A signal-driven AI advertisement effect analysis cloud system and method based on Meta Marketing API

Abstract

A signal-driven AI advertising effect analysis cloud system and method based on Meta Marketing API belong to the field of big data advertising data analysis. The system consists of a five-layer hardware layered architecture composed of an API gateway server, a signal calculation server, an AI computing server, a prescription cache server and a database cluster. The method collects advertising data regularly and verifies the delivery threshold; constructs a dynamic cycle benchmark in three delivery time lengths, filters and selects abnormal signals through three-layer progressive filtering and multi-signal resonance; identifies audience overlapping delivery intention through two rounds of AI diagnosis and outputs a confidence problem list; retrieves historical cache or generates a standardized five-section prescription at the prescription layer and outputs it in stages; automatically opens a 72-hour retest window after user rectification, and iterates the prescription library with Beta distribution. This system can distinguish between human operation and natural traffic fluctuations, reduce abnormal misjudgment and server computing power consumption, and achieve full-link automation of diagnosis, optimization and retest.

Description

Technical Field

[0001] This invention belongs to the field of big data advertising data analysis technology, specifically involving a signal-driven AI advertising performance analysis cloud system and analysis method based on the MetaMarketing API. Background Technology

[0002] The existing Meta (Facebook) official advertising backend and third-party overseas advertising tools on the market only have basic API data retrieval and simple data visualization functions, and have several inherent technical defects.

[0003] Patent document CN117829914B discloses a digital media advertising effectiveness evaluation system, which achieves comprehensive evaluation of advertising effectiveness through a causal relationship establishment module, a counterfactual analysis module, a user profile construction module, and an anomaly detection module. This patent only optimizes advertising placement based on user behavior data and does not design a complete technical solution for operational indicators such as advertising consumption and conversion, including layered effectiveness verification, multi-signal resonance joint anomaly judgment, periodic dynamic benchmark calibration, and automatic retesting after optimization. Therefore, it lacks the core layered linkage processing logic of this invention.

[0004] Based on existing technologies, the following six major technical defects exist: (1) Lack of layered data validity verification and ad change snapshot verification mechanism. The existing system does not distinguish between ad campaign cycles (cold start period, stable period, long-term campaign period). When the exposure and spending amount do not reach the minimum threshold, it still forcibly calculates all indicators, generating a large number of invalid calculations and wasting cloud server CPU and disk computing resources; at the same time, it does not collect ad configuration change snapshots, making it impossible to distinguish whether data fluctuations are caused by natural traffic fluctuations or by human modification of materials, audiences, budgets, placements, and bids, and unable to distinguish high-quality statistical samples. (2) Only using single indicator fixed threshold alarms, lacking three-layer anomaly filtering and multi-signal resonance joint verification logic. Short-term traffic fluctuations and occasional daily consumption deviations will frequently trigger false alarms, increasing the cost of manual investigation. (3) Using a unified static judgment benchmark across the entire platform, it is impossible to dynamically calibrate the benchmark interval according to the ad campaign duration, and it is impossible to distinguish weekly periodic campaign fluctuations, resulting in distorted judgment of indicator anomalies. (iv) It only lists single abnormal data, cannot conduct causal attribution across indicators, cannot automatically distinguish whether audience overlap is due to A / B testing strategy, tiered funnel placement, or ineffective repeated placement, and cannot automatically identify placement risk intent. (v) It only outputs minimal text prompts, without standardized, tiered, and step-by-step optimization prescriptions, and lacks supporting explanations such as abnormal reasons, data evidence, step-by-step operations, expected optimization effects, and risk warnings, resulting in low guidance value. (vi) There is no automatic retesting loop after manually adjusting advertising parameters, requiring maintenance personnel to continuously and manually track the rectification effect, resulting in low automation of the entire process; and it lacks supporting modules such as material lifecycle early warning, global budget intelligent allocation, and event timeline linkage analysis, making it impossible to achieve integrated diagnosis and optimization of the entire advertising placement chain.

