An advertisement putting intelligent monitoring method and system based on big data analysis

Abstract

The application discloses an advertisement putting intelligent monitoring method and system based on big data analysis, comprising the following steps: collecting multi-source heterogeneous data of advertisement putting, preprocessing into monitoring data and multi-dimensional index sequence; constructing a monitoring object and generating characteristic data; constructing an improved causal Transformer model to obtain a monitoring result; performing dimension refinement analysis to generate a candidate disposal action set; obtaining an estimation result and a disassembly result based on a causal forest algorithm; determining a target disposal action to generate a disposal instruction. The application combines the disposal effect evaluation of the improved causal Transformer model and the causal forest algorithm, realizes accurate monitoring, rapid positioning and linkage disposal of advertisement putting abnormity.

Description

Technical Field

[0001] This invention relates to the field of big data analytics, and in particular to an intelligent monitoring method and system for advertising placement based on big data analytics. Background Technology

[0002] With the continuous expansion of internet advertising and the development of programmatic and cross-media advertising, the advertising delivery process typically involves multiple stages, including advertising platforms, campaign management systems, creative and landing page systems, client-side event tracking, and third-party monitoring and anti-fraud platforms. This results in a significant multi-source heterogeneity in advertising data sources. Multi-source heterogeneous data typically includes exposure logs, click logs, conversion logs, consumption data, and strategy change data. Different data sources differ in data structure, field naming, statistical definitions, attribution windows, deduplication rules, and timestamp systems. Furthermore, the collection, transmission, and storage processes are prone to issues such as duplicate reporting, data loss, out-of-order delivery, delayed data transmission, and noise interference. All these differences and anomalies make it difficult to align metrics for the same advertising target across different stages, further leading to inconsistent monitoring data definitions, unstable metric fluctuations, and difficulties in verifying and tracing results. This ultimately affects the accuracy of advertising performance evaluation, budget control, and risk identification.

[0003] Existing advertising monitoring solutions primarily rely on report statistics or fixed threshold alerts. Common practices include setting static thresholds for single metrics such as click-through rate (CTR), conversion rate (CTR), spend, and cost per conversion (CPC) and triggering alerts, or using simple year-on-year / month-on-month comparison rules to identify anomalies. However, with frequent changes in the advertising environment, rapid iteration of channels and creatives, fluctuations in traffic quality, and frequent strategy adjustments, fixed thresholds or single-metric rules struggle to adapt to the inherent volatility differences among different advertising targets. They also fail to promptly capture the interconnected changes of multiple metrics and structural abrupt changes in time series, leading to false positives, missed detections, and delayed alerts. Furthermore, existing solutions often lack the ability to fine-grainedly analyze the scope of anomaly impact after it is triggered, making it difficult to quickly pinpoint the correlation between anomalies and specific regions, devices, ad placements, or strategy changes. This results in handling strategies relying on manual experience, long response times, and uncontrollable risks, potentially leading to increased ineffective spending or decreased advertising revenue.

[0004] Therefore, how to provide an intelligent monitoring method and system for advertising placement based on big data analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0005] One objective of this invention is to propose an intelligent monitoring method and system for advertising placement based on big data analytics. This invention comprehensively utilizes data analysis and artificial intelligence methods, including unified alignment and correction of multi-source heterogeneous data, temporal causal modeling, and causal effect evaluation. It fully describes the entire process from advertising placement data collection and preprocessing, generation of unified monitoring data and multi-dimensional indicator sequences, construction of monitoring objects and generation of feature data, anomaly detection and root cause localization, generation of candidate actions, evaluation of action effects, to generation of action instructions. In terms of model structure, it innovatively constructs an improved causal Transformer model, including multi-path causal modeling, temporal causal attention fusion, and dynamic covariate filtering. It also constructs a causal forest algorithm, including a back-estimation structure fusion tree and a hierarchical effect decomposition tree structure, to achieve adaptive multi-indicator detection of placement anomalies, localization of the scope of anomaly impact, and incremental effect evaluation of action measures. Compared with existing technologies, this invention has the advantages of high consistency in monitoring standards, timely anomaly detection, high efficiency in localization and action, reduced false alarms and missed alarms, and ease of engineering deployment.

[0006] According to an embodiment of the present invention, an intelligent monitoring method for advertising placement based on big data analysis includes:

[0007] Collect multi-source heterogeneous data on advertising placement, preprocess the multi-source heterogeneous data, and generate unified monitoring data and multi-dimensional indicator sequences;

[0008] Based on monitoring data and multidimensional indicator sequences, monitoring objects are constructed, and feature data is generated. The feature data includes multi-indicator time series, deployment action sequences, and covariate features.

[0009] An improved causal Transformer model is constructed, and multi-path causal modeling is introduced to divide the feature data into multiple causal paths, which are then processed by sequence encoding. Temporal causal attention fusion is used to perform causal mask self-attention calculation and fusion. The volatility is calculated by dynamically filtering covariates to obtain the monitoring results.

[0010] Based on the monitoring results, a detailed dimensional analysis of the monitored objects is performed, and a set of candidate actions is generated after locating the scope of the anomaly's impact.

[0011] The causal forest algorithm is used to evaluate the treatment effect of the candidate treatment action set. The inverse estimation structure fusion tree is introduced to construct the forward estimation tree and the inverse estimation tree respectively. The hierarchical effect decomposition tree structure is decomposed to generate hierarchical effect components, and the incremental effect estimation result and hierarchical effect decomposition result are obtained.

[0012] Based on the incremental effect estimation results, hierarchical effect decomposition results, and monitoring results, target action actions are determined and action instructions are generated.

[0013] Optionally, the multi-source heterogeneous data includes exposure logs, click logs, conversion logs, consumption data, and strategy change data.

[0014] Optionally, the generation of unified monitoring data and multidimensional indicator sequences includes:

[0015] Acquire multi-source heterogeneous data and map various data records to a unified set of event fields. The unified set of event fields includes event type, event occurrence time, access time, activity identifier, channel identifier, material identifier, user identifier, device identifier, cost field, and conversion value field.

[0016] Preprocessing is performed on the data records corresponding to the unified event field set. The preprocessing includes format standardization, missing value marking, duplicate record identification and removal, and outlier identification and removal. Data records whose event occurrence time is earlier than the start time of the current processing window and whose access time is later than the end time of the current processing window are marked as late data.

[0017] The preprocessed data records are subjected to association and alignment processing, which includes link association processing based on event identifiers or tracking identifiers, time alignment processing based on event occurrence time, and recovery processing for late data. Based on the results of association and alignment processing, monitoring data with unified caliber and multi-dimensional indicator sequences are generated.

[0018] Optionally, the generation of feature data includes:

[0019] Based on the monitoring data, monitoring objects are constructed. The monitoring objects are determined by activity identifiers, channel identifiers, and material identifiers. The monitoring data is then grouped according to the monitoring objects.

