A multi-channel data monitoring, early warning and attribution method based on a time series prediction model

Abstract

The application discloses a kind of multi-channel data monitoring early warning and attribution method based on timing prediction model, it is related to data processing and analysis technical field, including the following steps: based on multidimensional Hawkes self-activation point process, the marketing event stream received from multiple marketing channels is jointly modeled, constructs the event arrival intensity function by the basic arrival intensity component and cross-channel mutual excitation intensity component superposition for each marketing channel, when mutual excitation contribution rate is lower than dynamic early warning threshold, generate marketing effect abnormal early warning signal;In response to early warning signal, based on synthetic control method, target marketing channel is specified as to-be-analyzed processing unit, and actual marketing conversion rate time sequence is subtracted every moment to obtain causal effect sequence as effect deviation attribution result output.The application realizes the complete early warning link from cross-channel synergistic effect anomaly detection to strategy change causal attribution.

Description

Technical Field

[0001] This invention relates to the field of data processing and analysis technology, and in particular to a multi-channel data monitoring, early warning and attribution method based on a time series prediction model. Background Technology

[0002] Currently, the industry has accumulated some technical expertise in marketing performance monitoring. A common practice is to set fixed alarm thresholds for key performance indicators (KPIs) across various channels, triggering alarm notifications when KPI values ​​exceed preset ranges. While this fixed-threshold-based monitoring method is simple to implement, it has significant limitations. The pace and intensity of marketing campaigns continuously change with business cycles, holidays, promotional activities, and other factors. Fixed thresholds cannot adapt to the dynamic changes in data distribution, easily generating numerous false alarms during peak business periods and missing genuine anomalies during off-peak periods. This results in the accuracy and practicality of alerts failing to meet the requirements of refined operations.

[0003] In time-series forecasting, existing technologies widely employ methods such as autoregressive moving average models, long short-term memory networks, Transformers and their variants to predict marketing metrics, and use the deviation between predicted and actual values ​​to determine whether anomalies have occurred. These methods treat the time series of metrics for each channel as independent forecast objects, essentially fitting and extrapolating the numerical trends of a single channel. However, in multi-channel marketing scenarios, there are complex cascading event triggering relationships between channels. For example, a promotional activity launched on one channel may trigger an increase in user clicks and conversions on other channels within hours. This dynamic synergistic effect between channels cannot be captured by time-series forecasting of aggregated metrics. Ignoring the mutually stimulating structure between channels makes it difficult for existing time-series forecasting methods to provide effective early warnings in the early stages of abnormal decay of cross-channel synergistic effects.

[0004] In anomaly detection, commonly used methods in existing technologies include control chart methods based on statistical process control, unsupervised anomaly detection methods based on isolated forests, and deviation detection methods based on sliding window statistics. Most of these methods require prior assumptions about the overall distribution of the data, such as assuming the data follows a normal distribution or a known family of parameters. In actual marketing data, due to the non-stationarity and suddenness of marketing activities, the tail characteristics of the data distribution drift over time. Anomaly detection methods based on fixed distribution assumptions are prone to performance degradation due to distribution mismatch. Furthermore, existing methods typically rely on fitting the overall distribution when determining alarm boundaries, resulting in insufficient modeling accuracy for extreme tail events, while anomaly warnings for marketing performance precisely require high sensitivity to extreme deviation events. Summary of the Invention

[0005] In view of this, the present invention provides a multi-channel data monitoring, early warning and attribution method based on a time series prediction model, which can effectively separate the true causal impact of strategy changes from the common fluctuations of the market environment without the need for a pre-designed control group, and provide quantifiable and traceable attribution basis for marketing operation decisions.

[0006] The technical solution adopted in this invention is as follows:

[0007] A multi-channel data monitoring, early warning, and attribution method based on a time-series prediction model includes the following steps:

[0008] Step 1: Based on the multidimensional Hawkes self-excitation point process, jointly model the marketing event flow received from multiple marketing channels. Construct an event arrival intensity function for each marketing channel, which is composed of the basic arrival intensity component and the cross-channel mutual excitation intensity component. Estimate the parameters of the multidimensional Hawkes self-excitation point process. At each sampling time, perform intensity component decomposition for each marketing channel. Determine the mutual excitation contribution rate according to the proportion of the cross-channel mutual excitation intensity component in the total value of the event arrival intensity function. Arrange the mutual excitation contribution rates at each sampling time in chronological order to generate a time series of mutual excitation contribution rates for each marketing channel.

