An advertisement putting risk control method based on big data analysis

Abstract

The application discloses an advertisement putting risk control method based on big data analysis, comprising the following steps: acquiring advertisement putting multi-source behavior data, and constructing putting behavior time sequence; based on the putting behavior time sequence, establishing a multi-level time sequence structure; based on a Bayesian online variable point detection algorithm, constructing a multi-level variable point detection structure, generating a multi-level running length distribution and a multi-level variable point posterior; based on the multi-level running length distribution and the multi-level variable point posterior, generating a multi-level running length state; constructing a risk hypothesis set and generating a risk hypothesis posterior weight set; constructing a risk state structured representation; based on the risk state structured representation, generating a risk control action parameter; and applying the risk control action parameter to the advertisement putting process and performing additional update. The application realizes multi-level accurate identification and closed-loop adaptive control of advertisement putting risks.

Description

Technical Field

[0001] This invention relates to the field of digital advertising delivery and programmatic advertising control, and in particular to a method for controlling advertising delivery risks based on big data analysis. Background Technology

[0002] With the rapid development of digital advertising and programmatic advertising technologies, advertising delivery systems are characterized by large data volumes, multiple structural layers, and frequent dynamic changes in delivery strategies. Advertisers typically manage and control their ads through multi-level structures such as accounts, ad units, and creatives. Delivery behavior forms a complex sequence of business metrics over time, and abnormal fluctuations, strategy failures, and risk injections during the delivery process are becoming increasingly prominent, directly impacting the stability of advertising delivery and the security of funds.

[0003] Existing advertising risk control technologies largely rely on single-indicator threshold judgments, rule triggers, or static model analysis, typically identifying anomalies at a single level or time scale. These methods struggle to depict the interrelationships between different business levels and cannot distinguish between anomalies caused by normal business fluctuations, strategy adjustments, or external risk factors, easily leading to misjudgments or delayed responses. Furthermore, some technologies only focus on whether anomalies occur, lacking a depiction of the anomaly evolution process and its cumulative effects, making it difficult to support risk status assessment in continuous operation scenarios.

[0004] Furthermore, existing technologies often employ fixed rules or simple feedback control methods in the risk management process, lacking structured decision-making basis driven by the inference results of risk status. This results in a lack of clear correspondence between control actions and actual risk sources, making it difficult to achieve refined risk control while ensuring delivery efficiency. This is particularly prominent in multi-level, large-scale, and online advertising delivery scenarios, thus restricting the effectiveness and reliability of advertising delivery risk control technology.

[0005] Therefore, how to provide a method for controlling advertising placement risks based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a risk control method for advertising placement based on big data analysis. This invention constructs a time series of advertising placement behavior data from multiple sources, forming a multi-level time series structure through hierarchical mapping and time aggregation. A Bayesian online change point detection algorithm is introduced to construct a multi-level change point detection structure. Combining multi-level runtime distribution and multi-level change point posterior, the abnormal evolution process of placement behavior is continuously inferred. Furthermore, based on the multi-level runtime state, the posterior weight set of risk assumptions, and the accumulated anomaly, a structured representation of the risk state is constructed, and risk control action parameters are generated accordingly. This achieves deterministic modulation of placement rate parameters, review intensity parameters, and freeze path parameters. Through feedback updates of risk control actions, an online closed-loop risk control process is formed. This invention can distinguish abnormal evolution states under different business levels and different risk generation mechanisms, possessing the advantages of precise risk identification, continuous response, and clear control logic.

[0007] An advertising placement risk control method based on big data analysis according to an embodiment of the present invention includes the following steps:

[0008] Acquire multi-source behavioral data on ad placement, perform time alignment and structuring processing, and construct a time series of placement behaviors;

[0009] Based on the time series of delivery behavior, hierarchical mapping and time aggregation processing are performed to establish a multi-level time series structure;

[0010] Based on the Bayesian online change point detection algorithm, a multi-level change point detection structure is constructed. Combined with the multi-level time series structure, online change point inference processing is performed to generate a multi-level run length distribution. Joint updates are then performed to generate a multi-level change point posterior.

[0011] Based on the multi-level running length distribution and the multi-level variable point posterior, the running length state determination process is performed to determine the multi-level running length state;

[0012] Based on multi-level running length states and multi-level variable point posteriors, hypothesis space splitting inference processing is performed to construct a risk hypothesis set. For each risk hypothesis, posterior update and evidence accumulation are performed to generate a risk hypothesis posterior weight set.

[0013] Based on multi-level runtime states, a set of posterior weights for risk assumptions, and anomaly accumulation, a structured representation of risk states is constructed.

[0014] Based on the structured representation of risk status, risk control action parameters are generated through the mapping rules from risk status to control parameters.

[0015] Apply risk control action parameters to the advertising delivery process, generate risk control action feedback, and execute additional updates.

[0016] Optionally, the generation of the delivery behavior time series includes:

[0017] Collect multi-source behavioral data on advertising delivery corresponding to the advertising delivery process from the advertising delivery system, including exposure data, click data, conversion data, and control data;

[0018] Perform time alignment processing on multi-source behavioral data of advertising under a unified time benchmark, and perform synchronous merging processing with a fixed time window as the alignment granularity;

[0019] After completing the time alignment process, perform structured processing to convert exposure data into exposure statistics, click data into click statistics, conversion data into conversion statistics, and control data into control statistics.

[0020] A time series of campaign activities is constructed based on exposure statistics, click statistics, conversion statistics, and control statistics.

[0021] Optionally, the generation of the multi-level time series structure includes:

[0022] Based on each behavior record in the time series of the delivery behavior, hierarchical mapping processing is performed according to the pre-determined hierarchical identification relationship, so that each behavior record is uniquely mapped to a material level, a delivery unit level and an account media level.

