A data processing method and system based on a business management platform

Abstract

This invention discloses a data processing method and system based on a business management platform, specifically relating to the field of advertising placement business management technology. It addresses the problems of difficulty in anomaly attribution and low decision-making efficiency caused by isolated dimensional data in existing systems. The method involves synchronously acquiring channel-level click volume and touchpoint-level blacklist interception volume. When channel click volume abnormally decreases, it extracts the corresponding channel-related blacklist interception logs and product-level conversion rates. Based on the interception logs, it verifies the intensity of user behavior path topology distortion and decision manifold tightness caused by the blacklist strategy. When the distortion intensity exceeds a threshold or the tightness falls below a threshold, it backtracks the advertising exposure user path of the corresponding channel product. It analyzes the temporal correlation between blacklist interception nodes and product exposure nodes in this path to generate a candidate root cause set. It detects the direction of product conversion rate fluctuations during abnormal periods and determines the root cause from the candidate set accordingly, significantly improving the efficiency and accuracy of operational decision-making.

Description

Technical Field

[0001] This invention relates to the field of advertising placement business management technology, and more specifically, to a data processing method and system based on a business management platform. Background Technology

[0002] In the advertising management system of the business management platform, operations personnel need to make decisions based on the statistical data generated by the system. Current mainstream systems support the statistical analysis of business data from multiple independent dimensions, such as conversion rate by product, click volume by channel, and interception volume by touchpoint, and output independent analytical reports for each dimension.

[0003] When business metrics show anomalies, the existing system can only provide single-dimensional statistical results because it lacks effective causal correlation analysis capabilities between data from different dimensions. For example, when the conversion rate of a certain channel drops abnormally, the system cannot automatically identify whether the anomaly is caused by the blacklist blocking of related touchpoints or by the failure of product and channel matching. This forces operations personnel to manually backtrack multi-level log data to infer the causal link, resulting in low decision-making efficiency and difficulty in verifying the attribution results. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a data processing method and system based on a business management platform to solve the problems mentioned in the background art.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A data processing method based on a business management platform includes:

[0007] S1. Obtain the click volume from the channel dimension and the blacklist blocking volume from the touchpoint dimension of the ad placement;

[0008] S2. When the click volume at the channel level drops abnormally, extract the blacklist blocking logs associated with the corresponding channel and the conversion rate at the product level.

[0009] S3. Based on the blacklist interception logs, verify the strength of the user behavior path topology distortion caused by the blacklist policy, and verify the tightness of the decision manifold between the channel target and the interception policy.

[0010] S4. When the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold, backtrack the user path of the corresponding channel product's advertising exposure.

[0011] S5. Analyze the temporal correlation between the blacklist blocking nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set.

[0012] S6. Detect the direction of fluctuation in product conversion rate during abnormal channel periods to determine the root cause from the candidate root cause set.

[0013] In a preferred embodiment, obtaining the click volume at the channel level and the blacklist block volume at the touchpoint level for ad delivery includes:

[0014] Based on the log storage module of the advertising delivery management business system, user click behavior sequence data under the channel dimension is collected synchronously to generate channel dimension click volume statistics results with the channel identifier as the key value.

[0015] Parallel parsing of blacklist policy execution logs at the touchpoint dimension, extracting log entries containing blacklist interception action type fields, and aggregating them to generate blacklist interception volume at the touchpoint dimension indexed by touchpoint identifier;

[0016] The click volume at the channel level and the blacklist block volume at the touchpoint level are aligned through a unified time window, and the two share the same ad session identifier to achieve data association.

[0017] In a preferred embodiment, when the click volume at the channel level drops abnormally, the blacklist blocking logs associated with the corresponding channel and the product-level conversion rate are extracted, including:

[0018] The system monitors click statistics for different channels. When the click volume corresponding to a channel identifier decreases beyond the preset anomaly detection rules within a continuous fixed time window, it triggers the associated data extraction operation.

[0019] Based on the intersection of the ad serving session identifier sets, blacklist blocking log entries containing ad serving session identifiers associated with the corresponding channel identifiers are filtered from the original logs corresponding to the blacklist blocking volume at the touchpoint dimension.

[0020] The product dimension conversion rate is obtained in parallel with the set of ad placement session identifiers. The product dimension conversion rate is generated by the ratio of the number of conversion events associated with the same ad placement session identifier to the number of display events.

[0021] The blacklist interception log entries and product-level conversion rates are linked through ad session identifiers.

[0022] In a preferred embodiment, based on blacklist interception logs, the strength of user behavior path topology distortion caused by the blacklist policy is verified, and the tightness of the decision manifold between the channel target and the interception policy is verified, including:

[0023] Based on the interception policy codes in the blacklist interception log entry set, count the number of ad delivery session identifiers corresponding to each type of blacklist policy.

[0024] Divide the number of ad session identifiers for each blacklist strategy type by the total number of ad session identifier subsets to obtain the path interruption ratio for each blacklist strategy type.

[0025] Construct a policy type distribution vector based on the path interruption ratio of all blacklist policy types;

[0026] Perform L2 norm operation on the strategy type distribution vector to generate path topology distortion intensity;

[0027] Based on the product dimension conversion rate values ​​corresponding to the subset of ad delivery session identifiers, a target conversion rate vector is generated by combining the channel target conversion rate threshold.

[0028] Calculate the Pearson correlation coefficient between the strategy type distribution vector and the target conversion rate vector to generate a decision manifold tightness index between the strategy type distribution and channel targets.

[0029] In a preferred embodiment, when the path topology distortion intensity exceeds a first threshold or the decision manifold density is lower than a second threshold, the user path of the corresponding channel product's advertising exposure is traced back, including:

[0030] Based on the channel product identifier associated with the path topology distortion intensity exceeding the first threshold or the decision manifold density being lower than the second threshold, extract the corresponding advertising session identifier subset from the blacklist interception log entry set;

[0031] For each ad serving session identifier in the subset of ad serving session identifiers, perform a join query operation with the user behavior log table;

[0032] Obtain the user path sequence for ad exposure through related query operations.

[0033] In a preferred embodiment, the ad exposure user path sequence includes an ad exposure event timestamp, an exposure channel identifier, and a sequence of user interaction event types.

[0034] In a preferred embodiment, the temporal correlation between the blacklist blocking nodes and product exposure nodes in the user path of ad exposure is analyzed to generate a candidate root cause set, including:

[0035] Extract the exposure channel identifier corresponding to each exposure event timestamp in the user path sequence of ad exposure;

[0036] Iterate through the blacklist of blocked log entries, and for each log entry matching the channel product identifier, extract the blocking timestamp and the blocking reason code;

[0037] Within the time window defined by the timestamps of adjacent exposure events, the number of blacklisted blocks under the same exposure channel identifier is counted, and the interception statistics results at the touchpoint dimension are generated.

[0038] When the number of interceptions at the touchpoint dimension corresponding to the exposure channel identifier exceeds the third threshold, the corresponding exposure event timestamp and the associated exposure channel identifier will be added to the candidate root cause set.

[0039] In a preferred embodiment, based on the temporal distribution characteristics of the timestamps of each exposure event in the candidate root cause set, the abnormal interception-dense segment within a set time range before the exposure event of the target channel product occurs is marked.

[0040] In a preferred embodiment, detecting the direction of fluctuation in product-level conversion rates during periods of channel anomalies to determine root causes from a set of candidate root causes includes:

[0041] If click volume decreases in tandem with channel-level clicks, then the root cause should be identified as the blacklist blocking action at the touchpoint level from the candidate root cause pool; if it increases against the trend, then the root cause should be identified as the product-channel matching failure from the candidate root cause pool.

[0042] On the other hand, the present invention provides a data processing system based on a business management platform, comprising:

[0043] The data acquisition module is used to obtain the number of clicks from the advertising channels and the number of blacklisted clicks from the touchpoints.