[0020] For each monitored object, a multi-indicator time series is generated according to the time window. The multi-indicator time series includes exposure series, click series, conversion series, consumption series, and derived indicator series are generated.

[0021] The deployment action sequence is generated based on the strategy change data. Covariate features are extracted from the monitoring data. The multi-indicator time series, deployment action sequence and covariate features are aligned by time to form feature data.

[0022] Optionally, the obtained monitoring results include:

[0023] An improved causal Transformer model is constructed, including a multi-path causal modeling module, a temporal causal attention fusion module, and a covariate dynamic filtering module;

[0024] The multi-path causal modeling module splits the feature data into three causal paths and performs sequence encoding processing on each of them. The three causal paths include the action path, the indicator path, and the covariate path. The action path consists of the sequence of delivery actions, the indicator path consists of the time series of multiple indicators, and the covariate path consists of the covariate features, resulting in the action path encoding sequence, the indicator path encoding sequence, and the covariate path encoding sequence.

[0025] The temporal causal attention fusion module performs self-attention calculation with causal mask on the action path encoding sequence, indicator path encoding sequence and covariate path encoding sequence respectively, and performs fusion processing. The causal mask satisfies that when attention calculation is performed at any time step, it only allows association with the current time step and the sequence positions before it, and generates a fused temporal representation sequence of the monitored object.

[0026] The covariate dynamic filtering module calculates the volatility of the covariate path encoding sequence using a sliding time window. The volatility is the difference between the maximum and minimum values ​​of the covariate encoding values ​​within the sliding time window. The volatility is compared with a preset threshold. For covariate path encoding values ​​with volatility exceeding the preset threshold, a masking process is performed to obtain the filtered fused time series representation sequence.

[0027] Monitoring results are generated based on the filtered fused time-series representation sequence. The monitoring results include anomaly scores, anomaly types, and a set of root cause candidate dimensions.

[0028] Optionally, the generation of the candidate disposal action set includes:

[0029] Based on the monitoring results, a set of abnormal monitoring objects is determined, and an abnormal occurrence time window and abnormal type are determined for each abnormal monitoring object;

[0030] Detailed analysis was performed on the abnormal monitoring objects according to the regional identifier, equipment identifier, and advertising space identifier. The change of indicators in each detailed dimension within the time window of the abnormality was calculated, and the scope of the abnormality was determined based on the change of indicators.

[0031] Based on the anomaly type and the scope of its impact, a set of candidate actions is generated. The set of candidate actions includes pausing delivery, limiting delivery, adjusting budget, adjusting bid, and switching delivery targets. For each candidate action, the action target identifier, the action effective time window, and the action parameters are recorded.

[0032] Optionally, obtaining the incremental effect estimation results and the hierarchical effect decomposition results includes:

[0033] A causal forest algorithm is constructed to receive a set of candidate actions and feature data. For each candidate action, a processing variable, an outcome variable, and an evaluation time window are determined. The processing variable is whether the candidate action occurs, and the outcome variable is the change value of the indicator within the evaluation time window.

[0034] A forward estimation tree is generated based on the back estimation structure fusion tree. The forward estimation tree takes feature data and treatment variables as input and outputs the treatment effect estimate of the outcome variable. The treatment effect estimate of the leaf node is calculated at the leaf node. The treatment effect estimate of the leaf node is the difference between the mean of the outcome variable of the sample set in which the treatment variable occurs and the mean of the outcome variable of the sample set in which the treatment variable does not occur.

[0035] A reverse estimation tree is generated by fusing the reverse estimation structure tree. The reverse estimation tree takes feature data and outcome variables as input and outputs the occurrence estimate of the processing variable. Based on the leaf node processing effect estimate of the forward estimation tree and the occurrence estimate of the processing variable of the reverse estimation tree, a consistency judgment result is generated. The leaf node processing effect estimate is fused according to the consistency judgment result to obtain the fused effect estimate result.

[0036] The fusion effect estimation results are decomposed using a hierarchical effect decomposition tree structure. The decomposition process includes generating hierarchical divisions according to channel layer, material layer, and audience layer, calculating hierarchical effect components within each hierarchical division, and the hierarchical effect components being the statistical values ​​of the fusion effect estimation results within the corresponding hierarchical division. Based on the hierarchical effect components, the incremental effect estimation results of candidate disposal actions and the hierarchical effect decomposition results are generated.

[0037] Optionally, determining the target action and generating the action instruction includes:

[0038] For each candidate action in the candidate action set, read the incremental effect estimation results, hierarchical effect decomposition results and monitoring results, and determine the target indicator type and target indicator direction for each candidate action.

[0039] A comprehensive score is calculated for each candidate action based on the type and direction of the target indicator. The comprehensive score consists of the change value of the target indicator corresponding to the incremental effect estimation result and the influence degree value corresponding to the hierarchical effect decomposition result. The change value of the target indicator is the change value of the target indicator within the evaluation time window, and the influence degree value is the hierarchical effect component in the hierarchical effect decomposition result that matches the scope of abnormal influence.

[0040] The candidate action with the best comprehensive score is selected from the set of candidate actions and used as the target action. An action instruction is generated, which includes the action object identifier, action type, action parameters, action effective time window, and action duration.

[0041] An intelligent monitoring system for advertising placement based on big data analysis according to an embodiment of the present invention includes the following modules:

[0042] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data and complete preprocessing to generate unified-caliber monitoring data and multi-dimensional indicator sequences.

[0043] The object construction feature module is used to construct monitoring objects based on monitoring data and multidimensional indicator sequences, and generate feature data;

[0044] An improved causal Transformer analysis module was developed to perform multi-path encoding, causal mask attention fusion, and dynamic covariate filtering on feature data, and output monitoring results.

[0045] The abnormal impact analysis module is used to refine the analysis based on monitoring results, locate the scope of abnormal impact, and generate a set of candidate response actions.

[0046] The causal forest evaluation module is used to evaluate the candidate action set based on forward and backward estimation trees, decompose hierarchical effects, and output incremental effect estimation results and hierarchical effect decomposition results.

[0047] The disposal decision instruction module is used to determine the target disposal actions based on the assessment results and monitoring results, and generate disposal instructions.

[0048] The beneficial effects of this invention are:

[0049] This invention unifies the mapping of event fields, standardizes formats, marks missing data, and removes duplicates and anomalies from multi-source heterogeneous data such as exposure logs, click logs, conversion logs, consumption data, and strategy change data. Combined with link correlation, time alignment, and late arrival compensation, it generates unified monitoring data and multi-dimensional indicator sequences. This improves indicator consistency and traceability at the data level, reducing the impact of differences in caliber, delays, and noise on monitoring results. Compared to monitoring methods relying on a single data source, manual reconciliation, or simple summary statistics, this invention employs multi-path causal modeling, time-series causal attention fusion, and dynamic covariate filtering in the improved causal Transformer model. It adaptively and jointly models multiple indicator time series, campaign action sequences, and covariate features, enabling stable monitoring results under different campaign targets and fluctuation levels. These results include anomaly scores, anomaly types, and root cause candidate dimension sets, thereby improving the timeliness of anomaly detection and reducing false positives and false negatives.