[0023] After completing the hierarchical mapping process, based on the time window boundaries to which each behavior record belongs, time aggregation processing is performed on the behavior records mapped to the same level to form material-level time series, delivery unit-level time series and account media-level time series respectively;

[0024] A multi-level time series structure is constructed based on the material-level time series, the delivery unit-level time series, and the account-media-level time series.

[0025] Optionally, the generation of the multi-level running length distribution and the multi-level variable point posterior includes:

[0026] Based on a multi-level time series structure, online change point inference processing is performed, and the corresponding run length distribution is maintained independently at each level to obtain the run length distribution at the material level, the run length distribution at the delivery unit level, and the run length distribution at the account media level, thus forming a multi-level run length distribution.

[0027] Perform joint updates on multi-level runtime length distributions based on hierarchical dependency update rules;

[0028] After completing the joint update of hierarchical dependencies, based on the online update results of the runtime distribution of each level, the post-change posterior of the material level, the post-change posterior of the delivery unit level, and the post-change posterior of the account media level are calculated respectively, and multi-level post-change posteriors are generated by combining them.

[0029] Optionally, the generation of the multi-level running length state includes:

[0030] Based on the multi-level running length distribution and the multi-level change point posterior, a comprehensive judgment is made on the running length values ​​of each level at the current time point and their corresponding change point probability states to determine the multi-level running length state.

[0031] Based on the determination of the multi-level running length state, cumulative processing is performed on the changes of the multi-level running length state and the posterior change results of the multi-level change point along the time dimension to calculate the abnormal cumulative amount.

[0032] Modulation processing is performed on the posterior update rate of the runtime distribution at the delivery unit level and the runtime distribution at the account media level based on the abnormal cumulative amount.

[0033] Optionally, the generation of the risk hypothesis posterior weight set includes:

[0034] At the current point in time, a triggering decision is made based on the multi-level variable point posterior and the preset variable point posterior threshold triggering condition. When the multi-level variable point posterior meets the preset variable point posterior threshold triggering condition, the hypothesis space splitting inference state is entered.

[0035] After entering the hypothesis space splitting inference state, a risk hypothesis set is constructed, which is composed of a normal change hypothesis branch, a strategy change hypothesis branch, and a risk injection hypothesis branch in parallel.

[0036] Based on the multi-level runtime state, evidence accumulation processing is performed on each risk hypothesis in the risk hypothesis set. The change trend of the multi-level runtime state in the time dimension is used as input, and the corresponding evidence accumulation results are maintained for each risk hypothesis branch.

[0037] Based on the evidence accumulation process, a posterior update process is performed on each risk hypothesis branch in the risk hypothesis set to map the evidence accumulation result to the corresponding risk hypothesis posterior weight, thereby generating a risk hypothesis posterior weight set.

[0038] Optionally, the generation of the structured representation of the risk state includes:

[0039] Based on the running length values ​​and their evolution relationships at different levels, a hierarchical priority comparison and judgment process is performed to determine the risk level position.

[0040] Based on the relative weight relationships of each risk hypothesis branch in the posterior weight set of risk hypotheses, the dominant weight determination process is performed to determine the type of risk generation mechanism;

[0041] Based on the cumulative results of abnormal accumulation over time, a time accumulation trend determination process is performed to determine the evolution intensity.

[0042] By combining the risk level position, risk generation mechanism type, and evolution intensity, a structured representation of risk state is constructed.

[0043] Optionally, the generation of the risk control action parameters includes:

[0044] Based on the risk level position in the structured representation of risk status, perform hierarchical positioning and magnitude selection judgment processing based on the risk level position to determine the adjustment range of the delivery rate parameter;

[0045] Based on the risk generation mechanism type, the dominant mechanism determination and review level selection process based on the risk generation mechanism type are performed to determine the value level of the review intensity parameter;

[0046] Based on the evolution intensity in the structured representation of risk states, the evolution stage determination and execution process determination are performed based on the evolution intensity, and the execution process of the frozen path parameters is determined.

[0047] By combining the deployment rate parameter, review intensity parameter, and freeze path parameter, risk control action parameters are generated.

[0048] Optionally, the supplementary update includes:

[0049] Apply risk control action parameters to the advertising delivery process, perform corresponding delivery rate modulation, review intervention configuration and freeze path advancement operations on the current advertising delivery behavior, and generate risk control action feedback.

[0050] Based on risk control action feedback, obtain corresponding multi-source advertising behavior data, and add it to the advertising behavior time series according to time alignment and structure processing rules, and perform append update;

[0051] The updated delivery behavior time series is input into the multi-level change point detection structure, and online update processing is performed on the multi-level run length distribution and multi-level change point posterior, updating the multi-level run length status.

[0052] Based on the multi-level runtime state update, the risk hypothesis posterior weight set is updated based on the updated multi-level variable point posterior.

[0053] Based on the updated multi-level running length state, the posterior weight set of risk assumptions, and the anomaly accumulation, the risk state structure representation is reconstructed, and the risk control action parameters are updated.

[0054] The beneficial effects of this invention are:

[0055] First, this invention constructs a multi-level time series structure and introduces a multi-level change point detection structure to perform online change point inference processing on multi-source advertising behavior data. This enables real-time mapping of the structural changes in advertising behavior as data continues to arrive. Through joint inference of runtime distribution and multi-level change point posterior, this invention avoids the coarse-grained judgment of advertising anomalies by a single indicator or fixed threshold, achieving fine identification of continuous stable segments and abrupt changes in advertising behavior. This significantly improves the timeliness and accuracy of advertising risk identification.

[0056] Secondly, this invention introduces hierarchical dependency update rules and an anomaly accumulation mechanism based on multi-level runtime distribution. This asymmetrically couples the runtime state changes between different business levels and modulates the posterior update rate through anomaly accumulation, achieving temporal memory and intensity characterization of the risk evolution process. This structure enables risk identification to no longer rely on instantaneous change point probabilities but can distinguish between short-term fluctuations and continuous anomaly evolution, effectively reducing false alarm risk and enhancing the ability to discriminate complex deployment risk scenarios.