[0044] The anomaly response module is used to extract the blacklist blocking logs associated with the corresponding channel and the conversion rate of the product dimension when the click volume drops abnormally at the channel level.

[0045] The policy verification module is used to verify the strength of user behavior path topology distortion caused by blacklist policies based on blacklist interception logs, and to verify the tightness of the decision manifold between channel targets and interception policies.

[0046] The path backtracking module is used to backtrack the user path of the corresponding channel product when the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold.

[0047] The root cause candidate module is used to analyze the temporal correlation between the blacklist interception nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set.

[0048] The root cause identification module is used to detect the direction of fluctuation in product-level conversion rates during periods of channel anomalies in order to identify the root cause from the candidate root cause set.

[0049] Compared with the prior art, the present invention has the following beneficial effects:

[0050] The cross-dimensional data association diagnostic mechanism constructed in this invention effectively solves the pain point of ambiguous attribution of anomalies in advertising business, breaks through the limitations of traditional single-dimensional statistics, and achieves automated root cause tracing of abnormal events by dynamically coupling multi-source data such as channel click volume, touchpoint interception volume and product conversion rate (corresponding to S1, S2 and S6) and designing specific analysis steps (S3-S5). This significantly improves diagnostic accuracy and efficiency and avoids inefficient manual backtracking and guessing by operations personnel.

[0051] Specifically, by introducing quantitative verification of topological distortions in user behavior paths and the tightness of strategy decision-making manifolds, it is possible to accurately identify whether the execution of blacklist strategies deviates from the core objectives of the channel or causes abnormal distortions in the path. Furthermore, by combining in-depth analysis of the temporal correlation of key nodes in the exposure path and collaborative judgment of the direction of conversion rate fluctuations, it successfully distinguishes between two fundamentally different root causes: anomalies caused by excessive blacklist blocking and traffic degradation caused by product-channel mismatch failure. This diagnostic logic based on multi-dimensional data dynamic correlation and in-depth verification not only significantly reduces the risk of misjudgment but also directly outputs actionable attribution conclusions, providing reliable technical support for operational decisions. Attached Figure Description

[0052] Figure 1 This is a flowchart of a data processing method based on a business management platform according to the present invention;

[0053] Figure 2 This is a schematic diagram of the structure of a data processing system based on a business management platform according to the present invention. Detailed Implementation

[0054] 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.

[0055] Example 1: Figure 1 This invention provides a data processing method based on a business management platform, comprising:

[0056] S1. Obtain the click volume from the channel dimension and the blacklist blocking volume from the touchpoint dimension of the ad placement;

[0057] S2. When the click volume at the channel level drops abnormally, extract the blacklist blocking logs associated with the corresponding channel and the conversion rate at the product level.

[0058] S3. Based on the blacklist interception logs, verify the strength of the user behavior path topology distortion caused by the blacklist policy, and verify the tightness of the decision manifold between the channel target and the interception policy.

[0059] S4. When the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold, backtrack the user path of the corresponding channel product's advertising exposure.

[0060] S5. Analyze the temporal correlation between the blacklist blocking nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set.

[0061] S6. Detect the direction of fluctuation in product conversion rate during abnormal channel periods to determine the root cause from the candidate root cause set.

[0062] S1. Obtain the click volume from the advertising channel and the blacklist block volume from the touchpoint dimension. The specific implementation is as follows:

[0063] The user interaction event stream is captured in real time through the log storage function of the advertising delivery management business system. This event stream includes click behavior records triggered by the front-end user interface and blacklist interception records executed by the back-end strategy engine. The specific process of generating channel-level click volume includes: filtering raw log entries from the log storage function that are marked as "click" and whose dimension attributes contain a valid channel identifier, where the channel identifier adopts a unique encoding system consisting of a letter prefix and a number; grouping and aggregating the filtered log entries by channel identifier, with the aggregation period set to a fixed time window, and counting the number of log entries in the group corresponding to each channel identifier. This number is the channel-level click volume statistics result.

[0064] The process of generating the blacklist interception volume at the touchpoint dimension specifically includes: extracting log records from the same log storage function whose action type field value is "blacklist interception" and contains a non-empty touchpoint identifier, where the touchpoint identifier is composed of touchpoint type code and geographic location hash value; indexing the extracted records with the touchpoint identifier as the primary key, and calculating the number of times the same touchpoint identifier appears within the same fixed time window, which is the blacklist interception volume at the touchpoint dimension.

[0065] The data association mechanism includes: assigning a globally unique ad delivery session identifier to each user session during the log collection phase. This identifier is generated using an irreversible encryption algorithm based on the user device fingerprint and timestamp, where the user device fingerprint is obtained through the browser's native interface to obtain hardware feature codes; when generating channel-level click volume, the statistical results are associated with the set of ad delivery session identifiers corresponding to the log entries used during generation; when generating touchpoint-level blacklist interception volume, the corresponding set of ad delivery session identifiers is synchronously associated; in subsequent analysis processes, data association between the channel dimension and the touchpoint dimension is established by comparing the intersection of the ad delivery session identifier sets.

[0066] The length of the fixed time window is dynamically configured according to business needs, such as 5 minutes or 60 minutes. The start time of the time window is strictly synchronized with the statistical operations of all dimensions, and the time deviation of all server nodes is ensured to be less than 50 milliseconds through the global clock protocol of the distributed system. The collection process of user click behavior sequence data specifically includes: embedding a lightweight monitoring script in the user client, generating a log message containing a channel identifier and a millisecond-level timestamp when an ad click event is detected, and transmitting it to the log collection server through an encrypted channel; the collection process of blacklist policy execution logs specifically includes: when the policy engine triggers an interception action, it generates a structured log containing a touchpoint identifier and an interception reason code, and writes it directly to the distributed message queue.

[0067] The data alignment operation is performed after the statistical results are generated. The specific method is as follows: expand the channel dimension click volume statistics results into session-level click records according to the ad placement session identifier; expand the touchpoint dimension blacklist interception volume into session-level interception records according to the ad placement session identifier; sort the session-level click records and interception records with the same ad placement session identifier by timestamp, and establish an association mapping when the time difference between the two is less than the preset tolerance threshold. The tolerance threshold is dynamically adjusted according to the network latency, for example, it is set to 2000 milliseconds.

[0068] Among them, the channel identifier is a unique identifier for the advertising channel, consisting of the letters CH followed by a 6-digit number; the touchpoint identifier is a unique identifier for the user touchpoint, consisting of a 3-digit touchpoint type code followed by an 8-digit geohash value; the advertising session identifier is a unique identifier for a single user session period, with a fixed length of 32 characters; the fixed time window is an equal-length interval used for statistical analysis, with a length configuration range of 1 to 60 minutes; the irreversible encryption algorithm is, for example, the SHA-256 hash algorithm; the global clock protocol is, for example, the Network Time Protocol; and the tolerance threshold is the maximum allowed event time deviation, with a configuration range of 500-3000 milliseconds.

[0069] For example, within a certain time window, the click count for channel identifier CH202301 was 150, while the interception count for touchpoint identifier CT880 was 80. Correlation with ad delivery session identifiers revealed that 78 identifiers from the set of session identifiers corresponding to the 150 clicks also appeared in the interception session set. Time alignment verification showed that the average interval between click events and interception events was 800 milliseconds, lower than the 2000 millisecond tolerance threshold. All operations were implemented using a distributed computing framework, such as using Spark SQL to perform grouped aggregations and utilizing Redis to store the session identifier mapping relationships.