[0009] Step 2: The SPOT streaming extreme value detection method is used to process the time series of mutual stimulation contribution rates of each marketing channel. For data points below the screening threshold, a generalized Pareto distribution is fitted to the lower limit deviation. The tail distribution parameters of the generalized Pareto distribution are dynamically updated as new data points arrive. Based on the updated generalized Pareto distribution, a dynamic warning threshold is determined at a preset false alarm rate level. When the mutual stimulation contribution rate data points are lower than the dynamic warning threshold, an abnormal marketing effect warning signal carrying the target marketing channel identifier and the warning trigger time is generated.

[0010] Step 3: In response to the marketing performance abnormality warning signal, the target marketing channel is designated as the unit to be analyzed and processed based on the synthetic control method. The remaining marketing channels that have not triggered the marketing performance abnormality warning signal are formed into a donor pool. The time series of marketing conversion rates of each channel in the donor pool are used to solve the synthetic coefficient and construct the counterfactual control sequence of the unit to be analyzed and processed. The actual marketing conversion rate time series of the unit to be analyzed and processed after the warning trigger time is compared with the counterfactual control sequence time by time to obtain the causal effect sequence and output it as the effect deviation attribution result.

[0011] Furthermore, each marketing event in the marketing event stream carries the event occurrence time and the channel identifier; the basic arrival intensity component is the constant rate at which each marketing channel spontaneously generates marketing events; the cross-channel mutual stimulation intensity component is the sum of the stimulation effects exerted by each marketing event that has occurred in other channels on each marketing channel, and the stimulation effect of a single marketing event continues to decrease over time in an exponential decay manner from the moment the marketing event occurs.

[0012] Furthermore, the expected maximization algorithm is used to estimate the parameters of the multidimensional Hawkes self-excitation point process. In each iteration, the expected phase and the maximization phase are executed. In the expected phase, for each marketing event in the marketing event stream, the posterior attribution probability of each marketing event triggered by the stimuli effect of each source channel is calculated. The posterior attribution probability is determined based on the share of the stimuli effect contributed by each source channel to the channel to which the current marketing event belongs at the time of the current marketing event.

[0013] Furthermore, in the maximization phase, the basic arrival intensity components of each channel and the excitation decay rate between each channel pair are re-estimated based on the posterior attribution probability. The updated parameters are then substituted into the expectation phase of the next iteration until the parameter change between two consecutive iterations is less than the preset convergence threshold, at which point the iteration stops.

[0014] Furthermore, the intensity component decomposition includes: at each sampling time, calculating the cross-channel mutual stimulation intensity component of the current marketing channel and the total value of the event arrival intensity function of the current marketing channel respectively, and dividing the cross-channel mutual stimulation intensity component by the total value of the event arrival intensity function to obtain the mutual stimulation contribution rate of the current marketing channel at the current sampling time.

[0015] Furthermore, the SPOT streaming extreme value detection method includes an initialization phase: a preset length initialization window is extracted from the beginning segment of the mutual excitation contribution rate time series; all mutual excitation contribution rate data points below the preset initial screening threshold are selected within the initialization window; the difference between each selected mutual excitation contribution rate data point and the preset initial screening threshold is recorded as the lower limit deviation of the corresponding data point; a generalized Pareto distribution is fitted using the lower limit deviations of all selected mutual excitation contribution rate data points to obtain the initial tail shape parameter value and tail scale parameter value.

[0016] Furthermore, after the initialization window ends, the point-by-point streaming detection phase begins. Upon receiving each newly arrived mutual-stimulation contribution rate data point, the following steps are performed: First, determine if the newly arrived mutual-stimulation contribution rate data point is below the current filtering threshold. If it is, calculate the lower limit deviation of the newly arrived mutual-stimulation contribution rate data point and include it in the existing lower limit deviation set. Then, refit the generalized Pareto distribution to update the tail shape parameter and tail scale parameter values. Based on the updated generalized Pareto distribution, determine the dynamic warning threshold at the current time, under the warning false alarm rate level. Next, determine if the newly arrived mutual-stimulation contribution rate data point is below the dynamic warning threshold. If it is, determine that the cross-channel mutual-stimulation effect of the corresponding marketing channel has abnormally decayed, and generate a marketing effect abnormality warning signal carrying the target marketing channel identifier and the warning trigger time.

[0017] Furthermore, in step 3, a preset matching interval is set before the warning trigger time, and the marketing conversion rate time series of each channel in the unit to be analyzed and the donor pool are extracted within the preset matching interval for solving the synthesis coefficient.