[0050] This invention, after anomaly identification, uses dimensional refinement analysis to locate the scope of anomaly impact and generates a set of candidate action plans. This provides a structured description of the anomaly from its occurrence to its affected objects, facilitating further assessment and execution of actions. Addressing the difficulty in quantifying the effectiveness of actions, this invention utilizes a causal forest algorithm for effectiveness evaluation. It simultaneously constructs forward and backward estimation trees through a back-estimation structure fusion tree, performing consistency checks and fusion processing to improve the stability of effectiveness estimation. Furthermore, it decomposes the fusion effect into channel, material, and audience layers using a hierarchical effect decomposition tree structure, outputting incremental effect estimation results and hierarchical effect decomposition results, thereby supporting differentiated actions and priority ranking. Compared to traditional methods that rely on manual experience for investigation and handling after fixed threshold alarms, this invention can generate executable action instructions and shorten response time, reducing ineffective consumption and budget waste caused by persistent anomalies. It also improves the interpretability and engineering implementation capabilities of action decisions, facilitating closed-loop management in conjunction with deployment platforms. Attached Figure Description

[0051] 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:

[0052] Figure 1 This is a flowchart of an intelligent monitoring method for advertising placement based on big data analysis proposed in this invention;

[0053] Figure 2 This is a structural block diagram of an improved causal Transformer model for an intelligent monitoring method of advertising placement based on big data analysis proposed in this invention.

[0054] Figure 3 This is a functional diagram of an intelligent advertising placement monitoring system based on big data analysis proposed in this invention. Detailed Implementation

[0055] 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.

[0056] refer to Figure 1 and Figure 2 A smart monitoring method for advertising placement based on big data analysis includes:

[0057] Collect multi-source heterogeneous data on advertising placement, preprocess the multi-source heterogeneous data, and generate unified monitoring data and multi-dimensional indicator sequences;

[0058] Based on monitoring data and multidimensional indicator sequences, monitoring objects are constructed, and feature data is generated. The feature data includes multi-indicator time series, deployment action sequences, and covariate features.

[0059] An improved causal Transformer model is constructed, and multi-path causal modeling is introduced to divide the feature data into multiple causal paths, which are then processed by sequence encoding. Temporal causal attention fusion is used to perform causal mask self-attention calculation and fusion. The volatility is calculated by dynamically filtering covariates to obtain the monitoring results.

[0060] Based on the monitoring results, a detailed dimensional analysis of the monitored objects is performed, and a set of candidate actions is generated after locating the scope of the anomaly's impact.

[0061] The causal forest algorithm is used to evaluate the treatment effect of the candidate treatment action set. The inverse estimation structure fusion tree is introduced to construct the forward estimation tree and the inverse estimation tree respectively. The hierarchical effect decomposition tree structure is decomposed to generate hierarchical effect components, and the incremental effect estimation result and hierarchical effect decomposition result are obtained.

[0062] Based on the incremental effect estimation results, hierarchical effect decomposition results, and monitoring results, target action actions are determined and action instructions are generated.

[0063] In this embodiment, the multi-source heterogeneous data includes exposure logs, click logs, conversion logs, consumption data, and strategy change data.

[0064] In this embodiment, generating unified monitoring data and multidimensional indicator sequences includes:

[0065] Acquire multi-source heterogeneous data and map various data records to a unified set of event fields. The unified set of event fields includes event type, event occurrence time, access time, activity identifier, channel identifier, material identifier, user identifier, device identifier, cost field, and conversion value field.

[0066] Preprocessing is performed on the data records corresponding to the unified event field set. The preprocessing includes format standardization, missing value marking, duplicate record identification and removal, and outlier identification and removal. Data records whose event occurrence time is earlier than the start time of the current processing window and whose access time is later than the end time of the current processing window are marked as late data.

[0067] The preprocessed data records undergo association and alignment processing. This includes link association processing based on event identifiers or tracking identifiers, time alignment processing based on event occurrence time, and data recovery processing for late data. Based on the association and alignment processing results, unified-caliber monitoring data and multi-dimensional indicator sequences are generated. Specifically, generating unified-caliber monitoring data and multi-dimensional indicator sequences involves:

[0068] Based on the correlation and alignment processing results, various event records are summarized by event occurrence time according to a fixed time window length. Monitoring data is generated using activity identifier, channel identifier, and material identifier as statistical primary keys. Within each time window, exposure events, click events, and conversion events are deduplicated and counted. Consumption records are summed to obtain exposure, click, conversion, and consumption. Within the same time window, click-through rate is obtained by dividing click by exposure, conversion rate is obtained by dividing conversion by click, and cost per conversion is obtained by dividing consumption by conversion. These are arranged in the order of time windows to form a multidimensional indicator sequence. For late data, it is assigned to the corresponding time window according to the event occurrence time, and incremental updates are performed on the counting and summing results of the time windows, thereby outputting consistent monitoring data and a multidimensional indicator sequence.

[0069] In this embodiment, the generation of feature data includes:

[0070] Based on monitoring data, monitoring objects are constructed. These objects are identified by activity identifiers, channel identifiers, and material identifiers. Monitoring data is then grouped according to these objects. The construction of monitoring objects specifically involves:

[0071] Based on the activity identifier, channel identifier, and material identifier carried in each record of the monitoring data, the three are combined to generate a combination key. After deduplication of the combination key in the monitoring data, a set of monitoring objects is obtained. A unique monitoring object number is assigned to each combination key. The monitoring object number is used as the grouping key. Records in the monitoring data whose activity identifier, channel identifier, and material identifier are consistent with the combination key are grouped into the same group to form the data group of the corresponding monitoring object. Records that are missing any identifier are marked with a preset missing value and removed from the scope of the monitoring object construction.

[0072] For each monitored object, a multi-indicator time series is generated according to the time window. The multi-indicator time series includes exposure series, click series, conversion series, consumption series, and derived indicator series are generated.