[0070] S2. When the click volume at the channel level drops abnormally, extract the blacklist blocking logs associated with the corresponding channel and the conversion rate at the product level. The specific implementation is as follows:

[0071] First, based on real-time updated channel-level click statistics, a continuous analysis is performed on the click sequence corresponding to each channel identifier. Continuous analysis refers to arranging click data into a time series according to the generation order of fixed time windows. When it is detected that the click volume corresponding to a specific channel identifier shows a continuous downward trend within three consecutive fixed time windows, and the cumulative decline exceeds the preset anomaly judgment rule, subsequent operations are triggered. The preset anomaly judgment rule sets a threshold for the decline magnitude by analyzing the distribution characteristics of historical data. For example, it is set to be an anomaly when the decline rate of two adjacent time windows exceeds 15% and the total decline rate of the third window compared to the first window exceeds 30%. This rule supports automatically matching different threshold percentages according to the channel type, and the specific matching relationship is stored in the threshold configuration table.

[0072] Based on the intersection of the established sets of ad placement session identifiers, a blacklist interception log extraction operation is performed. According to the channel-level click statistics corresponding to the abnormal channel identifier, the complete set of ad placement session identifiers associated with generating the statistics is obtained. This complete set is then intersected with the ad placement session identifiers in the session-level interception records to obtain a subset of ad placement session identifiers that exist simultaneously in both click and interception sessions. Based on this subset, complete log entries are retrieved from the original log storage system corresponding to the blacklist interception volume at the touchpoint level. The original log storage system uses a distributed storage cluster based on ad placement session identifier hash sharding to implement log indexing, where the ad placement session identifier serves as the sharding key. During retrieval, a parallel query interface is used to obtain a set of blacklist interception log entries containing all target identifiers. Each log entry includes an interception timestamp, touchpoint identifier, and interception policy code. The interception policy code is completely consistent with the policy type code defined in Example 1.

[0073] The process of obtaining product-dimensional conversion rates in parallel includes the following steps. Based on the aforementioned subset of ad session identifiers, relevant conversion event records are extracted from the conversion event repository of the user behavior analysis system. The conversion event repository continuously receives tracking data from the ad landing page via a Kafka message queue. Conversion behavior is defined as a user completing a product purchase or submitting a valid inquiry. A product purchase event corresponds to an order creation message, and a valid inquiry event corresponds to a form submission success message. Simultaneously, display event records corresponding to the same subset of ad session identifiers are extracted from the ad display log repository. These display event records include ad exposure timestamps and exposure location codes, which match the ad placement hierarchy structure defined in Example 1. Within the same fixed time window, the total number of conversion events and the total number of display events corresponding to the subset of ad session identifiers are counted. The total number of conversion events is divided by the total number of display events to obtain the product-dimensional conversion rate. This calculation process uses a distributed aggregation operation performed by the Spark engine. For example, when the total number of display events is 1000 and the total number of conversion events is 50, the product-dimensional conversion rate is calculated to be 0.05, and the result is rounded to four decimal places.

[0074] The implementation details of the data association mechanism include physical storage structure and time alignment control. The blacklist interception log entries and product-level conversion rate values ​​are both associated using the same subset of advertising session identifiers. The physical storage of this association adopts a relational database association table structure, where the advertising session identifier serves as the association primary key and is indexed using a B+ tree. Regarding time range alignment, the generation timestamps of the blacklist interception log entries and the timestamps of events involved in the product-level conversion rate statistics are required to be within a continuous fixed time window that triggered the anomaly. The boundary values ​​of the time window are synchronized to all relevant systems through the ZooKeeper distributed coordination service, with boundary time errors controlled within 100 milliseconds. Specifically, a timestamp comparison function verifies whether the log timestamps are within the range of the window start time plus 100 milliseconds and the window end time minus 100 milliseconds. The verification method for associated datasets includes checking the coverage of a subset of ad serving session identifiers in the intercept logs and conversion records. For example, when the subset contains 1,000 identifiers, it is necessary to verify that the blacklist log retrieval results contain log entries corresponding to all 1,000 identifiers and that the conversion rate calculation covers the event records associated with all 1,000 identifiers.

[0075] The key parameter configuration rules are as follows. The continuous fixed time window length inherits the configuration parameters from Example 1, with a standard setting of 30 minutes, adjustable within the range of 10 to 60 minutes. Anomaly detection uses a setting of three consecutive windows, configurable as 2 to 5 windows. The dynamic adjustment of the decline threshold is based on the channel's historical fluctuation characteristics. Specifically, it is calculated by obtaining the standard deviation σ of the click volume over the past 30 windows for that channel, setting the threshold to 3σ, and using the lowest threshold of 5% when σ is less than 5%. The intersection calculation of the subset of ad placement session identifiers is implemented using the SINTER command in the Redis cluster, with a single identifier strictly limited to a 32-character hexadecimal string. Conversion event types are explicitly defined as including two categories: order creation events and form submission events, corresponding to event codes ORDER_CREATE and FORM_SUBMIT, respectively. The time alignment tolerance mechanism uses the NTP network time protocol for synchronization, maintaining the clock deviation of each node within 50 milliseconds. Nodes that exceed the timeout period automatically trigger a clock synchronization procedure.

[0076] The verification process is illustrated by the following example. When the click statistics for channel identifier CH202305 in window sequence W1-W3 are [1500, 1200, 900], the decrease rate of window W2 compared to window W1 is calculated as (1500-1200) / 1500 = 20%, and the total decrease rate of window W3 compared to window W1 is calculated as (1500-900) / 1500 = 40%. The historical fluctuation standard deviation of this channel is σ = 12%, and the trigger threshold is 3 × 12% = 36%. The actual decrease rate of 40% exceeds the threshold, so it is judged as abnormal. A total of 1200 advertising session identifiers associated with this channel are extracted. The intersection of these identifiers with the 800 identifiers in the session-level interception records yields a subset S (cardinality 780). 780 blacklist interception logs are retrieved from the distributed storage cluster, each log containing complete interception policy codes. The total number of display events corresponding to the parallel statistical subset S is 8500, and the total number of conversion events is 425, including 400 order creation events and 25 form submission events. The product-dimensional conversion rate is calculated as 425 / 8500 = 0.0500. The timestamps of all log entries are verified to be between 100 milliseconds after the start time of W1 and 100 milliseconds before the end time of W3. The conversion event timestamps are verified using the same rules. The coverage of the 780 identifiers in subset S is 100% in both the intercept logs and the conversion event records.

[0077] S3. Based on the blacklist interception logs, verify the strength of the user behavior path topology distortion caused by the blacklist policy, and verify the tightness of the decision manifold between the channel target and the interception policy. The specific implementation is as follows:

[0078] When performing user behavior path topology distortion strength verification based on the blacklist interception log entry set, the interception policy code recorded in the log entry set is parsed first. The interception policy code consists of an alphanumeric string, where the first three digits are fixed as the policy type identifier. For example, "BLK001" represents a regional blacklist policy, and "BLK002" represents a device blacklist policy. For each ad delivery session identifier in the log entry set, the first three digits of its associated interception policy code are extracted, and the number of ad delivery session identifiers corresponding to each type of blacklist policy is counted. The ad delivery session identifier is a hexadecimal string with a fixed length of 32 characters, obtained through the session_id field of the log entry. This field is generated and written to the log by the ad delivery system during the log collection phase.

[0079] The path interruption ratio for each blacklist strategy type is obtained by dividing the number of ad session identifiers for each blacklist strategy type by the total number of ad session identifier subsets. The ad session identifier subset refers to the set of sessions blocked by the blacklist strategy, and its total number is obtained by counting the unique instances of the `session_id` field in the statistical log entries, specifically using the `COUNT(DISTINCT)` aggregation function of the distributed database. The path interruption ratio calculation result is rounded to four decimal places. For example, if a certain strategy blocks 1250 sessions and the total number of ad session identifier subsets is 5000, then the path interruption ratio is 0.2500. When the total number of ad session identifier subsets is zero, the error code "ERR_DIV_ZERO" is returned and the calculation process terminates.