[0018] Furthermore, the composite coefficients are solved by constrained quadratic programming. The constrained quadratic programming requires that all composite coefficients are non-negative and that the sum of all composite coefficients equals 1. The optimization objective is to minimize the sum of squared deviations between the composite sequence obtained by linearly combining the marketing conversion rate time series of each channel in the donor pool according to the composite coefficients and the marketing conversion rate time series of the unit to be analyzed within the preset matching interval.

[0019] Furthermore, the obtained composite coefficients are applied to the marketing conversion rate time series of each channel in the donor pool after the warning trigger time, and a linear combination is performed to obtain the counterfactual control sequence. The counterfactual control sequence represents the expected marketing conversion rate trend of the target marketing channel under the assumption that no strategy change has occurred.

[0020] By employing the above technical solutions, this invention achieves the following beneficial effects: This invention uses a multi-dimensional Hawkes self-excitation point process to jointly model multi-channel marketing event flows, enabling the depiction of dynamic excitation and transmission relationships between various marketing channels at the event granularity level, and explicitly incorporating the cross-channel synergistic effect, which is often overlooked in traditional methods, into the modeling framework. Through iterative inference of the implicit event attribution relationships using the expectation-maximization algorithm, accurate estimation of the basic arrival intensity components of each channel and the excitation decay rate between channel pairs is achieved. Based on this, the intensity component decomposition decomposes the event arrival rate of each channel at each moment into a spontaneously generated part and a cross-channel excitation part. The resulting mutual excitation contribution rate time series provides a novel monitoring perspective, capable of capturing abnormal signals in the early stages of the decline in cross-channel synergistic effects, rather than triggering alarms only after a significant decline in marketing conversion rates, thus providing a more ample response window for operational decisions.

[0021] This invention employs the SPOT streaming extreme value detection method in the early warning stage. Based on the generalized Pareto distribution in extreme value theory, it models the tail extreme values ​​of the mutual stimulation contribution rate time series. It eliminates the need for prior assumptions about the overall data distribution; the alarm boundary is automatically determined solely based on the statistical characteristics of the tail data. As new data points continuously flow in, the tail distribution parameters of the generalized Pareto distribution are dynamically updated, allowing the early warning threshold to adaptively track the slow drift of the data distribution characteristics. This avoids the problem of increased false alarm or false negative rates caused by distribution mismatch in non-stationary marketing data under fixed threshold conditions. It achieves high-sensitivity detection of abnormal decay events while maintaining a preset false alarm rate level.

[0022] This invention employs a synthetic control method in the attribution phase. It utilizes the remaining marketing channels that have not triggered warnings to form a donor pool. By solving for the synthetic coefficients through constrained quadratic programming, a counterfactual control sequence is constructed for the target marketing channels under the assumption that no strategy change has occurred. This method eliminates the need for pre-designing a control group before a strategy change, meeting the practical needs of post-hoc attribution. Furthermore, non-negativity and normalization constraints ensure that the counterfactual control sequence always falls within the convex hull of the values ​​taken by each channel in the donor pool, guaranteeing the economic interpretability of the attribution conclusions. The causal effect sequence obtained by subtracting the actual marketing conversion rate time series from the counterfactual control sequence at each time step effectively separates the true causal impact of strategy changes from the common fluctuations of the external market environment, providing operational decision-makers with quantifiable and traceable attribution evidence for effect deviations. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the component decomposition of the event arrival intensity function of marketing channel A in an embodiment of the present invention;

[0024] Figure 2This is a schematic diagram of the dynamic early warning process for SPOT streaming extreme value detection in an embodiment of the present invention;

[0025] Figure 3 This is a schematic diagram of the tail fitting of the generalized Pareto distribution of the lower-side over-limit deviation in an embodiment of the present invention;

[0026] Figure 4 This is a schematic diagram of the construction of a counterfactual control sequence using the synthetic control method in an embodiment of the present invention. Detailed Implementation

[0027] A multi-channel data monitoring, early warning, and attribution method based on a time-series prediction model includes the following steps:

[0028] Step 1: Based on the multidimensional Hawkes self-excitation point process, jointly model the marketing event flow received from multiple marketing channels. Construct an event arrival intensity function for each marketing channel, which is composed of the basic arrival intensity component and the cross-channel mutual excitation intensity component. Estimate the parameters of the multidimensional Hawkes self-excitation point process. At each sampling time, perform intensity component decomposition for each marketing channel. Determine the mutual excitation contribution rate according to the proportion of the cross-channel mutual excitation intensity component in the total value of the event arrival intensity function. Arrange the mutual excitation contribution rates at each sampling time in chronological order to generate a time series of mutual excitation contribution rates for each marketing channel.