[0057] Furthermore, this invention transforms multi-level inference results into structured states such as risk level position, risk generation mechanism type, and evolution intensity through a risk hypothesis set, a risk hypothesis posterior weight set, and a structured representation of risk states. It then generates risk control action parameters based on predetermined mapping rules. This approach achieves a deterministic mapping between risk identification results and ad placement control actions, making the risk control process interpretable and traceable, and forming an online closed-loop risk control process. This enhances the stability, controllability, and automation level of ad placement risk control. Attached Figure Description

[0058] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0059] Figure 1 This is an overall flowchart of an advertising placement risk control method based on big data analysis proposed in this invention;

[0060] Figure 2 This is a schematic diagram of the multi-level change point detection structure constructed based on the Bayesian online change point detection algorithm in this invention;

[0061] Figure 3This is a schematic diagram illustrating the mapping relationship between the structured representation of risk status and the parameters of risk control actions in this invention. Detailed Implementation

[0062] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0063] refer to Figure 1-3 A method for controlling advertising placement risks based on big data analysis includes the following steps:

[0064] Acquire multi-source advertising behavior data, perform time alignment and structuring processing on the multi-source advertising behavior data, and construct a time series of advertising behavior, which includes exposure statistics, click statistics, conversion statistics, and control statistics;

[0065] Based on the time series of delivery behavior, hierarchical mapping and time aggregation processing are performed to establish a multi-level time series structure, which includes material-level time series, delivery unit-level time series, and account media-level time series.

[0066] Based on the Bayesian online change point detection algorithm, a multi-level change point detection structure is constructed. The multi-level time series structure is input into this structure, and online change point inference processing is performed on the time series at the material level, the delivery unit level, and the account media level, respectively. This generates corresponding material-level, delivery unit-level, and account media-level runtime length distributions, forming a multi-level runtime length distribution. A joint update is performed on these multi-level runtime length distributions based on hierarchical dependency update rules. These rules modulate the prior transfer of the delivery unit-level runtime length distribution within the account media-level runtime length distribution, and vice versa. Based on the online update results of each level's runtime length distribution, corresponding material-level, delivery unit-level, and account media-level change point posteriors are generated, forming a multi-level change point posterior.

[0067] Based on the multi-level running length distribution and the multi-level change point posterior, the running length state determination process is performed to determine the multi-level running length state. The multi-level running length state is used to uniformly characterize the evolution state of the continuous and stable segments of the current delivery behavior under different levels. Based on the multi-level running length state and the multi-level change point posterior, the abnormal accumulation amount is calculated. The abnormal accumulation amount is used to modulate the posterior update rate of the delivery unit level running length distribution and the posterior update rate of the account media level running length distribution. The multi-level running length state and the abnormal accumulation amount are output.

[0068] Based on the multi-level running length state and the multi-level change point posterior, the hypothesis space splitting inference process is performed. When the multi-level change point posterior meets the change point posterior threshold triggering condition, a risk hypothesis set is constructed. The risk hypothesis set includes normal change hypothesis branch, strategy change hypothesis branch, and risk injection hypothesis branch. Combined with the multi-level running length state, the posterior update and evidence accumulation are performed on each risk hypothesis in the risk hypothesis set to generate a risk hypothesis posterior weight set.

[0069] Based on the multi-level running length state, the risk hypothesis posterior weight set, and the anomaly accumulation, a structured representation of the risk state is constructed. The structured representation of the risk state includes the risk level position, the risk generation mechanism type, and the evolution intensity.

[0070] Based on the structured representation of risk status, risk control action parameters are generated through the mapping rules from risk status to control parameters. The mapping rules from risk status to control parameters include: determining the adjustment range of the delivery rate parameter according to the risk level position, determining the value level of the review intensity parameter according to the risk generation mechanism type, and determining the execution process of the freeze path parameter according to the evolution intensity, thereby generating risk control action parameters. The risk control action parameters include the delivery rate parameter, the review intensity parameter, and the freeze path parameter.

[0071] The risk control action parameters are applied to the advertising delivery process to generate risk control action feedback. The multi-source behavioral data of advertising delivery corresponding to the risk control action feedback is obtained and the time series of delivery behavior is updated. The updated delivery behavior time series is input into the multi-level change point detection structure to update the multi-level run length state, multi-level change point posterior, hypothesis branch posterior weight set, and the risk control action parameters are updated based on the updated risk state structured representation to form an online closed-loop risk control process.

[0072] In this embodiment, the generation of the delivery behavior time series includes:

[0073] The system collects multi-source behavioral data on advertising campaigns corresponding to the advertising campaign process. This multi-source behavioral data includes exposure data, click data, conversion data, and control data. The control data consists of campaign rate adjustment records, review action records, and freeze action records, which are used to characterize the risk control actions that have been performed during the advertising campaign process.

[0074] Time alignment processing is performed on multi-source behavioral data of advertising placement under a unified time benchmark. The time alignment processing uses a fixed time window as the alignment granularity. Within the same time window, exposure data, click data, conversion data and control data are synchronously merged to ensure the consistency of different types of behavioral data in the time dimension.

[0075] After completing the time alignment process, the multi-source behavior data of advertising placement merged into the same time window is subjected to structured processing, which converts the exposure data into exposure statistics, the click data into click statistics, the conversion data into conversion statistics, and the control data into control statistics.

[0076] Based on exposure statistics, click statistics, conversion statistics, and control statistics, a time series of campaign behavior is constructed in the order of time windows. This time series of campaign behavior serves as the unified input data basis for subsequent construction of multi-level time series structures, construction of multi-level change point detection structures, and risk control closed-loop updates.