[0073] Based on strategy change data, a deployment action sequence is generated. Covariate features are extracted from monitoring data. The multi-indicator time series, deployment action sequence, and covariate features are then aligned by time to form feature data. Specifically, the composition of feature data includes:

[0074] Using the starting time series of each monitored object's time window as the alignment benchmark, strategy change data is mapped to the same time window. Budget adjustment amount and bid adjustment amount are generated in each time window. Budget adjustment amount is the budget after the change minus the budget before the change, and bid adjustment amount is the bid after the change minus the bid before the change. Action flags for pausing and starting campaigns are generated, with one being recorded when it occurs and zero being recorded when it does not occur. Covariate features are extracted from the monitoring data as region identifier, device identifier, and ad placement identifier, and fixed values ​​are assigned to the same monitored object. For each time window, the time series values ​​of multiple indicators, action values, and covariate features corresponding to the current time window are concatenated in order to form a feature vector, which is then arranged in the order of the time window to obtain feature data.

[0075] In this embodiment, obtaining the monitoring results includes:

[0076] An improved causal Transformer model is constructed, including a multi-path causal modeling module, a temporal causal attention fusion module, and a covariate dynamic filtering module, wherein:

[0077] In the sequence encoding, a multi-path causal modeling module is set up to allocate the input feature data to the action path, indicator path, and covariate path according to the action sequence, multi-indicator time series, and covariate features. Embedding and position encoding are performed respectively. In the attention calculation of each path, a time-series causal attention fusion module is set up to perform self-attention calculation with causal masking on the embedded sequence of each path. The outputs of each path are fused to obtain a fused time-series representation. A covariate dynamic filtering module is set up between the fused output and the covariate path to calculate the volatility of the covariate path encoded sequence within a sliding time window. Covariate encoded values ​​exceeding the threshold are masked. The filtering result and the fused time-series representation are input into the output module to generate the monitoring result.

[0078] The multi-path causal modeling module splits the feature data into three causal paths, each undergoing sequence encoding. These three paths include an action path, an indicator path, and a covariate path. The action path consists of a sequence of deployment actions, the indicator path consists of multiple indicator time series, and the covariate path consists of covariate features. This results in the action path encoding sequence, the indicator path encoding sequence, and the covariate path encoding sequence.

[0079] The process of splitting feature data into three causal paths is as follows: For each monitored object, the feature data consists of a sequence of feature vectors arranged in the order of time windows. Each feature vector contains multiple indicator time series fields, campaign action sequence fields, and covariate feature fields in a fixed field order. The feature vectors are sliced ​​according to the category to which the fields belong. Fields containing exposure, clicks, conversions, consumption, and derived indicators are used as indicator path inputs. Fields containing budget adjustment, bid adjustment, paused campaign flag, and enabled campaign flag are used as action path inputs. Fields containing geographic identifier, device identifier, and ad slot identifier are used as covariate path inputs. The slicing is repeated for each time window to obtain the action path sequence, indicator path sequence, and covariate path sequence, respectively. The three sequences correspond in the order of time windows.

[0080] The process of obtaining the action path encoding sequence, indicator path encoding sequence, and covariate path encoding sequence is as follows:

[0081] Sequence encoding is performed on the action path sequence, indicator path sequence, and covariate path sequence respectively. Normalization is performed on the numerical fields. The normalization process is to calculate the maximum and minimum values ​​of each field within 30 consecutive time windows, and then subtract the minimum value from the current value of the field and divide by the difference between the maximum and minimum values ​​to obtain the normalized value. Embedding mapping is performed on the identifier field. The embedding vector dimension is set to 32. Position encoding is superimposed on the vector of each time window of the three paths. The position encoding is a learnable vector corresponding to the time window number.

[0082] The normalized numerical vector and the embedded identifier vector are concatenated in a fixed order within the same time window to obtain the time window encoding vector of the corresponding path. The time window encoding vectors are arranged in the order of the time window to obtain the action path encoding sequence, the indicator path encoding sequence and the covariate path encoding sequence, respectively.

[0083] The temporal causal attention fusion module performs self-attention computation with causal masks on the action path encoding sequence, indicator path encoding sequence, and covariate path encoding sequence, respectively, and then performs fusion processing. The causal mask satisfies that when attention computation is performed at any time step, it only allows association with the current time step and the sequence positions before it, generating a fused temporal representation sequence of the monitored object. The self-attention computation with causal masks specifically involves:

[0084] Self-attention computation units with the same structure are set up for the action path encoding sequence, indicator path encoding sequence and covariate path encoding sequence. Each self-attention computation unit uses four attention heads to map the encoding vector into query vector, key vector and value vector. For any time step, the dot product of the query vector of the time step and the key vectors of all time steps is calculated to obtain the relevance score. The relevance scores corresponding to the key vectors later than the time step are assigned to the minimum value to form a causal mask. The relevance scores after the mask are normalized to obtain the attention weight. The value vectors of the corresponding time steps are weighted and summed using the attention weight to obtain the self-attention output of the current time step.

[0085] The self-attention output sequences of the three paths are obtained by repeating the calculation at all time steps. The self-attention output vectors of the three paths at the same time step are concatenated in order and input into the linear mapping layer to obtain the fused temporal representation sequence.

[0086] The covariate dynamic filtering module calculates the volatility of the covariate path encoding sequence using a sliding time window. The volatility is the difference between the maximum and minimum values ​​of the covariate encoding values ​​within the sliding time window. The volatility is compared with a preset threshold. Covariate path encoding values ​​with volatility exceeding the preset threshold are masked to obtain the filtered fused time series representation sequence. Specifically, the masking process for covariate path encoding values ​​with volatility exceeding the preset threshold is as follows:

[0087] For each time step, the covariate encoding value of the corresponding window is retrieved from the covariate path encoding sequence using a sliding time window of length ten time windows. The difference between the maximum and minimum values ​​of the covariate encoding value of each dimension in the current window is calculated to obtain the volatility. The volatility is compared with a preset threshold of 0.5. When the volatility of any dimension is greater than 0.5, a masking mark of the same dimension is generated. The masking mark is set to zero, otherwise it is set to one. The masking mark is multiplied by the covariate encoding value of the time step dimension by dimension to obtain the masked covariate encoding value.

[0088] The masked covariate encoded values ​​and the fused temporal representation vector of the time step are concatenated in a fixed field order and then input into the linear mapping layer. The output is the filtered fused temporal representation vector of the time step, which is then arranged in the order of the time steps to obtain the filtered fused temporal representation sequence.

[0089] Monitoring results are generated based on the filtered fused time-series representation sequence. These results include anomaly scores, anomaly types, and a root cause candidate dimension set. Specifically, generating the monitoring results based on the filtered fused time-series representation sequence involves:

[0090] The filtered fused temporal representation sequence is input into the output layer. The output layer includes an anomaly score prediction head and an anomaly type classification head. The anomaly score prediction head consists of two fully connected layers and outputs an anomaly score sequence between zero and one. The anomaly type classification head consists of one fully connected layer and a normalization layer and outputs the probability distribution of each anomaly type. At each time step, the time step with an anomaly score greater than 0.7 is selected as the anomaly time step, and the anomaly type with the highest probability in the anomaly time step is selected as the anomaly type.