[0005] In summary, existing tools can only perform basic data statistics functions and cannot complete the full automated closed-loop process from API data collection, ad change snapshot recognition, hierarchical indicator calculation and filtering, AI root cause attribution, standardized prescription generation to automatic retesting of rectification. Summary of the Invention

[0006] 3.1 Technical problem to be solved by the present invention To address the technical shortcomings of existing overseas Meta advertising analysis tools, such as wasted computing power, high false alarm rates, static benchmark judgments, inability to identify advertising configuration change interference, lack of cross-metric causal attribution, lack of standardized optimization solutions, lack of automatic rectification and retesting closed loop, and lack of supporting global analysis modules, this paper provides a signal-driven AI advertising performance analysis cloud system and analysis method. This system achieves integrated, end-to-end automated processing, including advertising change snapshot recognition, layered data verification, dynamic benchmark calculation, multi-signal resonance anomaly recognition, AI-driven ad placement intent attribution, layered prescription output, and automatic retesting verification. This reduces server computing power consumption, minimizes manual intervention, and improves the accuracy of ad placement diagnosis. 3.2.1 Five-layer hardware cloud system architecture (corresponding to Figure 1)

[0007] A signal-driven AI advertising effectiveness analysis cloud system is deployed on a cloud server cluster, with a front-end PC web client for interaction. The hardware includes an API gateway server, signal computing server, AI computing power server, prescription caching server, and a distributed persistent database cluster. The system is divided into five independent functional modules, with unidirectional data flow between modules and isolated, uncoupled responsibilities. Data Acquisition Layer: Deployed on the API Gateway Server, it periodically calls the Meta Marketing API to collect raw ad delivery data in batches, synchronously captures snapshots of ad configuration changes, identifies five types of change operations: creatives, audiences, budgets, placements, and bids, limits the rate of collected data and stores it in batches into the database, and is only responsible for data retrieval and storage, without carrying out business calculation logic.

[0008] Signal Calculation Layer: Deployed on the signal calculation server, it is a pure numerical calculation module without AI models; it performs hierarchical data validity verification, advertising change window period environment verification, multi-period dynamic benchmark calculation, full index calculation, three-layer anomaly filtering, and multi-signal resonance grouping; it also marks sample weights according to window period changes, holidays, and disturbances in the same account's advertising group. All logic is implemented through deterministic code, and the calculation results are stable and reproducible.

[0009] AI Diagnostic Layer: Deployed on an independent AI computing server, it only undertakes pattern recognition and causal reasoning tasks and does not process raw numerical calculations; it performs two rounds of independent AI calls on the filtered abnormal signals. The first round identifies the root cause of the surface abnormality, and the second round extracts the overlapping signals of the audience separately and calls a dedicated AI to complete the classification of the advertising intent, and outputs a list of questions with confidence.

[0010] AI Prescription Layer: Deployed on the prescription cache server, it has a built-in historical prescription cache library and a unified five-segment prescription template; it receives the problem list output by the AI ​​diagnosis layer, first searches the cache to match historical optimization solutions, and if no matching cache is found, it calls the template to generate a standardized optimized prescription. It independently calls AI to ensure the quality of prescription output and distinguishes the output content according to three levels: severe, warning, and attention.

[0011] Front-end presentation layer: The interactive carrier is a PC web client, which visually displays the health status of advertisements, abnormal resonance signals, and graded optimization prescriptions. It supports users to mark advertisement rectification actions and view automatic retest results.

[0012] Distributed database cluster provides unified persistent storage for: original delivery data, ad configuration change snapshots, periodic indicator snapshots, multi-signal resonance anomaly records, historical optimization prescriptions, and retest verification records. 3.2.2 Signal-Driven AI Advertising Effectiveness Analysis Method (corresponding to Figure 2 for the complete process)

[0013] A signal-driven AI-based advertising effectiveness analysis method includes the following complete execution steps: S1. A scheduled task triggers the cloud API gateway server to call the Meta Marketing API to collect raw data of ad delivery in batches, and simultaneously captures snapshots of ad configuration changes. It identifies five types of change behaviors: changing creatives, adjusting audiences, modifying budgets, adjusting ad placements, and adjusting bids. The raw data and change snapshots are then stored in a persistent database.