[0077] In this embodiment, the generation of the multi-level time series structure includes:

[0078] Based on each behavior record in the time series of the delivery behavior, hierarchical mapping processing is performed on the behavior record according to the predetermined hierarchical identification relationship. Within any time window, each behavior record is mapped to a material level, a delivery unit level and an account media level respectively and uniquely.

[0079] After completing the hierarchical mapping process, based on the time window boundaries of each behavior record in the delivery behavior time series, time aggregation processing is performed on behavior records mapped to the same material level, the same delivery unit level, and the same account media level. The time aggregation process uses a consistent time boundary within the same time window. During the time aggregation process, delivery behavior time series segments mapped to the same material level and located within the same time window are aggregated to form a material level time series, delivery behavior time series segments mapped to the same delivery unit level and located within the same time window are aggregated to form a delivery unit level time series, and delivery behavior time series segments mapped to the same account media level and located within the same time window are aggregated to form an account media level time series.

[0080] Based on the time series of the material level, the time series of the delivery unit level, and the time series of the account media level, a multi-level time series structure is constructed. This multi-level time series structure serves as the unified time-aligned input for the online change point inference processing of each level in the subsequent multi-level change point detection structure.

[0081] In this embodiment, the generation of the multi-level running length distribution and the multi-level variable point posterior includes:

[0082] Based on a multi-level time series structure, online change point inference processing is performed on the time series at the material level, the time series at the delivery unit level, and the time series at the account media level, respectively. The corresponding runtime distribution is maintained independently at each level, resulting in the runtime distributions at the material level, the delivery unit level, and the account media level. The runtime distributions at the material level, the delivery unit level, and the account media level are then combined to form a multi-level runtime distribution.

[0083] The online change point inference processing steps include: receiving new observation data hourly for the material-level time series, the delivery unit-level time series, and the account media-level time series according to the update order of the delivery behavior time series in the time dimension; performing conditional inference processing on the new observation data in the time series corresponding to the current moment based on the running length distribution maintained under each level, calculating the running length update result under the condition of maintaining the continuous extension of the current running length, and the running length reset result under the condition of a new change point; and performing normalization update on the running length distribution under each level based on the above two types of running length update results to obtain the material-level running length distribution, the delivery unit-level running length distribution, and the account media-level running length distribution corresponding to the current moment.

[0084] Based on the completed online change-point inference processing, a joint update is performed on the runtime distributions of the media layer, delivery unit layer, and account media layer in the multi-level runtime distribution based on hierarchical dependency update rules. Specifically, the runtime values ​​and probability distribution patterns reflected in the account media layer runtime distribution are used to modulate the prior transition probability of the delivery unit layer runtime distribution, so that when performing online change-point inference processing, the delivery unit layer adjusts its prior tendency to change points based on the degree of continuous operation represented by the account media layer runtime distribution. Similarly, the runtime values ​​and probability distribution patterns reflected in the delivery unit layer runtime distribution are used to modulate the prior transition probability of the media layer runtime distribution, so that when performing online change-point inference processing, the media layer adjusts its prior tendency to change points based on the degree of continuous operation represented by the delivery unit layer runtime distribution. The hierarchical dependency update rules restrict lower-level runtime distributions from having a reverse effect on the prior structure of upper-level runtime distributions.

[0085] After completing the hierarchical dependency joint update, based on the online update results of the runtime distribution of each level, the material level change post-hoc, the delivery unit level change post-hoc, and the account media level change post-hoc are calculated respectively. The material level change post-hoc, the delivery unit level change post-hoc, and the account media level change post-hoc are combined to generate a multi-level change post-hoc, which is used to characterize the probability state of structural changes in delivery behavior under different business levels.

[0086] In this embodiment, the generation of the multi-level running length state includes:

[0087] Based on the multi-level running length distribution and the multi-level change point posterior, a comprehensive judgment is made on the running length values ​​of each level and their corresponding change point probability states at the current time point to determine the multi-level running length state. The multi-level running length state is used to uniformly characterize the evolution state of the continuous and stable segment of the current deployment behavior at different levels.

[0088] The comprehensive determination includes, at the current time point, extracting the corresponding running length value state of each level from the multi-level running length distribution, and extracting the corresponding change point probability state of each level from the multi-level change point posterior. Based on the running length value state of each level, a consistency check is performed on the continuity of the current delivery behavior's continuous and stable segment. The consistency check is used to determine whether the running length value state of the material level, the running length value state of the delivery unit level, and the running length value state of the account media level satisfy a unidirectional evolution relationship at the same time point. If the consistency check passes, the running length value states of each level are jointly merged according to the hierarchical dependency chain to obtain the multi-level running length state. If the consistency check fails, the running length value states of each level are checked based on the multi-level change point posterior. Conflict resolution processing is performed. Under the condition that the posterior of the change point at the account media level is higher than that at the delivery unit level, the running length value at the account media level is taken as the dominant state and the running length value at the delivery unit level is reset. Under the condition that the posterior of the change point at the delivery unit level is higher than that at the material level, the running length value at the delivery unit level is taken as the dominant state and the running length value at the material level is reset. After the above conflict resolution processing is completed, a multi-level running length state is formed. The resetting means that when the dominant condition of the upper-level change point is met, the continuous extension of the lower-level running length state under the current generation mechanism is terminated and the lower-level running length state is switched to a new starting state so that the subsequent update of the lower-level running length distribution is based on the new structural assumptions.

[0089] Based on the determination of the multi-level running length state, the changes of the multi-level running length state and the posterior of the multi-level change point are cumulatively processed along the time dimension to calculate the abnormal cumulative amount. The abnormal cumulative amount represents the degree of time accumulation of the multi-level running length state deviating from the continuous and stable evolution trend.