[0080] Based on the path interruption ratios of all blacklisted policy types, a policy type distribution vector is constructed. This vector is a numerical vector with a dimension equal to the total number of blacklisted policy types. The vector dimension is determined by querying the policy configuration table to obtain all valid policy types. Each element in the vector is arranged in lexicographical order according to the policy type encoding, generated based on ascending ASCII code values. For example, if there are three policy types BLK001, BLK002, and BLK003 with path interruption ratios of 0.3, 0.5, and 0.2 respectively, the constructed vector would be [0.3, 0.5, 0.2]. For policy types not appearing in the logs, their path interruption ratio is recorded as 0.0000. After vector construction, the vectors are stored in a double-precision floating-point array, with the array indices strictly corresponding to the policy type sorting positions.

[0081] The L2 norm operation is performed on the strategy type distribution vector to generate the path topology distortion intensity. The L2 norm operation process consists of three sub-steps: first, the square value of each element in the vector is calculated; second, the square values ​​of all elements are summed to obtain the sum of squares; and finally, the square root of the sum of squares is calculated. The operation process follows the IEEE 754 double-precision floating-point calculation standard. For example, the L2 norm calculation process for the vector [0.3, 0.5, 0.2] is as follows: the square value is calculated to obtain [0.09, 0.25, 0.04]; the sum of squares is calculated to be 0.09 + 0.25 + 0.04 = 0.38; and the square root is approximately 0.6164. The calculation result is retained to six significant figures, and this value is the path topology distortion intensity, used to quantify the interruption concentration of user behavior paths.

[0082] Based on the product-level conversion rate values ​​corresponding to a subset of ad session identifiers, a target conversion rate vector is generated by combining the channel target conversion rate threshold. The product-level conversion rate values ​​are obtained as follows: for each identifier in the ad session identifier subset, the `conversion_status` field of the user behavior log table is queried. This field is a boolean value indicating whether the conversion was successful; the conversion rate equals the number of successfully converted sessions divided by the total number of sessions. The channel target conversion rate threshold is read from the system configuration table. The threshold setting rule is: take the 75th percentile value of the historical conversion rate data from the most recent 30 days, which is reviewed and confirmed by operations personnel before being stored in the database. The construction rules for the target conversion rate vector are: the vector dimension is consistent with the strategy type distribution vector; for each strategy type, the average conversion rate of its corresponding session is calculated; if this average value is greater than the channel target conversion rate threshold, the vector element value is 1.0; if it is less than or equal to the threshold, the vector element value is 0.0. For example, if the channel target conversion rate threshold is 0.15, and the average conversion rate for strategy type BLK001 is 0.18, then the element value corresponding to this strategy type in the target conversion rate vector is 1.0.

[0083] The Pearson correlation coefficient between the strategy type distribution vector and the target conversion rate vector is used to generate the decision manifold tightness index. The Pearson correlation coefficient calculation process includes the following steps: First, verify whether the dimensions of the two vectors are consistent; if the dimensions do not match, an anomaly of dimension mismatch is raised. Second, calculate the mean of the elements in the strategy type distribution vector by summing all element values ​​and dividing by the number of elements. Calculate the mean of the elements in the target conversion rate vector. Next, calculate the covariance, which is equal to the difference between each element of the strategy type distribution vector and its mean multiplied by the difference between the corresponding element of the target conversion rate vector and its mean, then summing all products and dividing by the number of elements minus 1. Then, calculate the standard deviation of each vector, which is equal to the square root of the sum of the squares of the differences between each vector element and its own mean, divided by the number of elements minus 1. Finally, divide the covariance by the product of the two standard deviations to obtain the correlation coefficient. This coefficient ranges from -1 to +1; for example, a result of 0.85 indicates a strong positive correlation between the strategy distribution and target achievement. The calculation result is the decision manifold tightness index, stored to four decimal places in the analysis results database.

[0084] In practice, a subset of ad session identifiers is extracted from the blacklist interception log entries using an SQL query. The query conditions are that the `log_type` field equals "intercept" and the `timestamp` field is within a specified time range. Product-level conversion rate values ​​are obtained by using a LEFT JOIN operation to link the ad session identifiers with the user behavior log table. A hash mapping is established between the `partition_key` field of the user behavior log table and the ad session identifiers. The construction of the strategy type distribution vector requires a full set of strategy type validation. When a new strategy type is added to the strategy configuration table, the vector dimensions are automatically expanded. The square root calculation in the L2 norm operation uses Newton's iteration method, and the iteration terminates when the difference between two adjacent iteration results is less than 1e-6. A null value check is performed before calculating the Pearson correlation coefficient; if any vector contains a NaN value, a data cleaning process is triggered.

[0085] All intermediate data is stored in an in-memory data structure. The strategy type distribution vector and the target conversion rate vector use a Float64 array of the same dimension. The path interruption ratio calculation module has a built-in division-by-zero protection mechanism. When the total number of ad delivery session identifier subsets is detected to be zero, a preset error code is returned and subsequent calculations are skipped. After the decision manifold tightness index is generated, the following metadata is associated and recorded: calculation timestamp, channel target conversion rate threshold version number, and log data time range identifier. The final index is output to the strategy optimization engine. The engine triggers strategy adjustment alarms based on the path topology distortion intensity value. When the decision manifold tightness index is below 0.5, a strategy re-evaluation task is generated.

[0086] Through the above implementation process, a closed-loop technical chain is achieved from raw logs to quantitative indicators. The lexicographical arrangement of the strategy type distribution vector ensures the reproducibility of the strategy order, and L2 norm operations compress the multidimensional distribution into a comparable scalar indicator. Binarization of the target conversion rate vector establishes a discretized evaluation benchmark, and the Pearson correlation coefficient calculation uses an unbiased estimator (divided by n-1) to ensure statistical validity. The entire process includes 12 anomaly handling branches, covering boundary scenarios such as missing data, dimensional mismatch, and computational overflow. The input and output data of each calculation module are timestamped and version-tagged, supporting backtracking verification of the calculation process.

[0087] For example, in an e-commerce advertising scenario, blacklist strategies include three types: geographic strategy (BLK_GEO), device strategy (BLK_DEV), and content strategy (BLK_CT). The total subset of ad delivery session identifiers is 10,000, with BLK_GEO blocking 4,500, BLK_DEV blocking 3,000, and BLK_CT blocking 2,500. The path interruption ratios are 0.4500, 0.3000, and 0.2500, respectively, constructing a strategy type distribution vector [0.45, 0.30, 0.25]. The L2 norm is approximately 0.596. The channel target conversion rate threshold is set to 0.12, and the actual conversion rates for the three strategies are 0.13, 0.11, and 0.10, respectively. Therefore, the target conversion rate vector is [1.0, 0.0, 0.0]. The covariance of the two vectors was calculated as follows: the mean of the strategy distribution vector was (0.45 + 0.30 + 0.25) / 3 = 0.333, and the mean of the target vector was (1 + 0 + 0) / 3 = 0.333; the covariance was 0.0285; the standard deviation of the strategy distribution vector was 0.098, and the standard deviation of the target vector was 0.577; the correlation coefficient was 0.0285 / (0.098 * 0.577) ≈ 0.504. This result indicates a moderate positive correlation between the strategy distribution and target achievement.