[0029] Step 2: The SPOT streaming extreme value detection method is used to process the time series of mutual stimulation contribution rates of each marketing channel. For data points below the screening threshold, a generalized Pareto distribution is fitted to the lower limit deviation. The tail distribution parameters of the generalized Pareto distribution are dynamically updated as new data points arrive. Based on the updated generalized Pareto distribution, a dynamic warning threshold is determined at a preset false alarm rate level. When the mutual stimulation contribution rate data points are lower than the dynamic warning threshold, an abnormal marketing effect warning signal carrying the target marketing channel identifier and the warning trigger time is generated.

[0030] Step 3: In response to the marketing performance abnormality warning signal, the target marketing channel is designated as the unit to be analyzed and processed based on the synthetic control method. The remaining marketing channels that have not triggered the marketing performance abnormality warning signal are formed into a donor pool. The time series of marketing conversion rates of each channel in the donor pool are used to solve the synthetic coefficient and construct the counterfactual control sequence of the unit to be analyzed and processed. The actual marketing conversion rate time series of the unit to be analyzed and processed after the warning trigger time is compared with the counterfactual control sequence time by time to obtain the causal effect sequence and output it as the effect deviation attribution result.

[0031] In real-world multi-channel marketing operations, marketing events across different channels often exhibit mutually reinforcing relationships. For example, a promotional activity launched on one channel may temporarily boost user clicks and conversions on other channels. This cross-channel chain reaction is not an independent random process, but rather a temporary increase in the reach of events on other channels after an event occurs on one channel. If we model the event reach of each channel as an independent Poisson process, we cannot capture this dynamic coupling relationship between channels, and therefore cannot issue timely warnings when the synergistic effect between channels abnormally declines. The multidimensional Hawkes self-excitation point process can precisely characterize this dynamic transmission mechanism of "an event on one channel triggering a subsequent event on another channel." Therefore, this method chooses the multidimensional Hawkes self-excitation point process as the joint modeling framework for marketing event flows.

[0032] Assume the system has a total of One marketing channel are positive integers and In a specific application scenario The typical value range is 3 to 20. The system uses this... A continuous stream of marketing events is received from each marketing channel. Each marketing event carries two pieces of information: the time the event occurred and the channel identifier. For the first... One marketing channel Take 1 to Integers between these ranges are used to construct an event arrival intensity function based on a multidimensional Hawkes self-excitation point process. ,in Indicates the current moment. Event arrival intensity function. The physical meaning is: at time Within a very short period of time, the first The instantaneous rate at which a new marketing event occurs within a marketing channel. Event arrival intensity function. It is composed of the superposition of the basic arrival intensity component and the cross-channel mutual excitation intensity component, and is expressed as follows: In the above formula, For the first The basic reach intensity component of a marketing channel represents the reach of the first channel under the condition that no other channel events trigger an impact. The constant rate at which marketing events are spontaneously generated by each marketing channel It is a positive real number, and its dimension is the number of events per unit time. Indicates the first Each marketing channel at all times The first one that has already happened The moment a marketing event occurs. For the first The marketing channel for the first The activation amplitude coefficient of each marketing channel It is a non-negative real number, reflecting the first... Each channel instantly boosts its ranking after a marketing event. The initial magnitude of event reach rate for each channel. For the first The marketing channel for the first The excitation decay rate of each marketing channel It is a positive real number, reflecting the rate at which the stimulating effect decays over time. The term represents the exponentially decaying effect of a single marketing event from the moment the event occurs. The initial exponential decay continues to decrease over time. The double summation in the above formula covers the entirety. Each channel and each channel at any time All previous marketing events constitute the intensity of cross-channel synergy.

[0033] The exponential decay form is chosen to characterize the diminishing stimulative effect of a single marketing event because, in marketing practice, the cross-channel chain conversion behavior triggered by an activity launched on one channel is usually most concentrated in the first few hours, followed by a rapid decline. The shape of the exponential function closely matches this empirical pattern, and its analytical integrability facilitates the calculation of the log-likelihood function in subsequent parameter estimation. In an alternative implementation, a power-law decay form can be used instead of the exponential decay form to characterize the diminishing stimulative effect. Power-law decay offers more flexibility in fitting tail behavior in scenarios with longer time spans, but it correspondingly increases the computational complexity of parameter estimation.