[0090] Specifically, the calculation of the anomaly accumulation involves first acquiring the multi-level running length state at each time point to characterize the running state of the current deployment behavior in the continuous stable segment at each business level. Simultaneously, the multi-level change point posterior at the same time point is acquired to characterize the probability state of structural changes at each business level. When the multi-level running length state shows a continuous shortening trend or a switch from a continuous stable segment to the initial running state occurs, and the multi-level change point posterior at the corresponding time point is in a continuously non-zero probability state, this time point is included in the anomaly accumulation process. When the multi-level running length state maintains continuous extension in the time dimension, and the multi-level change point posterior maintains a low probability state within a continuous time window, no cumulative update is performed on the anomaly accumulation. The anomaly accumulation continuously accumulates over time the deviation of the multi-level running length state from the continuous stable evolution trend, characterizing the sustained intensity of anomaly evolution over time, rather than the instantaneous anomaly degree at a single time point.

[0091] Based on the abnormal accumulation, the posterior update rate of the delivery unit level running length distribution and the account media level running length distribution in the multi-level running length distribution is modulated. The abnormal accumulation is only used to modulate the posterior update rate, does not participate in the change point determination process, and is not used as the change point trigger threshold. The multi-level running length status and abnormal accumulation are output.

[0092] In this embodiment, the generation of the risk assumption posterior weight set includes:

[0093] At the current point in time, a triggering decision is made based on the multi-level variable point posterior and the preset variable point posterior threshold triggering condition. When the multi-level variable point posterior meets the preset variable point posterior threshold triggering condition, the hypothesis space splitting inference state is entered.

[0094] After entering the hypothesis space split inference state, a risk hypothesis set is constructed. The risk hypothesis set is composed of a normal change hypothesis branch, a strategy change hypothesis branch, and a risk injection hypothesis branch in parallel. The normal change hypothesis branch, the strategy change hypothesis branch, and the risk injection hypothesis branch exist simultaneously at the same point in time and are independent of each other. No uniqueness determination is performed on the risk type to which the deployment behavior belongs.

[0095] After the risk hypothesis set is constructed, evidence accumulation processing is performed on each risk hypothesis in the risk hypothesis set based on the multi-level running length state. The evidence accumulation processing takes the change trend of the multi-level running length state in the time dimension as input, and maintains the corresponding evidence accumulation results for each risk hypothesis branch to characterize the degree of support of the current deployment behavior evolution process for each risk generation mechanism.

[0096] Based on the evidence accumulation processing, a posteriori update processing is performed on each risk hypothesis branch in the risk hypothesis set, and the evidence accumulation results of each risk hypothesis branch are mapped to the corresponding risk hypothesis posterior weights to generate a risk hypothesis posterior weight set. The risk hypothesis posterior weight set is used to characterize the relative validity of different risk generation mechanisms at the current time point.

[0097] In this embodiment, the generation of the structured representation of the risk state includes:

[0098] Based on the multi-level operation length status, the risk level position is determined according to the operation length value status and evolution relationship at different levels. The risk level position is used to indicate the business level position corresponding to the abnormal evolution of the current deployment behavior.

[0099] Specifically, at the current time point, the system reads the running length values ​​for each level within the multi-level running length status. It then compares these running length values ​​according to the hierarchical dependency order of account media level, placement unit level, and creative material level to identify the level where the running length value status begins to change. When the running length value at the account media level begins to change, the risk level is determined to be the account media level. When the running length value at the account media level remains continuously extended and the running length value at the placement unit level begins to change, the risk level is determined to be the placement unit level. Finally, when the running length value at the account media level remains continuously extended and the running length value at the placement unit level remains continuously extended and the running length value at the creative material level begins to change, the risk level is determined to be the creative material level.

[0100] Based on the posterior weight set of risk assumptions, the risk generation mechanism type is uniquely determined by the relative weight relationship of each risk assumption branch in the posterior weight set of risk assumptions. The risk generation mechanism type is selected from the normal change assumption branch, the strategy change assumption branch and the risk injection assumption branch to select the corresponding dominant risk generation mechanism, which is used to characterize the risk generation mode corresponding to the abnormal evolution of the current deployment behavior.

[0101] Specifically, at the current point in time, the posterior weights of the risk assumptions corresponding to the normal change assumption branch, the strategy change assumption branch, and the risk injection assumption branch in the posterior weight set of risk assumptions are read. The relative importance of these three posterior weights is determined, and the risk assumption branch with the largest weight value is identified as the dominant risk generation mechanism. When the posterior weight of the risk assumption corresponding to the normal change assumption branch is the maximum value, the risk generation mechanism type is identified as the normal change assumption branch. When the posterior weight of the risk assumption corresponding to the strategy change assumption branch is the maximum value, the risk generation mechanism type is identified as the strategy change assumption branch. When the posterior weight of the risk assumption corresponding to the risk injection assumption branch is the maximum value, the risk generation mechanism type is identified as the risk injection assumption branch.

[0102] Based on the cumulative amount of anomalies, the evolution intensity is determined by the cumulative result of the cumulative amount of anomalies over time. The evolution intensity is used to characterize the persistence and evolution magnitude of the current risk state over time.

[0103] Specifically, at the current time point, the current value of the abnormal accumulation is read, and the cumulative change of the abnormal accumulation at adjacent time points is read. The cumulative growth process of the abnormal accumulation is segmented and judged. When the abnormal accumulation continues to accumulate and increase at consecutive time points, the evolution intensity is determined to be in a continuously increasing state. When the abnormal accumulation remains unchanged at consecutive time points, the evolution intensity is determined to be in a continuously stable state. When the abnormal accumulation shows a cumulative decline at consecutive time points, the evolution intensity is determined to be in a decaying and declining state.