[0088] The implementation of step S3 transforms discrete blacklist policy interception behaviors into continuous and computable mathematical representations by constructing a policy type distribution vector. Compared to simple counting statistics, this more comprehensively reflects the distribution characteristics of the interruption impact of different policy types on user paths. For example, using the L2 norm instead of the L1 norm to calculate distortion intensity is more sensitive to high-proportion policy types (square operation amplifies outliers), accurately identifying concentrated interruption risks. The binarization of the target conversion rate vector (1 / 0 threshold judgment) transforms business objectives into computable discrete indicators, forming a same-dimensional spatial mapping with the policy distribution vector. By using the Pearson correlation coefficient instead of simple difference comparison, the linear correlation between the policy distribution pattern and the target achievement can be captured, avoiding misjudgments caused by differences in data volume. The entire process design covers data anomaly scenarios (such as division by zero verification, NaN cleaning), and uses standardized mathematical operations (IEEE 754 floating-point specification, Newton's iteration method) to ensure consistency of calculation results across platforms. Compared to traditional manual policy evaluation, the calculation error is controlled within 1e-6 when automatically processing tens of thousands of log entries, and the results can be backtested. Compared to alternative solutions (such as simply counting the number of interceptions or manual sampling analysis), this implementation method, while ensuring computational efficiency, achieves multi-dimensional quantitative evaluation of the strategy's effectiveness through vector space modeling, providing a data-driven decision-making basis for dynamic strategy optimization.

[0089] S4. When the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold, backtrack the user path of the corresponding channel product's advertising exposure, as follows:

[0090] Based on the determination results where the path topology distortion intensity exceeds a first threshold or the decision manifold density is lower than a second threshold, the channel product identifier associated with this determination result is obtained. This channel product identifier is generated by the anomaly detection process in the preceding steps. The specific association logic is as follows: when the path topology distortion intensity calculation process detects that the anomaly of a channel product's user path connection relationship is greater than a preset first threshold, or when the decision manifold density calculation process detects that the cluster radius of the channel product's user decision behavior in the feature space is less than a preset second threshold, then the unique code of the channel product is output as the associated identifier. Specifically, the path topology distortion intensity is obtained by calculating the KL divergence between the actual jump probabilities between user path nodes and the baseline probability distribution. The first threshold is set based on the topology distortion intensity distribution of historical normal paths, for example, taking the 95th percentile value of this distribution (0.35). The decision manifold density is obtained by calculating the average Euclidean distance from all user behavior vectors to the cluster center. The second threshold is set based on the cluster radius distribution of historical normal channels, for example, taking the 5th percentile value of this distribution (1.8).

[0091] Based on channel product identifiers, a subset of corresponding ad placement session identifiers is extracted from the blacklist interception log entry set. The blacklist interception log entry set is generated by the pre-control risk process, and each log entry contains four fields: ad placement session identifier, channel product identifier, interception timestamp, and interception reason code. The specific extraction process is as follows: traverse all records in the log entry set, filter out entries whose channel product identifier field completely matches the currently associated identifier, store the ad placement session identifier field value in a temporary storage area, and eliminate duplicate identifiers using a hash deduplication algorithm, ultimately forming the ad placement session identifier subset. The ad placement session identifier is a globally unique string, generated by performing an MD5 hash operation on the channel ID, user ID, and exposure start timestamp, for example, forming an identifier in the format "CH123_U456_7a8b9c0d".

[0092] For each ad session identifier in the subset of ad session identifiers, a join query operation is performed with the user behavior log table. The user behavior log table is stored in a distributed columnar database. This table is established during the pre-data collection step, with the ad session identifier as the primary key. Each record contains five fields: user ID, device fingerprint, exposure event timestamp sequence, exposure channel identifier sequence, and user interaction event type sequence. The join query operation is implemented through a distributed computing framework and includes the following steps: First, the subset of ad session identifiers is loaded into the memory cache; second, a parallel task is started to scan the partition data of the user behavior log table; then, a Bloom filter is used to quickly match records whose ad session identifier field values ​​match the subset elements; finally, the successfully matched records are sorted in ascending order by exposure event timestamp, and log entries spanning multiple dates are merged to form a complete path record.

[0093] The user path sequence for ad impressions is obtained through a relational query operation. This sequence is a structured data set, where each element corresponds to the complete path information of an ad delivery session identifier, specifically comprising three components:

[0094] Ad exposure event timestamp sequence: A UNIX timestamp array sorted in ascending order by time, with the time unit being milliseconds. It records the precise time of each ad exposure in this ad delivery session. For example, a sequence containing two timestamps [1633046405123, 1633046452101] represents two exposures that occurred on October 1, 2021 at 08:00:05.123 and 08:00:52.101.

[0095] Exposure channel identifier sequence: A string array with the same length as the timestamp sequence. Each element identifies the media channel to which the exposure event belongs. The channel code adopts the format of "platform type_ad slot ID", for example, ["search engine A_keyword ad slot B", "social media C_feed ad slot D"];

[0096] User interaction event type sequence: A sequence of user behavior tags that occur between adjacent exposure events. The event type uses predefined enumeration values, including four basic types: click, conversion, closing pop-up, and scrolling / staying. For example, the behavior sequence ["click", "scrolling / staying"] that occurs between two exposures indicates that the user performs a click operation after the first exposure, which is then triggered by the scrolling / staying behavior.

[0097] The user path sequence of ad exposure will be analyzed further to pinpoint the root cause of the anomalies.

[0098] The exposure channel identifier sequence in the sequence is converted into a state transition matrix, and its KL divergence value with the baseline path template is calculated. When the deviation between two consecutive state transition probabilities exceeds 0.3, it is marked as a path distortion point. The user interaction event type sequence is vectorized by converting the event type count in each time window into a four-dimensional vector (click count, conversion count, pop-up close count, scroll dwell count). Then, the cluster density of all path vectors of the same channel is calculated by the DBSCAN clustering algorithm, with the density radius parameter ε set to 1.5 and the minimum sample size set to 20. When more than 15% of the path sequences in the backtracking results have two consecutive instances of failing to trigger the expected decision event (e.g., no click or conversion event is detected within 5 seconds after exposure), a channel strategy adjustment instruction is generated and sent to the advertising delivery system.

[0099] S5. Analyze the temporal correlation between the blacklist blocking nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set. The specific implementation is as follows:

[0100] First, the exposure event timestamp sequence and exposure channel identifier sequence recorded in the ad exposure user path sequence are read. This sequence structure has been fully defined in the S4 embodiment, containing an array of UNIX millisecond-level timestamps arranged in ascending order of time and an array of channel identifier strings that strictly correspond to them. For example, for a certain path sequence, the exposure event timestamp sequence contains three elements, namely 1633046405123, 1633046452101, and 1633046489000, and the exposure channel identifier sequence contains three elements, namely search engine A_ad slot B, social media C_ad slot D, and search engine A_ad slot E. Each element of the two sequences corresponds strictly one-to-one according to its index position.

[0101] The system iterates through the blacklist of blocked log entries, traversing all log entries that match the current channel product identifier. In embodiment S4, this set's data structure is defined as containing four fields: a blocking timestamp field, a channel product identifier field, an ad session identifier field, and a blocking reason code field. For each log entry, the system extracts the numerical time data from the blocking timestamp field and the string encoding from the blocking reason code field. The blocking reason code uses a three-letter alphanumeric combination to represent the specific risk control rule type; for example, blocking reason code F12 represents blocking fraudulent traffic. It's important to note that the channel product identifier and the exposure channel identifier are different dimensional data: the former is a product-level identifier, such as product X_search engine channel, while the latter is an ad placement-level identifier, such as search engine A_ad placement B. They are associated through a channel product identifier and exposure channel identifier mapping table established in embodiment S4, which is stored in a channel mapping table in a relational database.