[0014] S2. The signal calculation server reads the original data and change snapshots, and performs dual filtering for data validity and sample quality: ① Threshold verification for tiered placement: Divide the placement into three tiers based on the number of days of placement: cold start, stable, and long-term placement. Set minimum thresholds for exposure and cumulative consumption for each tier. Samples that do not meet the thresholds are directly filtered and marked as data accumulation, and do not participate in subsequent indicator calculations; ② Environmental quality verification during the window period: Detect and analyze whether there are other changes in ad configuration, statutory holidays, or synchronous adjustments of parallel ad groups within the same account during the window period. If all verification conditions are free from external disturbances, the sample is marked as a high-quality sample and assigned a sample weight of 1.0; if external interference exists, the sample weight is reduced to weaken its impact on the benchmark calculation.

[0015] S3. For the selected high-quality samples, calculate the multi-period dynamic benchmark and indicator trend for different tiers. The benchmark calculation strategy distinguishes three tiers of campaign duration: 1) Campaign duration < 30 days: Use the current average of the same target ad group as the comparison benchmark; 2) 30 days ≤ Campaign duration < 90 days: Use the weighted moving average of the historical data index of the ad group itself, with the data of the most recent 7 days having a higher weight than the data of the distant past; 3) Campaign duration ≥ 90 days: Split the weekly cycle benchmark, store the independent historical averages from Monday to Sunday, and use the average of the corresponding weekday on the day of ad campaign as the benchmark to eliminate the periodic fluctuation deviation of weekdays and weekends.

[0016] S4. Based on the calculated indicators and dynamic benchmarks, implement a three-layer progressive anomaly filtering rule: First layer: Basic data threshold filtering, eliminating invalid calculation results due to insufficient sample size; Second layer: Continuous filtering, cost and efficiency indicators must deviate from the benchmark threshold for three consecutive analysis periods to be considered potential anomalies; budget utilization indicators must be abnormal for five consecutive periods; configuration-related structural errors are directly marked upon discovery without continuous observation; trend signals require a linear regression fit R² > 0.6 to proceed to the next layer of verification; Third layer: Multi-signal resonance verification, classifying potential anomalies that have passed the first two layers of filtering into resonance groups, each group containing one core signal + at least one supporting signal; the existence of a core signal alone does not trigger an anomaly, it must be accompanied by at least one supporting signal to generate a valid anomaly signal.

[0017] S5. The system determines whether a valid abnormal signal has been generated. If no valid abnormal signal is detected, the advertisement is deemed healthy and the process ends directly. If a valid abnormal signal exists, step S6 is executed.

[0018] S6. Perform multi-signal resonance grouping for all valid abnormal signals, and classify them into three levels of problem levels based on the number of supporting signals and the degree of influence of indicators: severe level, early warning level, and attention level; display them in order of priority score, and force configuration-related structural errors to be at the top.

[0019] The S7 AI computing server initiates the first round of AI diagnostics, analyzes the surface-level fault causes corresponding to each group of resonance signals, and outputs a preliminary problem list; it extracts audience overlap signals separately, calls an independent AI model to perform audience overlap intent recognition, distinguishes four types of delivery intents: repeated A / B testing, layered funnel overlay, meaningless repeated audiences, and targeted audience conflict, and overlays them onto the problem list.

[0020] S8. Perform the second round of AI attribution analysis, combine the ad change snapshot and audience overlap intent results to complete causal reasoning, distinguish between symptoms and root causes, label the confidence level for each question, and output a complete list of questions with confidence level labels.

[0021] S9. The prescription cache server receives a list of issues with confidence levels and searches the local historical prescription cache database item by item to determine if a matching historical optimization solution exists: ① If a cache record is found, the prescription template skeleton is directly reused, and the current advertising data evidence is filled in to quickly generate an optimized prescription; ② If the cache is not found, the standard five-segment prescription template is called, and AI is independently invoked to generate a completely new optimized prescription item by item. The five-segment prescription always includes: ① A clear and easy-to-understand explanation of the issue; ② 2-3 pieces of abnormal data evidence (current value + benchmark comparison value); ③ Precise operation steps sorted by priority; ④ Expected improvement indicators and time to effect; ⑤ Up to 2 common precautions against misoperation.