[0104] By combining the risk level position, risk generation mechanism type, and evolution intensity, a structured representation of the risk state is constructed. This structured representation of the risk state serves as the sole input for the risk control stage and is used to characterize the structured state of the current advertising placement risk.

[0105] In this embodiment, the generation of the risk control action parameters includes:

[0106] Based on the risk level position in the structured representation of risk status, the adjustment range of the delivery rate parameter is determined. The adjustment range of the delivery rate parameter is used to characterize the magnitude of continuous modulation of the advertising delivery rhythm at the current risk level position. Moreover, the risk level position only affects the adjustment range of the delivery rate parameter and does not participate in the determination of the review intensity parameter and the freeze path parameter.

[0107] Specifically, at the current point in time, the risk level position is read, and a level positioning determination is performed on the risk level position. When the risk level position is determined to be the account media level, the corresponding account media level's delivery rate parameter adjustment range is selected; when the risk level position is determined to be the delivery unit level, the corresponding delivery unit level's delivery rate parameter adjustment range is selected; when the risk level position is determined to be the creative level, the corresponding creative level's delivery rate parameter adjustment range is selected. The selected delivery rate parameter adjustment range serves as the benchmark for continuously modulating the current advertising delivery rhythm, and the delivery rate parameter is updated according to the adjustment range during subsequent advertising delivery.

[0108] Based on the risk generation mechanism type in the structured representation of risk status, the value level of the review intensity parameter is determined. The value level of the review intensity parameter is used to characterize the degree of review intervention corresponding to the current risk generation mechanism type. Moreover, the risk generation mechanism type only affects the value level of the review intensity parameter and does not participate in the determination of the delivery rate parameter and the freeze path parameter.

[0109] Specifically, at the current point in time, the risk generation mechanism type is read, and a mechanism type determination is performed. When the risk generation mechanism type is determined to be a normal change assumption branch, the first value level of the corresponding review intensity parameter is selected; when the risk generation mechanism type is determined to be a strategy change assumption branch, the second value level of the corresponding review intensity parameter is selected; when the risk generation mechanism type is determined to be a risk injection assumption branch, the third value level of the corresponding review intensity parameter is selected. The selected review intensity parameter value level is used to determine the degree of review intervention in the current advertising process and serves as the basis for configuring the intensity of subsequent review actions.

[0110] Based on the evolution intensity in the structured representation of risk status, the execution process of the freezing path parameters is determined. The execution process of the freezing path parameters is used to characterize the degree of progress of the freezing operation in the time dimension. The evolution intensity only acts on the execution process of the freezing path parameters and does not participate in the determination of the delivery rate parameter and the review intensity parameter.

[0111] Specifically, at the current time point, the evolution intensity is read, and an evolution stage determination is performed on the evolution intensity. When the evolution intensity is determined to be in a continuous stable state, the first execution process of the frozen path parameters is selected; when the evolution intensity is determined to be in a continuous enhancing state, the second execution process of the frozen path parameters is selected; when the evolution intensity is determined to be in a decaying and declining state, the third execution process of the frozen path parameters is selected. The selected execution process of the frozen path parameters is used to determine the advancement stage of the freezing operation in the time dimension, and the frozen path parameters are advanced or regressed according to the execution process during subsequent advertising.

[0112] The delivery rate parameter, review intensity parameter, and freeze path parameter are combined to generate risk control action parameters. These risk control action parameters serve as control inputs for the advertising delivery process, enabling continuous and adjustable risk control processing of advertising delivery behavior.

[0113] In this embodiment, the additional update includes:

[0114] Apply risk control action parameters to the advertising delivery process, perform corresponding delivery rate modulation, review intervention configuration and freeze path advancement operations on the current advertising delivery behavior, and generate risk control action feedback.

[0115] Based on risk control action feedback, the corresponding multi-source advertising behavior data is obtained, and the multi-source advertising behavior data is added to the advertising behavior time series according to time alignment and structured processing rules, and the advertising behavior time series is updated.

[0116] The updated delivery behavior time series is input into the multi-level change point detection structure, and online update processing is performed on the multi-level running length distribution and multi-level change point posterior. The multi-level running length status is updated based on the online update processing result.

[0117] Based on the multi-level running length state update, the risk hypothesis posterior weight set is updated based on the updated multi-level variable point posterior. The posterior update process only applies to the risk hypothesis posterior weight set and does not modify the structure of the risk hypothesis set.

[0118] Based on the updated multi-level running length status, risk hypothesis posterior weight set, and anomaly accumulation, a structured representation of risk status is reconstructed, and risk control action parameters are updated based on the updated risk status structure, forming an online closed-loop risk control process with a multi-level change point detection structure as the core recharge object.

[0119] Example 1: To verify the feasibility of this invention in practice, it was applied to a large-scale online advertising campaign. In this scenario, the advertising platform simultaneously hosts a large number of accounts, media, ad placement units, and creative materials in parallel. During the campaign, it continuously generates multi-source behavioral data on advertising, including exposure statistics, click statistics, conversion statistics, and control statistics. As the scale of advertising expands and the campaign strategy is frequently adjusted, advertising behavior exhibits significant non-stationary characteristics over time. It is difficult to promptly distinguish whether some abnormal campaign behaviors are caused by normal strategy adjustments or by abnormal risk injections, easily leading to budget waste, decreased conversion efficiency, and delayed platform risk control responses. This example focuses on addressing these issues and verifies the application effect of the method of this invention in a real business environment.

[0120] In practical applications, the system continuously receives multi-source advertising behavior data and performs time alignment and structuring processing on various types of behavior data at a unified time granularity, forming a time series of advertising behavior that can be used for subsequent analysis. This time series simultaneously reflects core advertising performance indicators such as exposure, clicks, and conversions, while retaining control statistics such as advertising rate adjustments, review action triggers, and freeze operations, thus forming a complete data foundation for subsequent risk control. Based on this, the system performs hierarchical mapping and time aggregation processing on the advertising behavior time series, mapping the original behavior sequences to creative-level time series, ad-hoc unit-level time series, and account-media-level time series, ensuring consistent data boundaries and comparability across different business levels within the same time window.