[0034] All parameters to be estimated for the multidimensional Hawkes self-excitation point process include: Each basic arrival intensity component ,as well as Each excitation amplitude coefficient and Excitation decay rate The estimation of the above parameters is accomplished using the Expectation-Maximization (EM) algorithm. The reason for choosing the EEM algorithm instead of directly maximizing the log-likelihood function is that in the multidimensional Hawkes self-excitation point process, the occurrence of each marketing event may be jointly contributed by the stimulating effects of multiple source channels, and the contribution proportion of each source channel is a latent variable that cannot be directly observed. The EEM algorithm, by introducing posterior estimation of the latent variables in the expectation phase, transforms the original optimization problem containing latent variables into a standard maximum likelihood problem under a series of complete data, ensuring that the log-likelihood function remains monotonically constant in each iteration.

[0035] The expectation-maximization algorithm performs the following two phases in each iteration.

[0036] In the expectation phase, for each marketing event in the marketing event stream, the posterior attribution probability of each marketing event being triggered by the stimuli from each source channel is calculated. Specifically, for the ... Each marketing channel at all times The first occurrence The first marketing event needs to determine whether it was initiated by the [number]th [organization / organization]. The basic reach intensity of each channel is generated spontaneously, or is it determined by a specific source channel? A previous marketing event The posterior attribution probability is determined by the share of the stimulating effect contributed by each source channel to the channel to which the current marketing event belongs at the time of the marketing event. The first channel The marketing event was attributed to the first... The first channel The posterior attribution probability triggered by a previous marketing event is denoted as . The calculation method is to take the first... The first channel A previous marketing event at the time For the first The instantaneous excitation intensity applied by the first channel, divided by the first... Each channel at all times The total value of the event arrival intensity function Similarly, the first The marketing event was attributed to the first... The posterior attribution probability spontaneously generated by the basic arrival intensity component of each channel is: Divide by The sum of all posterior attribution probabilities equals 1, ensuring the normalization of probabilities.

[0037] In the maximization phase, the posterior attribution probabilities obtained in the expectation phase are used to re-estimate the base reach intensity components of each channel and the excitation decay rate between channel pairs using maximum likelihood estimation. Intuitively, in the maximization phase, each marketing event is no longer entirely attributed to a single source, but rather distributed among the sources using a "soft allocation" method based on the posterior attribution probability. Based on this, the base reach intensity components... The updated value is equal to the value assigned to the first The sum of the posterior attribution probabilities spontaneously generated by each channel, divided by the total length of the observation time interval; excitation amplitude coefficient. and excitation decay rate The updated value is assigned to the channel pair The excitation time interval of all soft-assignment events is obtained by maximum likelihood estimation.

[0038] The updated parameters from the maximization phase are substituted into the expected phase of the next iteration, and this process is repeated alternately until the parameter change between two consecutive iterations is less than a preset convergence threshold, at which point the iteration stops. A typical value for the preset convergence threshold is... arrive Between 30 and 80 iterations. On a test dataset containing 5 marketing channels and a total of approximately 50,000 marketing events, the Expectation-Maximization algorithm typically converges.

[0039] After parameter estimation, intensity component decomposition is performed for each marketing channel at each sampling time. The sampling time is set as follows: points are taken at fixed intervals within the time range covered by the marketing event stream. Typical sampling intervals are 1 minute, 5 minutes, or 15 minutes, which can be flexibly adjusted according to the density of the marketing event stream. At each sampling time... , for the For each marketing channel, calculate the cross-channel mutual stimulation intensity component and the total event arrival intensity function value for that channel. The cross-channel mutual stimulation intensity component is the event arrival intensity function. Remove the basic strength component The remaining portion. Divide the cross-channel mutual excitation intensity component by the total value of the event arrival intensity function. The mutual stimulation contribution rate of the current marketing channel at the current sampling time is obtained. The value of the mutual stimulation contribution rate ranges from 0 to 1. The closer the value is to 1, the more dependent the arrival of events on the channel at the current time is on the stimulation driven by other channels; the closer the value is to 0, the more the arrival of events on the channel at the current time is driven by its own basic arrival intensity component. The mutual stimulation contribution rates of each sampling time are arranged in chronological order to generate a time series of mutual stimulation contribution rates for each marketing channel.

[0040] The reason for choosing the mutual stimulation contribution rate time series as the monitoring object for subsequent early warning, rather than directly monitoring the event arrival intensity function itself, is based on the following considerations: the absolute value of the event arrival intensity function is greatly affected by external factors such as marketing budget and seasonal fluctuations, and its normal range is difficult to define stably; while the mutual stimulation contribution rate reflects the relative proportion of inter-channel synergy in the current channel activity, and this proportion indicator has better stability when the external environment changes. When the mutual stimulation contribution rate of a certain channel continues to decline, it means that the driving force of other channels on that channel is weakening, which often indicates that the marketing effect of that channel is about to decline.