[0102] Interception statistics are performed within a time window defined by the timestamps of adjacent exposure events. The time window length is set to 5000 milliseconds, based on the maximum effective response time threshold between two ad exposures determined by experimental psychology research. Specifically, the starting point of the time interval in the exposure event timestamp sequence is set to the end point, which is the starting point of the time interval plus 5000 milliseconds. For the current exposure channel identifier (i.e., the channel identifier string corresponding to the starting point of the time interval), the following operations are performed: The blacklist log entry set is scanned, and entries that simultaneously meet two conditions are selected. The first condition is that the interception timestamp value of the log entry is greater than or equal to the starting point of the time interval and less than or equal to the starting point of the time interval plus 5000 milliseconds. The second condition is that the channel product identifier field of the log entry is converted to the exposure channel identifier through a channel mapping table. The total number of log entries meeting both conditions is counted, generating interception statistics results at the touchpoint level. For example, if the exposure channel is identified as search engine A_ad slot B, and 3 interception logs are counted within the time interval from the start point of the time interval equal to 1633046405123 to the start point of the time interval plus 5000 milliseconds equal to 1633046410123, then the interception count result of the touchpoint dimension is equal to 3.

[0103] When the interception statistics for the touchpoint dimension exceed the third threshold, the corresponding exposure event timestamp and associated exposure channel identifier are added to the candidate root cause set. The third threshold is set as follows: based on the interception data of the same exposure channel identifier within the same time period (e.g., 9:00-10:00 Beijing time) over the past 30 calendar days, the arithmetic mean μ and standard deviation σ of this dataset are calculated. The third threshold is then dynamically calculated using the formula μ + 3σ. For example, if the historical average interception μ for a certain ad placement is 2.1 times per 5 seconds and the standard deviation σ is 0.8, then the third threshold is 2.1 + 3 multiplied by 0.8, which equals 4.5. The addition operation is triggered when the touchpoint interception statistics result is greater than or equal to 5. The candidate root cause set data structure is a set of binary tuples. Each binary tuple element contains two attributes: the first attribute is the exposure event timestamp accurate to milliseconds, and the second attribute is the complete channel identifier string.

[0104] Based on the temporal distribution characteristics of the timestamps of each exposure event in the candidate root cause set, the abnormal interception-intensive segment within a set time range before the exposure event of the target channel product is marked. The set time range is fixed at 60,000 milliseconds before the exposure event timestamp, which is based on the minimum warm-up duration determined by the advertising delivery system's warm-up period technical document. The specific execution steps include: extracting the timestamp sequence composed of all exposure event timestamps in the candidate root cause set, sorting it in ascending order of value, and then scanning the time axis; using a sliding window length of 60,000 milliseconds and a step size of 1,000 milliseconds, the step size is set according to the minimum time granularity of system log collection, and calculating the number of candidate root cause entries contained in the interval from the start of each window to the start of the window plus 60,000 milliseconds; when the number of candidate root cause entries exceeds the fourth threshold, the time interval is marked as an abnormal interception-intensive segment. The formula for calculating the fourth threshold is: twice the arithmetic mean η of the number of candidate root cause entries of the same channel product within the 60,000 millisecond window before exposure in the past 30 days. For example, when η equals 8, the fourth threshold equals 16. The final output is a list of time intervals, each interval represented by a start timestamp and an end timestamp. For example, the start timestamp 1633046345000 to the end timestamp 1633046405000 indicates that there were abnormally dense interception events between 07:59:05 and 08:00:05 on October 1, 2021.

[0105] During implementation, the following data closed loop is established: the index correspondence between the exposure event timestamp sequence and the exposure channel identifier sequence ensures that each exposure event timestamp can accurately locate the exposure channel identifier; the channel product identifier is converted into the exposure channel identifier through a predefined channel mapping table to ensure the consistency between the blacklist log and the exposure event channel; the time window interval calculation adopts the exposure event timestamp plus a fixed number of milliseconds to avoid time zone conversion errors; during sliding window detection, the candidate root cause set is sorted by timestamp to ensure the correctness of time series analysis.

[0106] The technical parameters are set based on the following: the third threshold is calculated using the historical data μ+3σ formula, which conforms to the statistical outlier detection standard; the 60,000 millisecond warm-up period is taken from the default value of the ChannelPreheat parameter in the advertising system API documentation; and the sliding window step size of 1,000 milliseconds is consistent with the storage accuracy of the timestamp field in the log collection system.

[0107] The exception handling mechanism includes: when the timestamp sequence of exposure events contains data in reverse order, a sorting function is called to rearrange it in ascending order; when the channel mapping table lacks a corresponding relationship, the current log entry is skipped and the error code ERR_MAP_404 is recorded; when the candidate root cause set is empty, an empty exception interval list is directly returned. All numerical calculations use 64-bit floating-point arithmetic, and timestamp comparisons use long integer comparisons to avoid data type conversion errors.

[0108] The parameter adjustment range is clearly defined as follows: the time window length can be configured between 3000 milliseconds and 10000 milliseconds; the third threshold coefficient can be adjusted between 2σ and 4σ; the warm-up period duration supports a range of 30000 milliseconds to 120000 milliseconds; and the sliding window step size can be configured between 500 milliseconds and 5000 milliseconds. The adjustment is based on the actual throughput of the business system; for example, in high-concurrency scenarios, the time window can be shortened to 3000 milliseconds to reduce computational load.

[0109] S6. Detect the fluctuation direction of product conversion rate during abnormal channel periods to determine the root cause from the candidate root cause set. The specific implementation is as follows:

[0110] If click volume decreases in tandem with channel-level clicks, then the root cause should be identified as the blacklist blocking action at the touchpoint level from the candidate root cause pool; if it increases against the trend, then the root cause should be identified as the product-channel matching failure from the candidate root cause pool.

[0111] Based on the candidate root cause set and advertising conversion log data, the specific implementation process for determining the root cause type by performing product-level conversion rate fluctuation direction detection is as follows. First, the candidate root cause set generated in the S5 embodiment is obtained. This set contains multiple binary elements, each with two attributes: the first attribute is the exposure event timestamp accurate to milliseconds, and the second attribute is the complete channel identifier string. Simultaneously, the list of densely blocked anomaly segments output in the S5 embodiment is obtained. Each densely blocked anomaly segment represents a time interval from the start timestamp to the end timestamp. For example, the start timestamp 1633046345000 to the end timestamp 1633046405000 represents a specific time interval.

[0112] For each densely populated anomaly interception segment, perform the following correlation analysis. Extract conversion event records for the target channel product from the advertising conversion log database within a time range from milliseconds before to milliseconds after the densely populated anomaly interception segment. The comparison time range is a preset parameter, with a default setting of 300,000 milliseconds (5 minutes). This advertising conversion log data structure contains four fields: a conversion event timestamp field recording the UNIX millisecond-level timestamp of the event occurrence; a channel product identifier field storing the combined identifier of the product and channel; a user identifier field recording the unique user code that triggered the conversion; and a conversion type code field identifying the conversion event type. For example, for the time interval [1633046045000, 1633046705000], query all conversion event records where the channel product identifier field equals the target product's search engine channel.

[0113] Calculate key metrics within two time windows. The first window is defined as the time range of the currently analyzed anomaly interception-intensive segment, where the start timestamp of the first window equals the start timestamp of the anomaly interception-intensive segment, and the end timestamp of the first window equals the end timestamp of the anomaly interception-intensive segment. The second window is defined as the historical baseline window, which is the merged data of the same time period as the first window within the past 30 calendar days. For example, if the first window corresponds to the period from 14:00 to 14:01 Beijing time, then the historical baseline window is the collection of all data from the period from 14:00 to 14:01 in the past 30 days. Calculate the product-dimensional conversion rate according to the following logic: the current window conversion rate equals the total number of conversion events in the first window divided by the total number of exposure events in the first window; the historical baseline conversion rate equals the total number of conversion events in the historical baseline window divided by the total number of exposure events in the historical baseline window. The total number of exposure events is obtained by querying the number of records in the advertising exposure log where the channel product identifier field equals the target product identifier field. The data structure of this advertising exposure log has been defined in the S4 embodiment. For example, if the first window has a total of 2000 exposure events and 60 conversion events, then the current window conversion rate is equal to 60 divided by 2000, which equals 0.03; if the historical baseline window has a total of 60000 exposure events and 2400 conversion events, then the historical baseline conversion rate is equal to 2400 divided by 60000, which equals 0.04.