[0121] Subsequently, the system constructs a multi-level change point detection structure based on the Bayesian online change point detection algorithm, and inputs the multi-level time series structure into this structure to perform online change point inference processing. During the continuous reception of new observation data, the system maintains corresponding runtime distributions at the material level, delivery unit level, and account media level, respectively, and performs joint updates through hierarchical dependency update rules. This yields a multi-level runtime distribution reflecting the operational stability of different business levels, along with the corresponding multi-level change point posteriors. Through this structure, the system can not only capture short-term fluctuations of single indicators but also identify structural changes propagating across levels.

[0122] Building upon this foundation, the system generates multi-level runtime states based on multi-level runtime distribution and multi-level change point posteriors. These states characterize the continuous and stable segment evolution of the current deployment behavior across different business levels. By accumulating runtime states and change point posteriors along the time dimension, the system further calculates the accumulated anomaly amount. This ensures that anomaly judgment no longer relies on the sudden change at a single point in time but reflects the persistence and evolutionary trend of anomalous behavior. Combining multi-level runtime states and multi-level change point posteriors, the system constructs a set of risk hypotheses and performs evidence accumulation and posterior update on the normal change hypothesis branch, strategy change hypothesis branch, and risk injection hypothesis branch, respectively, generating a set of risk hypothesis posterior weights. This allows for the differentiation of anomaly generation mechanisms.

[0123] During the risk state generation phase, the system constructs a structured representation of the risk state based on multi-level runtime states, the posterior weight set of risk assumptions, and the accumulated anomalies, explicitly defining the risk level location, risk generation mechanism type, and evolution intensity. This structured representation serves as the sole input for the subsequent control phase, avoiding the uncertainty of manual rule overlay or experience-based judgment. Subsequently, the system generates risk control action parameters according to the mapping rules from risk states to control parameters, continuously modulating the delivery rate parameters, review intensity parameters, and freeze path parameters, and applying these risk control action parameters to the ad delivery process. Feedback data generated after the control actions are executed flows back to the multi-level change point detection structure, achieving online closed-loop updates and enabling the system to continuously adapt and adjust in a dynamic delivery environment.

[0124] During implementation, the same batch of advertising data was selected to compare the overall performance of the method of this invention, the anomaly monitoring method based on fixed thresholds, and the method based on single-level change point detection. Continuous advertising behavior covering multiple accounts, hundreds of ad placement units, and thousands of creative materials within the statistical period was used to comprehensively evaluate the risk identification and ad placement control effects. The experimental results are summarized in the table below.

[0125] Table 1. Comparison of the Effectiveness of Advertising Placement Risk Control Methods

[0126] Comparison indicators Fixed threshold anomaly monitoring method Single-level change point detection method Method of the present invention Risk identification accuracy (%) 78.4 83.1 86.9 False alarm rate (%) 14.6 11.2 9.3 Average risk response delay (minutes) 18.5 14.2 11.6 Conversion rate fluctuation range (%) 12.8 10.1 8.7 Percentage of wasted budget in a single cycle (%) 6.4 5.1 4.3

[0127] As shown in Table 1, the fixed threshold anomaly detection method, relying on static rules, has weak adaptability to structural changes in campaign behavior. It is prone to high false alarm rates when campaign strategies are frequently adjusted, resulting in long risk response delays and significant fluctuations in campaign conversion rates. Single-level change point detection methods improve the timeliness of anomaly identification to some extent, but because inference is only performed at a single business level, it is difficult to distinguish anomaly propagation paths, limiting the improvement in risk identification accuracy. The method of this invention, through joint inference of multi-level runtime distribution and multi-level change point posterior, improves the risk identification accuracy to 86.9% while keeping the false alarm rate at a low level.

[0128] Regarding risk response, the method of this invention utilizes the accumulated anomalies and structured representations of risk states to continuously characterize risk evolution, enabling earlier intervention in risk control actions and reducing the average risk response delay to 11.6 minutes. Because control actions employ a continuous modulation approach rather than sudden intervention, fluctuations in campaign conversion rates are effectively suppressed, significantly improving overall campaign stability. Simultaneously, by more accurately distinguishing between strategy changes and risk injection behaviors, the system avoids unnecessary over-intervention, further reducing the proportion of wasted budget per cycle.

[0129] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A method for controlling advertising placement risks based on big data analysis, characterized in that, Includes the following steps: Acquire multi-source behavioral data on ad placement, perform time alignment and structuring processing, and construct a time series of placement behaviors; Based on the time series of delivery behavior, hierarchical mapping and time aggregation processing are performed to establish a multi-level time series structure; Based on the Bayesian online change point detection algorithm, a multi-level change point detection structure is constructed. Combined with the multi-level time series structure, online change point inference processing is performed to generate a multi-level run length distribution. Joint updates are then performed to generate a multi-level change point posterior. Based on the multi-level running length distribution and the multi-level variable point posterior, the running length state determination process is performed to determine the multi-level running length state; Based on multi-level running length states and multi-level variable point posteriors, hypothesis space splitting inference processing is performed to construct a risk hypothesis set. For each risk hypothesis, posterior update and evidence accumulation are performed to generate a risk hypothesis posterior weight set. Based on multi-level runtime states, a set of posterior weights for risk assumptions, and anomaly accumulation, a structured representation of risk states is constructed. Based on the structured representation of risk status, risk control action parameters are generated through the mapping rules from risk status to control parameters. Apply risk control action parameters to the ad delivery process, generate risk control action feedback, and execute additional updates. The generation of the multi-level running length distribution and the multi-level change point posterior includes: Based on a multi-level time series structure, online change point inference processing is performed, and the corresponding run length distribution is maintained independently at each level to obtain the run length distribution at the material level, the run length distribution at the delivery unit level, and the run length distribution at the account media level, thus forming a multi-level run length distribution. Perform joint updates on multi-level runtime length distributions based on hierarchical dependency update rules; After completing the joint update of hierarchical dependencies, based on the online update results of the runtime distribution of each level, the posterior of the change point at the material level, the posterior of the change point at the delivery unit level, and the posterior of the change point at the account media level are calculated respectively, and combined to generate multi-level posterior of change points. The generation of the multi-level running length state includes: Based on the multi-level running length distribution and the multi-level change point posterior, a comprehensive judgment is made on the running length values ​​of each level at the current time point and their corresponding change point probability states to determine the multi-level running length state. Based on the determination of the multi-level running length state, cumulative processing is performed on the changes of the multi-level running length state and the posterior change results of the multi-level change point along the time dimension to calculate the abnormal cumulative amount. Modulation processing is performed on the posterior update rate of the runtime distribution at the delivery unit level and the runtime distribution at the account media level based on the abnormal cumulative amount.