[0041] refer to Figure 1The horizontal axis of the graph represents time in hours, ranging from 0 to 20; the vertical axis represents event arrival intensity in hours, ranging from 0 to 0.8. The graph also displays the event arrival intensity function. The overall curve and the stacked distribution of its constituent components.

[0042] At the bottom layer of the graph, a region of approximately constant height is laid out along the horizontal axis. The upper bound of this region is approximately 0.15 events per hour, representing the base arrival intensity component of marketing channel A. This refers to the constant rate at which marketing events spontaneously generate by marketing channel A without any other channel events influencing it. As can be seen, the basic arrival intensity component remains a horizontal band throughout the entire observation period, unchanged over time, which is consistent with the definition of the constant rate of the basic arrival intensity component.

[0043] Above the base arrival intensity component, three layers of dynamically changing regions are stacked sequentially. The first layer, immediately above the base arrival intensity component, represents the self-excitation component of subsequent events generated by events already occurring in marketing channel A. The second layer above that represents the cross-channel mutual excitation intensity component exerted by marketing channel B on marketing channel A, labeled as the B-to-A mutual excitation component. The topmost third layer represents the cross-channel mutual excitation intensity component exerted by marketing channel C on marketing channel A, labeled as the C-to-A mutual excitation component. These three layers together constitute the excitation effect portion of the event arrival intensity function, excluding the base arrival intensity component. The upper boundary after superimposing all components is the event arrival intensity function of marketing channel A. , indicated by solid lines.

[0044] As can be clearly observed from the graph, whenever a new marketing event occurs, the event arrival intensity function... A steep upward jump occurs, followed by an exponential decline. This sawtooth waveform of "jump followed by decay" repeats throughout the time axis, perfectly matching the characteristic that the excitation effect of a single marketing event in the multidimensional Hawkes self-excitation point process decreases exponentially over time from the moment the event occurs. At the bottom of the graph, three different shaped markers are marked below the horizontal axis, representing the event occurrence times of marketing channels A, B, and C as triangles, squares, and rhombuses, respectively. By comparing the bottom markers with the jump positions of the intensity function above, it can be verified that not only events of marketing channel A itself cause jumps in the intensity function, but the occurrence of events of marketing channels B and C also leads to a significant jump response in the intensity function of marketing channel A's events, demonstrating the actual mechanism of cross-channel mutual excitation effects.

[0045] After obtaining the time series of mutual stimulation contribution rates for each marketing channel, the SPOT streaming extreme value detection method was applied to each time series. Traditional fixed-threshold alerts require a pre-set alarm threshold. However, the pace and intensity of marketing activities change continuously with the business cycle. Fixed thresholds can easily generate a large number of false alarms during peak business periods, while potentially missing genuine anomalies during off-peak periods. The SPOT streaming extreme value detection method, based on the Pickands-Balkema-deHaan theorem in extreme value theory, can automatically determine alarm boundaries based solely on the statistical characteristics of the data tails, without making any assumptions about the overall distribution of the data. Therefore, it can adapt to the non-stationary nature of marketing data.

[0046] Since this method focuses on the abnormal direction of the mutual excitation contribution rate, that is, the abnormal decay of the cross-channel mutual excitation effect, the SPOT flow cytometry extreme value detection method monitors the lower tail extreme value of the mutual excitation contribution rate time series in this method.

[0047] The SPOT streaming extreme value detection method includes an initialization phase and a point-by-point streaming detection phase.

[0048] The purpose of the initialization phase is to establish an initial fit to the generalized Pareto distribution using data from a stable operating period. A pre-defined initialization window is extracted from the initial segment of the mutual-excitation contribution rate time series. The typical length of this window is 500 to 2000 data points, ensuring it contains a sufficient number of extremely low values ​​to support a reliable fit to the tail distribution. A pre-defined initial screening threshold is set within this window. This threshold is determined by arranging all mutual-excitation contribution rate data points in the window from smallest to largest, and using the values ​​corresponding to the 2nd to 5th percentiles as the initial screening threshold. The specific percentile is chosen based on the proportion of extreme values ​​expected to be included in the tail analysis. All mutual-excitation contribution rate data points below the pre-defined initial screening threshold are then selected within the initial window. The difference between each selected data point and the pre-defined initial screening threshold is recorded as the lower limit deviation of the corresponding data point.

