loop cycle exceeds a preset threshold or the policy execution deviation exceeds a preset tolerance, the system is triggered to perform a self-check and switch to stable attitude quantity mode.

2. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 1, characterized in that, The preprocessing operations of the attribution data standardization module include data cleaning, deduplication, and outlier removal; the core feature extraction includes extracting CID-related delivery channels, conversion cycle duration, precise user group tags, channel conversion efficiency, and creative click-through rate; the graph features include node degree centrality, average conversion path length, and community module degree.

3. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 1, characterized in that, The feature tag dimensions constructed by the user group feature library include platform attributes, user profiles, conversion cycle characteristics, and conversion efficiency levels; the dynamic community discovery algorithm adopts a variant of the Louvain algorithm that introduces a time decay factor, so that the community segmentation results reflect the recent dynamic evolution of user conversion behavior.

4. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 1, characterized in that, The reinforcement learning model adopts a hybrid architecture combining a deep Q-network (DQN) and a policy gradient method. The deep Q-network (DQN) is used to fit a hierarchical action value function conditioned on community labels. The policy gradient method is used to optimize a unary policy network, which receives the value distribution output by the DQN and outputs specific bid adjustment coefficients, directional scaling ranges, and material combination weights.

5. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 4, characterized in that, The reward function of the reinforcement learning model is designed as a weighted sum of two objectives: in, This represents the rate of return on investment at time t. This represents the conversion rate at time t. and These are adaptive weight parameters that are dynamically adjusted based on the business stage.

6. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 4, characterized in that, The online inference process of the self-iterative engine of the deployment strategy adopts an integrated inference mechanism, which loads the latest version of the model and the previous stable version of the model for parallel inference. When the difference between the output instructions of the two models exceeds a preset safety threshold, the output instructions of the previous stable version of the model are adopted, and the current state is marked as a state to be reinforced for training.

7. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 1, characterized in that, The multi-agent game theory model in the cross-platform collaborative scheduling module models each media platform as an agent, and its payoff function... ( )for: in, B represents the proportion of the budget allocated to platform i, and B represents the total budget. This represents the historical average conversion rate of platform i. The historical average order value of platform i As a time discount factor, The typical conversion delay days for platform i; the constraint is... and , The budget limit for platform i is defined; the model uses a distributed iterative algorithm to solve for the Nash equilibrium point as the benchmark for budget allocation, and reserves an elastic budget pool for rapid budget replenishment.

8. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 1, characterized in that, The system is configured to simultaneously process CID data and scheduling tasks from at least three heterogeneous media platforms, identify cross-conversion paths of users across different media platforms, and perform collaborative attribution and joint strategy optimization based on these paths.

9. The self-iterative closed-loop system for advertising delivery strategies based on CID attribution data according to claim 1, characterized in that, The system includes a visual configuration management platform for setting business phase goals, adjusting model parameters, viewing monitoring reports, and performing manual policy interventions in stable quantitative mode.

10. A self-iterative closed-loop method for advertising delivery strategies based on CID attribution data, applied to the system described in any one of claims 1 to 9, characterized in that, The method includes the following steps: S1: After the service provider completes the CID attribution, it inputs the original CID attribution data into the attribution data standardization module and outputs the standardized CID attribution core feature dataset. S2: The user group feature library receives the standardized dataset, performs dynamic community discovery and feature label update, and outputs user group feature label data; S3: The self-iterative engine of the advertising strategy receives the standardized dataset and user group feature label data, performs inference through the reinforcement learning model, and outputs preliminary advertising strategy adjustment instructions. S4: The cross-platform collaborative scheduling module receives the preliminary adjustment instruction and the conversion efficiency data of each platform channel, executes budget allocation and resource scheduling based on multi-agent game, and generates and distributes cross-platform advertising placement strategies; S5: Each media platform executes the cross-platform advertising strategy, generating new user behavior data; S6: The closed-loop monitoring module monitors the entire link to ensure that the single closed-loop cycle is less than the preset threshold T; if a timeout occurs or the strategy execution is abnormal, a degradation procedure is initiated. S7: The service provider completes a new round of CID attribution on the new user behavior data generated in S5, and uses the obtained raw data as input to repeat steps S1 to S6, forming a self-iterative closed loop.

Images