[0022] S10. Graded output of diagnostic results: Severe and warning levels output complete five-segment prescriptions; attention level outputs only concise abnormality prompts; all data is pushed to a PC web client for visualization.

[0023] S11. After the user completes the advertising parameter rectification marking on the front end, the system automatically opens a 72-hour monitoring window; after the window ends, the entire process of S1~S10 is automatically re-executed to complete the retest and verification of the rectification effect.

[0024] S12. Based on the advertising conversion and consumption data after retesting, a Beta distribution weak learning model is used to calculate the success rate of each prescription optimization, dynamically adjust the prescription retrieval priority in the cache library, and complete the system's self-iterative optimization.

[0025] The addition of an ad change snapshot recognition and sample weight calibration mechanism distinguishes between human-induced disturbances and natural traffic fluctuations, reducing the false alarm rate of abnormal indicators by more than 70% and significantly reducing the workload of manual invalid investigation; the hierarchical data filtering mechanism filters out 40% of invalid indicator calculations, significantly reducing the CPU and disk computing power consumption of cloud servers.

[0026] The system uses a two-dimensional dynamic benchmark calculation with three duration levels plus a weekly cycle to eliminate the judgment bias caused by cold start of advertising, long-term placement, and weekday / weekend cycles, effectively solving the problem of distorted indicator judgment.

[0027] The three-layer anomaly filtering and multi-signal resonance joint verification mechanism avoids invalid alarms caused by occasional daily fluctuations, and greatly improves the accuracy of anomaly identification.

[0028] Two rounds of layered AI diagnostics, combined with independent audience overlap intent recognition, automatically distinguish four types of overlapping placement risks, reducing the time spent on manually identifying ad placement problems by 85%.

[0029] The standardized five-stage graded prescription output provides actionable solutions for abnormalities of varying severity, significantly enhancing its guiding value.

[0030] 72-hour automatic retesting closed loop + Beta distributed weak learning iteration, fully automated advertising rectification and verification process, improved verification efficiency by 90%, and the solution continuously self-optimizes based on the data from the campaign.

[0031] The five-layer hardware architecture achieves full-link isolation of data collection, computing, AI diagnosis, prescription generation, and front-end interaction. Computing resources are allocated on demand, and supporting analysis modules such as material lifecycle, budget allocation, and event timeline are integrated, eliminating the need to switch between multiple tools and achieving a higher degree of integration. Attached Figure Description

[0032] Figure 1 is a schematic diagram of the five-layer hardware architecture of the signal-driven AI advertising effect analysis cloud system of the present invention; Figure 2 is a complete automated execution flowchart of the signal-driven AI advertising effect analysis method of the present invention. Detailed Implementation

[0033] This embodiment is used to fully illustrate the technical solution of the present invention. The hardware deployment environment consists of a distributed cluster composed of multiple cloud CPU servers, equipped with a distributed persistent disk database. The front-end interaction carrier is an ordinary office PC, which accesses the system through a web page.

[0034] Step 1: The data collection layer starts a scheduled collection task at 2:00 AM every day. It uses the Meta Marketing API to pull raw data on impressions, clicks, spend, and audience targeting for all ad campaigns, ad groups, and creatives under the ad account in batches, with a limit of 50 ads per batch. Simultaneously, it captures snapshots of ad configuration changes, recording all operations such as creative replacement, audience adjustment, budget modification, placement change, and bid modification. The raw data and change snapshots are stored together in the raw data table of the database.

[0035] Step 2: The signal calculation server reads the collected data and change snapshots for the day and performs dual verification: First, the threshold verification: For cold start ads (1-3 days), the minimum impressions are 500 and the minimum spend is $10; for stable ads (3-7 days), the minimum impressions are 2000 and the minimum spend is $30; for long-term ads (over 7 days), the minimum impressions are 5000 and the minimum spend is $80. Data that does not meet the threshold is filtered out and not included in the metric calculation. Second, the window period environment verification: This checks for other ad changes, statutory holidays, or synchronized adjustments by other ad groups within the same account during the analysis period. If there are no external disturbances, high-quality samples are marked with a weight of 1.0; if external interference exists, the sample weight is reduced to weaken its interference with the benchmark calculation.