[0049] According to the Pickands-Balkema-deHaan theorem in extreme value theory, when the screening threshold is sufficiently high, the conditional distribution of the out-of-limit deviation approximately follows a generalized Pareto distribution. Using the lower-side out-of-limit deviations of all screened mutual-excitation contribution rate data points, the generalized Pareto distribution is fitted using the maximum likelihood estimation method to obtain the initial tail shape parameter values. and tail scale parameter values Among them, the tail shape parameter value The decay rate at the tail of the distribution is determined: when When the distribution is equal to 0, the tails decay exponentially. When the value is greater than 0, the distribution has a heavy-tailed property. When the value is less than 0, the distribution has a finite tail. Tail scale parameter value. It is a positive real number, which controls the dispersion of the out-of-limit deviation.

[0050] After the initialization window is completed, the point-by-point streaming detection phase begins. Upon receiving each newly arrived mutual excitation contribution rate data point, the following processing is performed: First, it is determined whether the newly arrived mutual excitation contribution rate data point is below the current screening threshold. If it is below the current screening threshold, the lower bound deviation of the newly arrived mutual excitation contribution rate data point is calculated and included in the existing lower bound deviation set. A generalized Pareto distribution is then refitted to update the tail shape parameter and tail scale parameter values. This dynamic update mechanism allows the parameters of the generalized Pareto distribution to track the slow drift of the tail characteristics of the mutual excitation contribution rate, avoiding detection sensitivity degradation due to long-term use of fixed parameters.

[0051] refer to Figure 2 The horizontal axis of the graph represents the sampling time, ranging from 1 to 400; the vertical axis represents the mutual stimulation contribution rate, ranging from 0.3 to 0.8. The graph shows the complete detection process of the mutual stimulation contribution rate time series of a target marketing channel under the SPOT streaming extreme value detection method, including the initialization stage and the point-by-point streaming detection stage.

[0052] The initial segment of the horizontal axis, i.e., the interval from sampling time 1 to approximately 80, is marked as the initialization window, separated from subsequent intervals by a vertical dashed line. The mutual excitation contribution rate data points within the initialization window are used to establish an initial fit for the generalized Pareto distribution, providing initial tail distribution parameters for subsequent point-by-point flow cytometry detection. It can be seen that the mutual excitation contribution rate within the initialization window fluctuates between 0.60 and 0.70, indicating relatively stable overall operation.

[0053] After the initialization window is complete, the graph indicates the start of the point-by-point streaming detection phase. During this phase, two key reference lines are plotted. The first, represented by a horizontal dashed line at approximately 0.50, is the current filtering threshold. This threshold filters data points in the mutual stimulation contribution rate time series that fall below it; these selected data points are used for fitting and updating the generalized Pareto distribution. The second line, represented by a stepped solid line below the current filtering threshold, is the dynamic warning threshold. This threshold is determined based on the updated generalized Pareto distribution at a preset false alarm rate level. The dynamic warning threshold is not fixed but gradually adjusts as new extreme data points are included in the lower limit deviation set and the generalized Pareto distribution is refitted. Therefore, it appears as a stepped shape that changes progressively with the sampling time in the graph.

[0054] Within the sampling time interval from 1 to 280, although the mutual stimulation contribution rate time series exhibited some random fluctuations, it remained within the normal range of 0.55 to 0.70, far exceeding the dynamic warning threshold, and did not trigger any warnings. Starting around sampling time 280, the mutual stimulation contribution rate showed a continuous downward trend, indicating that the cross-channel mutual stimulation effect was gradually weakening. Within the sampling time interval from 300 to 400, several circularly marked data points were observed scattered between the current screening threshold and the dynamic warning threshold. These data points were below the current screening threshold but not yet below the dynamic warning threshold, and were included in the lower limit deviation set to update the generalized Pareto distribution parameters. As the mutual stimulation contribution rate continued to decline, after sampling time approximately 340, several warning trigger points marked with inverted triangles appeared. The mutual stimulation contribution rate of these data points had fallen below the dynamic warning threshold. Based on this, the system determined that the cross-channel mutual stimulation effect of the corresponding marketing channel had abnormally weakened, generating a marketing effect abnormality warning signal carrying the target marketing channel identifier and the warning trigger time.

[0055] Based on the updated generalized Pareto distribution, a dynamic warning threshold is determined for the current moment under a preset false alarm rate level. The false alarm rate level represents the upper limit of the probability that a single data point will be incorrectly identified as an anomaly under normal conditions, with a typical value being [value missing]. arrive This means that at most one false alarm is allowed per 1,000 to 10,000 normal data points. The dynamic warning threshold is determined as follows: based on the inverse function of the cumulative distribution function of the generalized Pareto distribution, substituting the warning false alarm rate level, the current tail shape parameter value, and the tail scale parameter value, the deviation threshold corresponding to the warning false alarm rate level under the current tail distribution condition is calculated. Subtracting the deviation threshold from the current screening threshold gives the dynamic warning threshold.