[0091] Based on the feature contribution of the covariate dimension, action dimension and indicator dimension corresponding to the filtered fused time series representation sequence, a root cause candidate dimension set is generated. The feature contribution is obtained by summing the absolute values ​​of the gradients of the output layer with respect to the vectors corresponding to each dimension. The top five dimensions with the highest contribution are selected as the root cause candidate dimension set to obtain the monitoring results.

[0092] In this embodiment, generating the candidate action set includes:

[0093] The set of abnormal monitoring objects is determined based on the monitoring results, and an abnormal occurrence time window and abnormality type are determined for each abnormal monitoring object. Specifically, the determination of the set of abnormal monitoring objects based on the monitoring results involves:

[0094] For each monitored object, read the abnormal score sequence and abnormal type probability distribution from the monitoring results. Set the number of consecutive trigger windows to three time windows. When the abnormal score of the same monitored object is greater than 0.7 in three consecutive time windows, add the monitored object to the abnormal monitored object set. The start time of the time window that first meets the consecutive trigger condition is determined as the start time of the abnormal occurrence time window, and the start time of the third consecutive trigger time window is determined as the end time of the abnormal occurrence time window.

[0095] Within the time window of an anomaly occurrence, the anomaly type with the highest probability is selected as the anomaly type of the time window, and the anomaly type that occurs most frequently is selected as the anomaly type of the anomaly monitoring object.

[0096] A detailed analysis is performed on the anomaly monitoring targets based on geographical identifiers, device identifiers, and advertising space identifiers. The change in indicators for each detailed dimension within the anomaly occurrence time window is calculated. The scope of the anomaly's impact is determined based on these changes. Specifically, the calculation of the change in indicators for each detailed dimension within the anomaly occurrence time window involves:

[0097] For each abnormal monitoring object, it is grouped according to the detailed dimensions of geographical identifier, device identifier, and advertising space identifier. Within the time window of the abnormality occurrence, the corresponding indicator sequence is extracted according to the detailed dimension. Consumption, conversion volume, and cost per conversion are selected as calculation indicators. At the same time, three consecutive time windows before the time window of the abnormality occurrence are taken as control time windows. The average value of the calculated indicators corresponding to each detailed dimension within the control time window is calculated. The change in indicator for each detailed dimension is calculated. The change in indicator is the average value of the calculated indicators within the time window of the abnormality occurrence minus the average value of the calculated indicators within the control time window. The absolute value of the change in consumption and cost per conversion is taken to obtain the change in indicator for each detailed dimension within the time window of the abnormality occurrence.

[0098] Based on the anomaly type and the scope of its impact, a set of candidate actions is generated. The set of candidate actions includes pausing delivery, limiting delivery, adjusting budget, adjusting bid, and switching delivery targets. For each candidate action, the action target identifier, the action effective time window, and the action parameters are recorded.

[0099] In this embodiment, obtaining the incremental effect estimation result and the hierarchical effect decomposition result includes:

[0100] A causal forest algorithm is constructed, which receives a set of candidate actions and feature data. For each candidate action, a processing variable, an outcome variable, and an evaluation time window are determined. The processing variable is whether the candidate action occurs, and the outcome variable is the change in the indicator value within the evaluation time window. Specifically, determining the processing variable, outcome variable, and evaluation time window for each candidate action involves:

[0101] For each candidate action in the candidate action set, read the action object identifier, action type, action effective time window, and action parameters. Use the feature data sample corresponding to the action object identifier as the evaluation sample set. Define the processing variable as whether the action type has occurred within the action effective time window. If it has occurred, record it as one; if it has not occurred, record it as zero. Whether the action has occurred is determined by the corresponding action record in the strategy change data. Set the evaluation time window to six consecutive time windows after the action effective time window. Define the change value of the target indicator within the evaluation time window as the result variable. The target indicator is the single conversion cost and conversion volume.

[0102] The change value of single conversion cost is the average value of single conversion cost within the evaluation time window minus the average value of single conversion cost within the three consecutive time windows before the action take-off time window. The change value of conversion volume is the average value of conversion volume within the evaluation time window minus the average value of conversion volume within the three consecutive time windows before the action take-off time window. This determines the processing variables, outcome variables, and evaluation time window for each candidate action.

[0103] A forward estimation tree is generated based on the back estimation structure fusion tree. The forward estimation tree takes feature data and treatment variables as input and outputs the treatment effect estimate of the outcome variable. At the leaf nodes, the leaf node treatment effect estimate is calculated. The leaf node treatment effect estimate is the difference between the mean of the outcome variable in the sample set where the treatment variable occurs and the mean of the outcome variable in the sample set where the treatment variable does not occur.

[0104] The process of generating a forward estimation tree is as follows: using the evaluation sample set as the input sample, the input features as the feature fields in the feature data, and the input labels as the processing variables and the result variables, a bootstrap sampling method is used to extract the same number of samples as the total number of samples from the evaluation sample set with replacement as the tree construction samples. The maximum depth of each tree is set to ten layers, the minimum number of samples in each leaf node is fifty, and the number of candidate features for each split is the square root of the total number of features rounded down. Recursively splitting is performed starting from the root node. At each node, the candidate splitting points of the candidate features are traversed, and the splitting point that maximizes the variance increment of the difference between the sample of the processing variable and the result variable of the non-processing variable within the node is selected as the splitting point, until the maximum depth or the minimum number of samples in the leaf node is met, thus obtaining the forward estimation tree.

[0105] The estimated treatment effect of the output variable is specifically as follows: For the sample to be evaluated, the feature data is input into the forward estimation tree, and the corresponding leaf node is reached along the splitting rules. Within the leaf node, the sample set where the treatment variable occurs and the sample set where it does not occur are taken respectively. The mean of the result variable of the two sets is calculated, and the difference is used to obtain the estimated treatment effect of the leaf node. The splitting process is repeated for the same candidate treatment action to generate a preset number of one hundred forward estimation trees. The arithmetic mean of the estimated treatment effect of the leaf node obtained by the sample to be evaluated on the one hundred forward estimation trees is calculated as the estimated treatment effect of the result variable of the sample to be evaluated.

[0106] A reverse estimation tree is generated by fusing the reverse estimation structure. The reverse estimation tree takes feature data and outcome variables as input and outputs the occurrence estimate of the processing variable. Based on the leaf node processing effect estimates of the forward estimation tree and the occurrence estimates of the processing variable from the reverse estimation tree, a consistency determination result is generated. The leaf node processing effect estimates are then fused according to the consistency determination result to obtain the fused effect estimate, where:

[0107] The process of generating the inverse estimation tree is as follows: starting from the root node, recursively splitting, traversing the candidate splitting points of the candidate features at each node, selecting the splitting point that maximizes the decrease in Gini impurity of the processed variables within the node as the splitting point, until the conditions of maximum depth or minimum number of samples in the leaf node are met, thus obtaining the inverse estimation tree. The Gini impurity is the product of the proportions of each category of samples within the node and the value obtained by subtracting from it, which is used to characterize the degree of mixing of sample categories in the node.