[0114] The click-through rate of change (CTR) is calculated synchronously across channels. Click event records from the same two time windows used for conversion rate calculation are extracted from the ad click log. This ad click log data structure includes a click timestamp field recording the UNIX millisecond-level timestamp of the click, an exposure channel identifier field recording the ad placement-level identifier, and a user identifier field recording the unique user code that triggered the click. The CTR is equal to the difference between the total number of click events in the first window and the total number of click events in the historical baseline window, divided by the total number of click events in the historical baseline window. For example, if the total number of click events in the first window is 400 and the average total number of click events in the historical baseline window is 500, then the CTR is equal to the difference between 400 and 500 divided by 500, which equals -0.2.

[0115] Establish root cause determination rules. When both conditions are met simultaneously, the root cause is determined to be a blacklisted entry at the touchpoint level; otherwise, it is determined to be a product-channel matching failure root cause. Condition 1 is that the click-through rate change is less than the fifth threshold, and condition 2 is that the conversion rate change (the difference between the current window conversion rate and the historical baseline conversion rate, divided by the historical baseline conversion rate) is less than the sixth threshold. The fifth threshold is set to -0.15, and the sixth threshold is set to -0.1. These thresholds are set based on the well-known significant fluctuation criteria in the field of digital advertising performance evaluation. For example, if a click-through rate change equals -0.2 (less than -0.15) and a conversion rate change equals the difference between 0.03 and 0.04 (divided by 0.04, equaling -0.25, less than -0.1), then the blacklisted entry root cause determination is triggered.

[0116] Perform root cause confirmation. If the root cause is determined to be a blacklist blocking issue, iterate through all binary elements in the candidate root cause set, filter out elements whose exposure event timestamps fall within the current densely blocked abnormal segment, and mark them as the final root cause event set. If the root cause is determined to be a product-channel matching failure, extract the exposure channel identifiers associated with the current densely blocked abnormal segment from the candidate root cause set, and combine them with the device type distribution data in the user profile database to generate a channel-device compatibility analysis report. This user profile database includes dimensional data such as device model field, operating system version field, and screen resolution field.

[0117] The implementation process establishes a closed-loop data association. The densely populated anomaly interception areas are derived from the S5 implementation example output; the total number of exposure events is associated with the target product through the channel product identifier field in the ad exposure log; the total number of click events is associated with the same identifier in the ad click log through the exposure channel identifier field; and conversion events are associated with the same identifier in the conversion log through the channel product identifier field. Time window alignment employs a UNIX timestamp value comparison mechanism, with all timestamps uniformly set to millisecond-level integer values ​​to avoid time zone conversion errors.

[0118] The parameter settings are based on the following: The comparison time range is 300,000 milliseconds, referencing the upper limit of a typical conversion cycle determined in user behavior analysis research; the fifth threshold of -0.15 represents the critical value for a significant drop in clicks, determined based on experience in analyzing performance fluctuations in the digital advertising field; the sixth threshold of -0.1 corresponds to the industry-experienced threshold for a significant drop in conversion rates. The adjustable range of parameters is clearly defined as follows: the comparison time range can be configured between 180,000 milliseconds and 600,000 milliseconds, the fifth threshold can be adjusted between -0.2 and -0.1, and the sixth threshold can be set between -0.15 and -0.05. Adjustments are based on the actual response latency characteristics of the advertising system, for example, expanding the comparison time range to 600,000 milliseconds in high-latency scenarios.

[0119] The anomaly handling mechanism includes three types of boundary scenario response strategies. When the historical baseline window data is less than 7 days, the data collection range is automatically expanded to the most recent 90 days of data. When the denominator (total number of exposure events) in the conversion rate calculation is zero, the current segment analysis is skipped and the error code ERR_DIV_ZERO is recorded. When the candidate root cause set has no temporal intersection with the current anomaly interception dense segment, the root cause unconfirmed status code WARN_NO_MATCH is output. All percentage calculations use 64-bit floating-point arithmetic and retain three decimal places; negative value comparisons use precise numerical comparisons to avoid floating-point errors.

[0120] For the final root cause event set marked as a blacklist blocking root cause, a blocking rule verification instruction containing a list of exposure event timestamps and a list of exposure channel identifiers is sent to the risk control service; for product-channel matching failure root causes, the channel matching optimization process of the advertising delivery strategy service is triggered. The output is a structured data object, including a root cause type code field, where code 1 represents a blacklist blocking root cause and code 2 represents a product-channel matching failure root cause; the associated exposure event timestamp list field stores all associated UNIX millisecond-level timestamps; the channel identifier string field records the complete channel identifier; and the fluctuation index value table field stores the specific values ​​of conversion rate change rate and click-through rate change rate.

[0121] In the data processing logic, the total number of exposure events is counted using a distributed counting framework. The advertising exposure logs are grouped by the channel product identifier field before counting. Historical baseline window data processing uses a sliding window aggregation algorithm to merge data from multiple days within the same time period. During the generation of the channel-device compatibility analysis report, the application programming interface of the user profile service is called to obtain a histogram of device type distribution. All time window boundary processing adopts the left-closed and right-closed interval principle, meaning that timestamps equal to boundary values ​​are included within the interval.

[0122] This embodiment achieves precise location of anomaly root causes by constructing a multi-source log real-time correlation analysis framework. Unlike the simple threshold alarm mechanism of conventional monitoring systems, it couples ad exposure anomaly detection (S1-S3) with conversion effect fluctuation analysis (S4-S6) across processes: First, it identifies micro-interception events based on the spatiotemporal characteristics of exposure logs (S1-S2), and then extracts abnormal segments through density clustering (S3). Simultaneously, a dual-window comparison mechanism is established at the conversion analysis layer (S4), combining click-through rate mutation indicators (S5) and product conversion rate changes (S6) for collaborative judgment, automatically distinguishing between blacklist interception and channel failure—two fundamental root causes. This dual-layer judgment architecture effectively overcomes the misjudgment problem caused by user behavior delays in the ad conversion chain, especially through dynamic baseline window settings and device profiling linkage mechanisms (S6), which solve the technical blind spots of traditional methods in cross-channel attribution.

[0123] Example 2: Figure 2 A schematic diagram of the structure of a data processing system based on a business management platform is provided according to the present invention. The data processing system based on a business management platform includes:

[0124] The data acquisition module is used to obtain the number of clicks from the advertising channels and the number of blacklisted clicks from the touchpoints.

[0125] The anomaly response module is used to extract the blacklist blocking logs associated with the corresponding channel and the conversion rate of the product dimension when the click volume drops abnormally at the channel level.

[0126] The policy verification module is used to verify the strength of user behavior path topology distortion caused by blacklist policies based on blacklist interception logs, and to verify the tightness of the decision manifold between channel targets and interception policies.

[0127] The path backtracking module is used to backtrack the user path of the corresponding channel product when the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold.

[0128] The root cause candidate module is used to analyze the temporal correlation between the blacklist interception nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set.

[0129] The root cause identification module is used to detect the direction of fluctuation in product-level conversion rates during periods of channel anomalies in order to identify the root cause from the candidate root cause set.

[0130] All calculations involved in the embodiments are dimensionless numerical calculations, and the preset parameters and thresholds in the calculations are set by those skilled in the art according to the actual situation.

[0131] It should be noted that this invention can be deployed on the device itself to realize embedded applications, or it can run on a PC or other terminal with a user interface, thereby meeting various hardware environments and usage requirements.