2. The advertising placement risk control method based on big data analysis according to claim 1, characterized in that, The generation of the time series of the delivery behavior includes: Collect multi-source behavioral data on advertising delivery corresponding to the advertising delivery process from the advertising delivery system, including exposure data, click data, conversion data, and control data; Perform time alignment processing on multi-source behavioral data of advertising under a unified time benchmark, and perform synchronous merging processing with a fixed time window as the alignment granularity; After completing the time alignment process, perform structured processing to convert exposure data into exposure statistics, click data into click statistics, conversion data into conversion statistics, and control data into control statistics. A time series of campaign activities is constructed based on exposure statistics, click statistics, conversion statistics, and control statistics.

3. The advertising placement risk control method based on big data analysis according to claim 1, characterized in that, The generation of the multi-level time series structure includes: Based on each behavior record in the time series of the delivery behavior, hierarchical mapping processing is performed according to the pre-determined hierarchical identification relationship, so that each behavior record is uniquely mapped to a material level, a delivery unit level and an account media level. After completing the hierarchical mapping process, based on the time window boundaries of each behavior record, time aggregation processing is performed on the behavior records mapped to the same level to form material-level time series, delivery unit-level time series and account media-level time series respectively; A multi-level time series structure is constructed based on the material-level time series, the delivery unit-level time series, and the account-media-level time series.

4. The advertising placement risk control method based on big data analysis according to claim 1, characterized in that, The generation of the posterior weight set of the risk assumptions includes: At the current point in time, a triggering decision is made based on the multi-level variable point posterior and the preset variable point posterior threshold triggering condition. When the multi-level variable point posterior meets the preset variable point posterior threshold triggering condition, the hypothesis space splitting inference state is entered. After entering the hypothesis space splitting inference state, a risk hypothesis set is constructed, which is composed of a normal change hypothesis branch, a strategy change hypothesis branch, and a risk injection hypothesis branch in parallel. Based on the multi-level runtime state, evidence accumulation processing is performed on each risk hypothesis in the risk hypothesis set. The change trend of the multi-level runtime state in the time dimension is used as input to generate corresponding evidence accumulation results for each risk hypothesis branch. Based on the evidence accumulation processing, a posterior update processing is performed on each risk hypothesis branch in the risk hypothesis set to map the evidence accumulation result to the corresponding risk hypothesis posterior weight, thereby generating a risk hypothesis posterior weight set.

5. The advertising placement risk control method based on big data analysis according to claim 1, characterized in that, The generation of the structured representation of the risk state includes: Based on the running length values ​​and their evolution relationships at different levels, a hierarchical priority comparison and judgment process is performed to determine the risk level position. Based on the relative weight relationships of each risk hypothesis branch in the posterior weight set of risk hypotheses, the dominant weight determination process is performed to determine the type of risk generation mechanism; Based on the cumulative results of abnormal accumulation over time, a time accumulation trend determination process is performed to determine the evolution intensity. By combining the risk level position, risk generation mechanism type, and evolution intensity, a structured representation of risk state is constructed.

6. The advertising placement risk control method based on big data analysis according to claim 1, characterized in that, The generation of the risk control action parameters includes: Based on the risk level position in the structured representation of risk status, perform hierarchical positioning and magnitude selection judgment processing based on the risk level position to determine the adjustment range of the delivery rate parameter; Based on the risk generation mechanism type, the dominant mechanism determination and review level selection process based on the risk generation mechanism type are performed to determine the value level of the review intensity parameter; Based on the evolution intensity in the structured representation of risk states, the evolution stage determination and execution process determination are performed based on the evolution intensity, and the execution process of the frozen path parameters is determined. By combining the deployment rate parameter, review intensity parameter, and freeze path parameter, risk control action parameters are generated.

7. The advertising placement risk control method based on big data analysis according to claim 1, characterized in that, The additional updates include: Apply risk control action parameters to the advertising delivery process, perform corresponding delivery rate modulation, review intervention configuration and freeze path advancement operations on the current advertising delivery behavior, and generate risk control action feedback. Based on risk control action feedback, obtain corresponding multi-source advertising behavior data, and add it to the advertising behavior time series according to time alignment and structure processing rules, and perform append update; The updated delivery behavior time series is input into the multi-level change point detection structure, and online update processing is performed on the multi-level run length distribution and multi-level change point posterior, updating the multi-level run length status. Based on the multi-level runtime state update, the risk hypothesis posterior weight set is updated based on the updated multi-level variable point posterior. Based on the updated multi-level running length state, the posterior weight set of risk assumptions, and the anomaly accumulation, the risk state structure representation is reconstructed, and the risk control action parameters are updated.

Images