[0108] The occurrence estimate of the output processing variable is specifically as follows: for the sample to be evaluated, the feature data and result variables are input into the inverse estimation tree. The splitting process is repeated for the same candidate treatment action to generate one hundred inverse estimation trees. The arithmetic mean of the occurrence estimates of the processing variable obtained by the sample to be evaluated on the one hundred inverse estimation trees is calculated as the occurrence estimate of the processing variable of the sample to be evaluated.

[0109] The process of generating consistency determination results is as follows: For each sample to be evaluated, the estimated value of the treatment effect of the result variable output by the forward estimation tree and the estimated value of the occurrence of the treatment variable output by the backward estimation tree are read. The occurrence estimation threshold is set to 0.6. When the estimated value of the occurrence of the treatment variable is greater than 0.6, the sample is determined to be consistent with the sample where the treatment variable occurs; otherwise, it is determined to be consistent with the sample where the treatment variable does not occur. A consistency determination flag is generated, with a value of 1 indicating consistency and a value of 0 indicating inconsistency. The leaf node treatment effect estimate is retained for samples with a consistency determination flag of 1, and the leaf node treatment effect estimate is set to zero for samples with a consistency determination flag of 0. The arithmetic mean of the leaf node treatment effect estimates of all samples corresponding to the same candidate treatment action is calculated to obtain the fusion effect estimation result.

[0110] The fusion effect estimation results are decomposed using a hierarchical effect decomposition tree structure. This decomposition includes generating hierarchical divisions based on channel, material, and audience layers; calculating hierarchical effect components within each hierarchical division; and generating incremental effect estimation results for candidate actions and hierarchical effect decomposition results based on these hierarchical effect components.

[0111] The decomposition process for the fusion effect estimation results is as follows: taking the set of evaluation samples corresponding to the candidate disposal actions as the decomposition object, the channel identifier, material identifier, and population identifier of each evaluation sample are read respectively, and a three-level hierarchical division is constructed in sequence. The first level is the channel layer, the second level is the material layer, and the third level is the population layer. The evaluation samples are grouped by channel identifier to obtain channel layer groups, and within each channel layer group, they are grouped by material identifier to obtain material layer groups. Within each material layer group, they are grouped by population identifier to obtain population layer groups. The fusion effect estimation results of all samples in the current group are aggregated for each finest-grained group, and the hierarchical effect components are calculated. The hierarchical effect components are the arithmetic mean of the fusion effect estimation results of the current group.

[0112] The incremental effect estimation results and hierarchical effect decomposition results of candidate action generation based on hierarchical effect components are as follows: For the same candidate action, after removing groups with fewer than fifty samples from the set of hierarchical effect components obtained at the finest granular population level, the weighted summation value of the hierarchical effect components of the remaining groups is calculated as the incremental effect estimation result of the candidate action. The weighted summation value is the sum of the products of the hierarchical effect components of each group and the corresponding sample number of the group, and then divided by the sample number. At the same time, the hierarchical effect decomposition results are output. The hierarchical effect decomposition results include the group identifiers of the channel layer, material layer and population layer, the corresponding hierarchical effect components and the sample number of the group. The hierarchical groups are sorted and output in descending order of the absolute value of the hierarchical effect components.

[0113] In this embodiment, determining the target action and generating the action command includes:

[0114] For each candidate action in the candidate action set, read the incremental effect estimation results, hierarchical effect decomposition results and monitoring results, and determine the target indicator type and target indicator direction for each candidate action.

[0115] A comprehensive score is calculated for each candidate action based on the type and direction of the target indicator. The comprehensive score consists of the change value of the target indicator corresponding to the incremental effect estimation result and the influence degree value corresponding to the hierarchical effect decomposition result. The change value of the target indicator is the change value of the target indicator within the evaluation time window, and the influence degree value is the hierarchical effect component in the hierarchical effect decomposition result that matches the scope of abnormal influence.

[0116] The candidate action with the best comprehensive score is selected from the set of candidate actions and used as the target action. An action instruction is generated, which includes the action object identifier, action type, action parameters, action effective time window, and action duration.

[0117] refer to Figure 3 An intelligent monitoring system for advertising placement based on big data analysis includes the following modules:

[0118] The data acquisition and preprocessing module is used to collect multi-source heterogeneous data and complete preprocessing to generate unified-caliber monitoring data and multi-dimensional indicator sequences.

[0119] The object construction feature module is used to construct monitoring objects based on monitoring data and multidimensional indicator sequences, and generate feature data;

[0120] An improved causal Transformer analysis module was developed to perform multi-path encoding, causal mask attention fusion, and dynamic covariate filtering on feature data, and output monitoring results.

[0121] The abnormal impact analysis module is used to refine the analysis based on monitoring results, locate the scope of abnormal impact, and generate a set of candidate response actions.

[0122] The causal forest evaluation module is used to evaluate the candidate action set based on forward and backward estimation trees, decompose hierarchical effects, and output incremental effect estimation results and hierarchical effect decomposition results.

[0123] The disposal decision instruction module is used to determine the target disposal actions based on the assessment results and monitoring results, and generate disposal instructions.

[0124] Example 1:

[0125] To verify the feasibility of this invention in practice, it was applied to a high-concurrency advertising scenario where multiple channels and ad placements were simultaneously covered. The target audience included a large number of creative materials and different audience targeting strategies, and the data chain consisted of multi-source heterogeneous data on impressions, clicks, conversions, spending, and strategy changes. Due to differences in field definitions, deduplication rules, attribution feedback links, and timestamp systems among various data sources, and the delays and late arrivals of conversion feedback and third-party receipts, the spending and conversions of the same target audience did not match on different reporting sides, and indicators experienced sudden jumps and rebounds. Traditional fixed threshold monitoring frequently generated false alarms during periods of high volatility. Furthermore, when real anomalies occurred, manual investigation of channels, creative materials, devices, and ad placements was often required, resulting in long location and handling times and potentially leading to an increase in ineffective spending.