[0132] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, as a computer program product. The computer program product includes one or more computer instructions or computer programs. When the computer instructions or computer programs are loaded or executed on a computer, all or part of the processes or functions described in the embodiments of this application are generated. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable device. The computer instructions can be stored in a computer-readable storage medium or transmitted from one computer-readable storage medium to another. For example, the computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center via wired (e.g., infrared, wireless, microwave, etc.) means. The computer-readable storage medium can be any available medium that a computer can access or a data storage device such as a server or data center that includes one or more sets of available media. The available medium can be a magnetic medium (e.g., floppy disk, hard disk, magnetic tape), an optical medium (e.g., DVD), or a semiconductor medium. The semiconductor medium can be a solid-state drive.

[0133] Those skilled in the art will understand that, for the sake of convenience and brevity, the specific working processes of the systems, devices, and modules described above can be referred to the corresponding processes in the foregoing method embodiments, and will not be repeated here.

[0134] In the several embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of modules is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple modules or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or modules may be electrical, mechanical, or other forms.

[0135] The modules described as separate components may or may not be physically separate. The components shown as modules may or may not be physical modules; they may be located in one place or distributed across multiple network modules. Some or all of the modules can be selected to achieve the purpose of this embodiment according to actual needs.

[0136] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0137] If the aforementioned functions are implemented as software functional modules and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or a portion of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0138] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0139] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A data processing method based on a business management platform, characterized in that, include: S1. Obtain the click volume from the channel dimension and the blacklist blocking volume from the touchpoint dimension of ad placement; S2. When the click volume at the channel level drops abnormally, extract the blacklist blocking logs associated with the corresponding channel and the conversion rate at the product level. S3. Based on blacklist interception logs, verify the strength of user behavior path topology distortion caused by the blacklist policy, and verify the tightness of the decision manifold between the channel target and the interception policy, including: Based on the interception policy codes in the blacklist interception log entry set, count the number of ad delivery session identifiers corresponding to each type of blacklist policy. Divide the number of ad session identifiers for each blacklist strategy type by the total number of ad session identifier subsets to obtain the path interruption ratio for each blacklist strategy type. Construct a policy type distribution vector based on the path interruption ratio of all blacklist policy types; Perform L2 norm operation on the strategy type distribution vector to generate path topology distortion intensity; Based on the product dimension conversion rate values ​​corresponding to the subset of ad delivery session identifiers, a target conversion rate vector is generated by combining the channel target conversion rate threshold. Calculate the Pearson correlation coefficient between the strategy type distribution vector and the target conversion rate vector, and generate a decision manifold tightness index between the strategy type distribution and channel targets; S4. When the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold, backtrack the user path of the corresponding channel product's advertising exposure. S5. Analyze the temporal correlation between the blacklist blocking nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set. S6. Detect the direction of fluctuation in product conversion rate during abnormal channel periods to determine the root cause from the candidate root cause set.

2. The data processing method based on a business management platform according to claim 1, characterized in that, Obtain click-through rates from advertising channels and blacklisted block rates from touchpoints, including: Based on the log storage module of the advertising delivery management business system, user click behavior sequence data under the channel dimension is collected synchronously to generate channel dimension click volume statistics results with the channel identifier as the key value. Parallel parsing of blacklist policy execution logs at the touchpoint dimension, extracting log entries containing blacklist interception action type fields, and aggregating them to generate blacklist interception volume at the touchpoint dimension indexed by touchpoint identifier; The click volume at the channel level and the blacklist block volume at the touchpoint level are aligned through a unified time window, and the two share the same ad session identifier to achieve data association.

3. The data processing method based on a business management platform according to claim 1, characterized in that, When click volume drops abnormally at the channel level, extract the blacklist blocking logs associated with the corresponding channel and the conversion rate at the product level, including: The system monitors click statistics for different channels. When the click volume corresponding to a channel identifier decreases beyond the preset anomaly detection rules within a continuous fixed time window, it triggers the associated data extraction operation. Based on the intersection of the ad serving session identifier sets, blacklist blocking log entries containing ad serving session identifiers associated with the corresponding channel identifiers are filtered from the original logs corresponding to the blacklist blocking volume at the touchpoint dimension. Parallel acquisition of product-dimensional conversion rates corresponding to the set of ad placement session identifiers. The product-dimensional conversion rate is generated by statistically analyzing the ratio of the number of conversion events associated with the same ad placement session identifier to the number of display events. The blacklist interception log entries and product-level conversion rates are linked through ad session identifiers.

4. The data processing method based on a business management platform according to claim 1, characterized in that, When the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold, the user path of the corresponding channel product's ad exposure is traced back, including: Based on the channel product identifier associated with the path topology distortion intensity exceeding the first threshold or the decision manifold density being lower than the second threshold, extract the corresponding advertising session identifier subset from the blacklist interception log entry set; For each ad serving session identifier in the subset of ad serving session identifiers, perform a join query operation with the user behavior log table; Obtain the user path sequence for ad exposure through related query operations.

5. The data processing method based on a business management platform according to claim 4, characterized in that, The user path sequence for ad exposure includes ad exposure event timestamps, exposure channel identifiers, and user interaction event type sequences.

6. The data processing method based on a business management platform according to claim 1, characterized in that, Analyze the temporal correlation between the blacklist blocking events and product exposure events in the user path of ad exposure, and generate a candidate root cause set, including: Extract the exposure channel identifier corresponding to each exposure event timestamp in the user path sequence of ad exposure; Iterate through the blacklist of blocked log entries, and for each log entry matching the channel product identifier, extract the blocking timestamp and the blocking reason code; Within the time window defined by the timestamps of adjacent exposure events, the number of blacklisted blocks under the same exposure channel identifier is counted, and the interception statistics results at the touchpoint dimension are generated. When the number of interceptions at the touchpoint dimension corresponding to the exposure channel identifier exceeds the third threshold, the corresponding exposure event timestamp and the associated exposure channel identifier will be added to the candidate root cause set.

7. The data processing method based on a business management platform according to claim 6, characterized in that, Based on the temporal distribution characteristics of the timestamps of each exposure event in the candidate root cause set, the abnormal interception-dense segments within a set time range before the exposure event of the target channel product are marked.

8. The data processing method based on a business management platform according to claim 1, characterized in that, Detecting the direction of fluctuations in product conversion rates during periods of channel anomalies in order to identify root causes from a set of candidate root causes, including: If clicks decrease in tandem with channel-level clicks, then the root cause should be identified as the blacklist blocking action at the touchpoint level from the candidate root cause pool; if clicks increase against the trend, then the root cause should be identified as the product-channel matching failure from the candidate root cause pool.

9. A data processing system based on a business management platform, used to implement the data processing method based on a business management platform as described in any one of claims 1-8, characterized in that, include: The data acquisition module is used to obtain the number of clicks from the advertising channels and the number of blacklisted clicks from the touchpoints. The anomaly response module is used to extract the blacklist blocking logs associated with the corresponding channel and the conversion rate of the product dimension when the click volume drops abnormally at the channel level. The policy verification module is used to verify the strength of user behavior path topology distortion caused by blacklist policies based on blacklist interception logs, and to verify the tightness of the decision manifold between channel targets and interception policies. The path backtracking module is used to backtrack the user path of the corresponding channel product when the path topology distortion intensity exceeds the first threshold or the decision manifold density is lower than the second threshold. The root cause candidate module is used to analyze the temporal correlation between the blacklist interception nodes and product exposure nodes in the user path of ad exposure, and generate a candidate root cause set. The root cause identification module is used to detect the direction of fluctuation in product-level conversion rates during periods of channel anomalies in order to identify the root cause from the candidate root cause set.

Images