[0036] Step 3: For valid high-quality samples, match the corresponding dynamic benchmark calculation rules according to the advertising duration. For short-term ads, use the same target mean; for medium-term ads, use a weighted moving average; and for long-term ads, split into independent weekday benchmarks from Monday to Sunday. Simultaneously calculate the month-on-month and year-on-year trend deviation of the indicators and the linear regression R² fitting coefficient.

[0037] Step 4: Perform three layers of anomaly filtering in sequence. First, remove data with insufficient samples. Then, filter temporary fluctuations through continuous periodic thresholds. Finally, filter real and valid anomalies based on the resonance rule of core signal + supporting signal, and filter invalid alarms caused by accidental fluctuations.

[0038] Step 5: If there is no valid resonance anomaly signal, the front end will only display the health tips in the advertisement; if there are multiple valid anomaly signals, they will be classified into three levels: severe, warning, and attention, according to the number of supporting signals and the degree of business impact, and the configuration-type structural error will be forcibly pinned to the top.

[0039] Step 6: The AI ​​computing server executes two rounds of AI inference. The first round identifies the root causes of surface-level anomalies, extracts overlapping audience data separately, and calls dedicated AI to complete the classification of ad delivery intent. The second round combines change snapshots and audience intent to complete causal attribution and labels each anomaly with a confidence value.

[0040] Step 7: The prescription layer searches the local cache library one by one to match historical optimized prescriptions with the same resonance type and confidence level; if the cache is hit, the template is reused to fill the real-time data; if no matching cache is found, the standard five-segment template is called to generate a new prescription; the severe and warning levels display the complete operation plan, while the attention level only provides a simplified prompt.

[0041] Step 8: The diagnostic grading results and optimized prescriptions are pushed to the PC webpage for visual display; after the user completes the adjustment and marking of the advertising parameters, the system automatically starts a 72-hour monitoring window; after the window ends, the complete collection, calculation and diagnosis process is automatically restarted to verify the rectification effect.

[0042] Step 9: Based on the advertising conversion and consumption optimization data after retesting, use the Beta distribution model to calculate the success rate of each prescription optimization, dynamically adjust the priority of prescription retrieval from the cache library, and continuously iterate and optimize the output scheme.

[0043] In this embodiment, all data collection, indicator calculation, AI inference, prescription generation, and retesting verification processes are completed using cloud-based hardware servers, eliminating purely subjective human judgment. Through a five-layered hardware architecture, ad change snapshot verification, multi-period dynamic benchmarks, three-layer resonance filtering, two rounds of AI attribution, and an automatic retesting closed-loop integrated technical solution, this embodiment completely solves various shortcomings of existing technologies and can be stably deployed and applied to overseas Meta advertising SaaS delivery analysis systems.

Claims

1. A signal-driven AI-based cloud system and analysis method for advertising effectiveness analysis based on the Meta Marketing API, characterized in that, This process is executed in conjunction with a cloud server cluster and a PC web client, and includes the following steps: S1. The API gateway server periodically calls the Meta Marketing API to collect raw ad delivery data, synchronously captures ad configuration change snapshots, identifies five types of change behaviors: changing creatives, adjusting audiences, modifying budgets, adjusting ad placements, and adjusting bids, and stores them uniformly in the database. S2. The signal calculation server reads data and change snapshots, and performs dual filtering verification: tiered deployment threshold verification: Deployment is divided into three tiers based on the number of days: cold start, stable, and long-term. Minimum thresholds for exposure and cumulative consumption are set for each tier. Samples that do not meet the thresholds are directly filtered. Window period environment verification: Additional changes, holidays, and disturbances from parallel ad groups within the same account are detected during the window period. Samples without external disturbances are marked as high-quality and assigned a weight of 1.0; samples with disturbances are weighted lower. S3. Calculate dynamic benchmarks for effective samples based on three tiers of campaign duration: campaign duration < 30 days, take the average of the same optimized target ad group; 30 days ≤ campaign duration < 90 days, use the weighted moving average of its own historical index; campaign duration ≥ 90 days, use independent weekday benchmarks from Monday to Sunday to eliminate cyclical fluctuations. S4. Perform a three-layer progressive anomaly filtering: The first layer filters data with insufficient sample size; The second layer filters potential anomalies through continuous multi-cycle thresholds; the third layer uses multi-signal resonance verification to generate valid anomaly signals only when a core signal and at least one supporting signal are present simultaneously. S5. If there is no valid abnormal signal, the process ends; if there is a valid abnormal signal, the abnormal signal is resonantly grouped and divided into three levels: severe, warning, and attention. The S6 AI computing server performs two rounds of AI diagnosis: the first round identifies the root cause of surface-level faults, extracts overlapping audience signals separately, and calls dedicated AI to identify four types of advertising intentions; the second round combines change snapshots and audience intentions to complete causal attribution and outputs a list of issues with confidence. S7. The prescription cache server retrieves the historical prescription cache library: if the cache is hit, the template is reused to fill the real-time data and generate a prescription. If the cache is not hit, the five-segment prescription template AI is used to generate a new prescription; for severe and warning levels, a complete five-segment prescription is output, and for attention level, a concise prompt is output. S8. Diagnostic and prescription data are pushed to the PC web client for display; After a user marks an ad rectification action, a 72-hour monitoring window will automatically open. Upon expiration, S1-S7 will be executed again to complete the retest. S9. Based on the retest data, the Beta distribution model is used to calculate the success rate of each prescription optimization and dynamically adjust the prescription retrieval priority in the cache library.