[0126] After multi-source heterogeneous data is connected to the system of this invention, various records are uniformly mapped to a set of event fields, and standardization, missing data marking, duplicate and anomaly removal are completed. Link association and time alignment are performed based on tracking identifiers and event occurrence time. Late feedback is filled back to the corresponding time window according to the event occurrence time and incremental updates are performed, outputting unified monitoring data and multi-dimensional indicator sequences. Monitoring objects are constructed using activity identifiers, channel identifiers, and material identifiers to generate feature data. This feature data is obtained by aligning and stitching multi-indicator time series, delivery action sequences, and covariate features according to time windows. An improved causal Transformer model is used to perform multi-path causal modeling and encoding of the feature data. A time-series causal attention fusion is used to perform causal masking self-attention and fusion representation. Covariate dynamic filtering is used to mask covariates with volatility exceeding the threshold, and monitoring results are output. Based on the monitoring results, abnormal objects are analyzed in detail to locate the scope of abnormal impact and generate a set of candidate actions. The causal forest algorithm is used to evaluate the effectiveness of the candidate actions. The back estimation structure is used to fuse the tree to obtain both positive and back estimates and perform consistency fusion. The hierarchical effect decomposition tree structure is used to decompose the effect components according to the channel layer, material layer, and audience layer, outputting incremental effect estimation results and hierarchical decomposition results, and generating disposal instructions to be issued to the delivery execution end.

[0127] In real-world operation across multiple campaign cycles, the system reliably handles high-throughput campaign logs and feedback data, uniformly aligning and correcting metric fluctuations caused by delayed feedback, inconsistent data definitions, and noise interference, significantly improving the consistency of window-level monitoring metrics. For events such as sudden increases in spending, sharp drops in conversions, abnormal feedback, and traffic quality anomalies, the system can output alerts and anomaly types earlier, narrowing the impact of anomalies to fewer regions, devices, or ad placement combinations. It also provides a ranking of incremental effects for candidate actions such as pausing campaigns, limiting spending, adjusting budgets, and adjusting bids. After executing recommended actions, the duration of anomalies is shortened, ineffective spending decreases, and the cost of manual investigation and handling is reduced. Furthermore, the handling conclusions have a hierarchical effect breakdown basis, facilitating review and strategy solidification.

[0128] Table 1 Summary of Comparison Indicators for Intelligent Monitoring Solutions for Advertising Placement

[0129] Comparison indicators Fixed threshold Year-on-year and month-on-month comparison rules Rules + Manual LSTM alarm Transformer warning Method of the present invention Alarm accuracy rate (%) 62.4 66.8 74.2 78.5 81.3 89.7 Alarm recall rate (%) 58.1 61.6 69.8 73.4 76.2 87.1 Discovery delay (min) 38.0 31.0 27.0 18.0 15.0 7.0 Root cause accuracy (%) 33.5 37.2 52.6 58.9 62.1 78.4 Processing time (min) 95.0 82.0 64.0 52.0 45.0 26.0 Ineffective consumption reduction (%) 4.8 6.1 9.7 12.5 14.3 22.6 Increase in revenue from advertising (%) 2.3 3.1 4.9 6.0 6.8 10.7 False alarm rate (%) 21.6 18.9 13.4 10.8 9.6 5.2

[0130] As shown in Table 1, in terms of alarm effectiveness, the method of this invention achieves the best performance in both alarm accuracy and recall, with an accuracy of 89.7% and a recall of 87.1%. This significantly outperforms traditional methods such as fixed thresholds, year-on-year / month-on-month comparison rules, and rule-based + manual methods. Compared to deep learning baselines, the method of this invention maintains its leading position in both accuracy and recall, and achieves a root cause accuracy of 78.4%. It can more stably output key dimensions related to the anomaly after alarm triggering, providing a reliable basis for impact range localization and action generation. This invention not only improves the reliability of anomaly identification but also enhances the interpretability of anomaly types and sources, upgrading alarms from simply indicating anomalies to providing locating anomalies.

[0131] In terms of response efficiency, the method of this invention performs best in terms of detection latency and handling time. The detection latency is 7.0 minutes, which can complete scoring, type determination, and root cause candidate output in a short time after the anomaly occurs, thereby preventing the anomaly from continuing to spread during the high-consumption stage. The handling time is 26.0 minutes, which shows that the link from alarm to the formation of handling instructions is shorter and the process is more closed-loop. Compared with the traditional method that relies on manual layer-by-layer investigation, this invention can quickly converge the scope of anomaly impact through dimensional refinement analysis driven by monitoring results, and then complete the selection of candidate actions through the handling effect evaluation module. This changes the response process from human problem finding to the system providing scope, suggestions, and instructions, which significantly reduces the time loss in the investigation and decision-making stages, and also reduces the cumulative risk caused by delayed handling.

[0132] From the perspective of business revenue and risk control, the method of this invention achieves optimal results in both reducing ineffective spending and improving campaign revenue. Ineffective spending is reduced by 22.6%, and campaign revenue is increased by 10.7%, indicating that it not only stops losses faster but also avoids revenue loss due to excessive contraction during the recovery period. The false positive rate of this invention is 5.2%, the lowest among the compared methods, indicating that it improves recall without significantly increasing false positives, reducing the interference of frequent alarms on operations personnel and lowering the risk of unnecessary manual review and mishandling. Overall, this invention demonstrates consistent advantages in monitoring accuracy, response efficiency, revenue improvement, and risk control, making it more suitable for long-term stable operation in advertising scenarios characterized by high volatility, strong heterogeneity, and high time sensitivity.

[0133] 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 intelligent monitoring of advertising placement based on big data analysis, characterized in that, include: Collect multi-source heterogeneous data on advertising placement, preprocess the multi-source heterogeneous data, and generate unified monitoring data and multi-dimensional indicator sequences; Based on monitoring data and multidimensional indicator sequences, monitoring objects are constructed, and feature data is generated. The feature data includes multi-indicator time series, deployment action sequences, and covariate features. An improved causal Transformer model is constructed, and multi-path causal modeling is introduced to divide the feature data into multiple causal paths, which are then processed by sequence encoding. Temporal causal attention fusion is used to perform causal mask self-attention calculation and fusion. The volatility is calculated by dynamically filtering covariates to obtain the monitoring results. Based on the monitoring results, a detailed dimensional analysis of the monitored objects is performed, and a set of candidate actions is generated after locating the scope of the anomaly's impact. The causal forest algorithm is used to evaluate the treatment effect of the candidate treatment action set. The inverse estimation structure fusion tree is introduced to construct the forward estimation tree and the inverse estimation tree respectively. The hierarchical effect decomposition tree structure is decomposed to generate hierarchical effect components, and the incremental effect estimation result and hierarchical effect decomposition result are obtained. Based on the incremental effect estimation results, hierarchical effect decomposition results, and monitoring results, target action actions are determined and action instructions are generated.

2. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The multi-source heterogeneous data includes exposure logs, click logs, conversion logs, consumption data, and strategy change data.

3. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The generation of unified-caliber monitoring data and multi-dimensional indicator sequences includes: Acquire multi-source heterogeneous data and map various data records to a unified set of event fields. The unified set of event fields includes event type, event occurrence time, access time, activity identifier, channel identifier, material identifier, user identifier, device identifier, cost field, and conversion value field. Preprocessing is performed on the data records corresponding to the unified event field set. The preprocessing includes format standardization, missing value marking, duplicate record identification and removal, and outlier identification and removal. Data records whose event occurrence time is earlier than the start time of the current processing window and whose access time is later than the end time of the current processing window are marked as late data. The preprocessed data records are subjected to association and alignment processing, which includes link association processing based on event identifiers or tracking identifiers, time alignment processing based on event occurrence time, and recovery processing for late data. Based on the results of association and alignment processing, monitoring data with unified caliber and multi-dimensional indicator sequences are generated.

4. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The generated feature data includes: Based on the monitoring data, monitoring objects are constructed. The monitoring objects are determined by activity identifiers, channel identifiers, and material identifiers. The monitoring data is then grouped according to the monitoring objects. For each monitored object, a multi-indicator time series is generated according to the time window. The multi-indicator time series includes exposure series, click series, conversion series, consumption series, and derived indicator series are generated. The deployment action sequence is generated based on the strategy change data. Covariate features are extracted from the monitoring data. The multi-indicator time series, deployment action sequence and covariate features are aligned by time to form feature data.

5. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The obtained monitoring results include: An improved causal Transformer model is constructed, including a multi-path causal modeling module, a temporal causal attention fusion module, and a covariate dynamic filtering module; The multi-path causal modeling module splits the feature data into three causal paths and performs sequence encoding processing on each of them. The three causal paths include the action path, the indicator path, and the covariate path. The action path consists of the sequence of delivery actions, the indicator path consists of the time series of multiple indicators, and the covariate path consists of the covariate features, resulting in the action path encoding sequence, the indicator path encoding sequence, and the covariate path encoding sequence. The temporal causal attention fusion module performs self-attention calculation with causal mask on the action path encoding sequence, indicator path encoding sequence and covariate path encoding sequence respectively, and performs fusion processing. The causal mask satisfies that when attention calculation is performed at any time step, it only allows association with the current time step and the sequence positions before it, and generates a fused temporal representation sequence of the monitored object. The covariate dynamic filtering module calculates the volatility of the covariate path encoding sequence using a sliding time window. The volatility is the difference between the maximum and minimum values ​​of the covariate encoding values ​​within the sliding time window. The volatility is compared with a preset threshold. For covariate path encoding values ​​with volatility exceeding the preset threshold, a masking process is performed to obtain the filtered fused time series representation sequence. Monitoring results are generated based on the filtered fused time-series representation sequence. The monitoring results include anomaly scores, anomaly types, and a set of root cause candidate dimensions.

6. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The generated candidate action set includes: Based on the monitoring results, a set of abnormal monitoring objects is determined, and an abnormal occurrence time window and abnormal type are determined for each abnormal monitoring object; Detailed analysis was performed on the abnormal monitoring objects according to the regional identifier, equipment identifier, and advertising space identifier. The change of indicators in each detailed dimension within the time window of the abnormality was calculated, and the scope of the abnormality was determined based on the change of indicators. Based on the anomaly type and the scope of its impact, a set of candidate actions is generated. The set of candidate actions includes pausing delivery, limiting delivery, adjusting budget, adjusting bid, and switching delivery targets. For each candidate action, the action target identifier, the action effective time window, and the action parameters are recorded.

7. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The obtained incremental effect estimation results and hierarchical effect decomposition results include: A causal forest algorithm is constructed to receive a set of candidate actions and feature data. For each candidate action, a processing variable, an outcome variable, and an evaluation time window are determined. The processing variable is whether the candidate action occurs, and the outcome variable is the change value of the indicator within the evaluation time window. A forward estimation tree is generated based on the back estimation structure fusion tree. The forward estimation tree takes feature data and treatment variables as input and outputs the treatment effect estimate of the outcome variable. The treatment effect estimate of the leaf node is calculated at the leaf node. The treatment effect estimate of the leaf node is the difference between the mean of the outcome variable of the sample set in which the treatment variable occurs and the mean of the outcome variable of the sample set in which the treatment variable does not occur. A reverse estimation tree is generated by fusing the reverse estimation structure tree. The reverse estimation tree takes feature data and outcome variables as input and outputs the occurrence estimate of the processing variable. Based on the leaf node processing effect estimate of the forward estimation tree and the occurrence estimate of the processing variable of the reverse estimation tree, a consistency judgment result is generated. The leaf node processing effect estimate is fused according to the consistency judgment result to obtain the fused effect estimate result. The fusion effect estimation results are decomposed using a hierarchical effect decomposition tree structure. The decomposition process includes generating hierarchical divisions according to channel layer, material layer, and audience layer, calculating hierarchical effect components within each hierarchical division, and the hierarchical effect components being the statistical values ​​of the fusion effect estimation results within the corresponding hierarchical division. Based on the hierarchical effect components, the incremental effect estimation results of candidate disposal actions and the hierarchical effect decomposition results are generated.

8. The intelligent monitoring method for advertising placement based on big data analysis according to claim 1, characterized in that, The process of determining the target action and generating the action instruction includes: For each candidate action in the candidate action set, read the incremental effect estimation results, hierarchical effect decomposition results and monitoring results, and determine the target indicator type and target indicator direction for each candidate action. A comprehensive score is calculated for each candidate action based on the type and direction of the target indicator. The comprehensive score consists of the change value of the target indicator corresponding to the incremental effect estimation result and the influence degree value corresponding to the hierarchical effect decomposition result. The change value of the target indicator is the change value of the target indicator within the evaluation time window, and the influence degree value is the hierarchical effect component in the hierarchical effect decomposition result that matches the scope of abnormal influence. The candidate action with the best overall score is selected from the set of candidate actions and used as the target action. An action instruction is generated, which includes the action object identifier, action type, action parameters, action effective time window, and action duration.

9. The intelligent advertising monitoring system based on big data analysis according to claim 1, performing the intelligent advertising monitoring method based on big data analysis according to any one of claims 1 to 8, characterized in that, Includes the following modules: The data acquisition and preprocessing module is used to collect multi-source heterogeneous data and complete preprocessing to generate unified-caliber monitoring data and multi-dimensional indicator sequences. The object construction feature module is used to construct monitoring objects based on monitoring data and multidimensional indicator sequences, and generate feature data; An improved causal Transformer analysis module was developed to perform multi-path encoding, causal mask attention fusion, and dynamic covariate filtering on feature data, and output monitoring results. The abnormal impact analysis module is used to refine the analysis based on monitoring results, locate the scope of abnormal impact, and generate a set of candidate response actions. The causal forest assessment module is used to perform a fusion assessment of the candidate action set based on forward and backward estimation trees, decompose hierarchical effects, and output incremental effect estimation results and hierarchical effect decomposition results. The disposal decision instruction module is used to determine the target disposal actions based on the assessment results and monitoring results, and generate disposal instructions.

Images