2. The signal-driven AI advertising effectiveness analysis method according to claim 1, characterized in that, In step S4, cost and efficiency indicators must deviate from the benchmark for three consecutive analysis periods to be considered potential anomalies, budget utilization indicators must be abnormal for five consecutive analysis periods, trend signals must have a linear regression fit R² greater than 0.6 to enter the resonance verification layer, and configuration errors are marked as anomalies as soon as they are discovered.

3. The signal-driven AI advertising effectiveness analysis method according to claim 1, characterized in that, The five-stage prescription includes, in sequence: problem description, data evidence, priority operation steps, expected improvement indicators and time to effect, and precautions for misoperation.

4. The signal-driven AI advertising effectiveness analysis method according to claim 1, characterized in that, Audience overlap targeting intent can be categorized into four types: repeated targeting in A / B testing, layered funnel targeting, meaningless repeated audience targeting, and conflicting target audiences.

5. A signal-driven AI advertising effectiveness analysis cloud system, used to implement the analysis method according to any one of claims 1 to 4, characterized in that, This includes PC web clients, cloud server clusters, and distributed database clusters; The cloud server cluster comprises an API gateway server, a signal calculation server, an AI computing power server, and a prescription cache server, deployed in five layers of functional modules: A data acquisition layer, deployed on the API gateway server, is used to periodically collect advertising data and configuration change snapshots, and store them with rate limiting; A signal calculation layer, deployed on the signal calculation server, is a pure numerical calculation module used to perform tiered threshold verification, window period sample weight marking, multi-cycle dynamic benchmark calculation, three-layer anomaly filtering, and multi-signal resonance grouping; An AI diagnostic layer, deployed on the AI ​​computing power server, is used for two rounds of AI diagnosis, independent identification of overlapping audience intents, and outputting a list of questions with confidence levels; An AI prescription layer, deployed on the prescription cache server, has a built-in historical prescription cache library and five-segment prescription templates for matching cached or AI-generated graded and optimized prescriptions; A front-end display layer, carried by a PC web client, is used to visually display the health status of advertisements, abnormal signals, and graded prescriptions, receive user rectification marks, and display retest results; The distributed database cluster is used to persistently store original advertising data, change snapshots, indicator snapshots, resonance anomaly records, historical prescriptions, and retest records.

6. The signal-driven AI advertising effectiveness analysis cloud system according to claim 5, characterized in that, The signal calculation layer does not carry an AI model; all computational logic is implemented by deterministic code. The AI ​​diagnostic layer only performs pattern recognition and causal reasoning, and does not process raw indicator numerical calculations.

7. The signal-driven AI advertising effectiveness analysis cloud system according to claim 5, characterized in that, The system includes built-in modules for material lifecycle early warning, intelligent global budget allocation, and event timeline linkage analysis.

Images