[0056] refer to Figure 3 The horizontal axis of the graph represents the lower limit deviation, ranging from 0.000 to 0.175; the vertical axis represents the probability density, ranging from 0 to 25. The graph also shows the empirical distribution histogram of the lower limit deviation and the fitted curve of the generalized Pareto distribution, as well as the deviation critical value and the corresponding tail probability region determined by the fitting results.

[0057] The histogram in the figure shows the empirical frequency distribution of the lower bound deviations of all data points below the screening threshold selected from the mutual excitation contribution rate time series. The height of each bar represents the normalized probability density of data points falling into the corresponding deviation interval. The distribution pattern of the histogram shows that the lower bound deviations are concentrated in a small region close to 0, with the probability density reaching its highest value of approximately 23 near the deviation of 0. As the deviation increases, the probability density generally decreases, but there are some local fluctuations in the middle region. This non-uniform distribution characteristic indicates that the simple normal distribution assumption is insufficient to accurately describe the actual distribution pattern of the lower bound deviations.

[0058] The smooth curve superimposed on the histogram is the generalized Pareto distribution curve obtained by fitting using the maximum likelihood estimation method. The generalized Pareto distribution curve monotonically decreases along the horizontal axis from its highest point near the vertical axis, and its decay rate is determined by the tail shape parameter. and tail size parameters This is determined jointly. The figure shows the fitted parameter values: tail shape parameters. =0.08, tail scale parameter =0.0397. Tail shape parameters A value greater than 0 indicates that the distribution of the lower-side deviation has heavy-tailed characteristics, meaning that the probability of extremely large deviations is higher than expected by the exponential distribution. This result suggests that it is necessary to use the generalized Pareto distribution, which can characterize the heavy-tailed characteristics, for tail modeling in marketing data.

[0059] At approximately 0.125 on the horizontal axis, the deviation threshold is marked with a vertical dashed line in the figure. The deviation threshold is calculated using the inverse function of the cumulative distribution function, based on a fitted generalized Pareto distribution, at a preset false alarm rate level. The area below the generalized Pareto distribution curve to the right of the deviation threshold is the tail probability region, which is marked in the figure along with its corresponding false alarm rate. The area of ​​the tail probability region equals the preset false alarm rate level, representing the probability that the lower limit deviation of the mutual-stimulation contribution rate data points exceeds the deviation threshold under normal conditions. Subtracting the deviation threshold from the current screening threshold yields the dynamic warning threshold; a warning is triggered when the mutual-stimulation contribution rate data points fall below the dynamic warning threshold. The figure clearly shows that the deviation threshold is located at the far tail of the histogram distribution; the vast majority of lower limit deviations are less than the deviation threshold, with only a very small number of extreme deviations falling into the tail probability region, consistent with the preset low false alarm rate level.

[0060] The system determines whether the newly arrived mutual stimulation contribution rate data point is below the dynamic warning threshold. If it is below the threshold, it determines that the cross-channel mutual stimulation effect of the corresponding marketing channel has abnormally weakened, generating a marketing effect anomaly warning signal carrying the target marketing channel identifier and the warning trigger time. In an optional implementation, a continuous trigger counting mechanism can be introduced, requiring that a marketing effect anomaly warning signal be generated only when the data point is below the dynamic warning threshold for a consecutive number of sampling times. For example, a warning is only confirmed after triggering for 3 or 5 consecutive sampling times, further reducing false alarms caused by instantaneous fluctuations. In another optional implementation, multiple levels of false alarm rates can be set, for example, setting separate false alarm rates for each level. , , Three levels are set, corresponding to three different levels of abnormal marketing performance warning signals: prompt level, warning level, and severe level.

[0061] After generating an abnormal marketing performance warning signal in step 2, the system needs to further answer a key question: to what extent does the abnormal decay of the mutual stimulation effect of the target marketing channel affect the actual marketing conversion effect of the channel, and whether this impact truly stems from a strategy change within the channel itself, rather than from fluctuations in the overall market environment? Simply observing the changes in the marketing conversion rate of the target marketing channel before and after the warning is triggered is insufficient, because even without any strategy changes, fluctuations in external market factors can lead to changes in the marketing conversion rate. Direct before-and-after comparisons cannot separate the true impact of strategy changes from environmental noise. The synthetic control method addresses this attribution problem by constructing a "trend in the marketing conversion rate that the target marketing channel should have shown if it had not undergone a strategy change." This virtual trend curve is the counterfactual control sequence, quantifying the deviation between the actual trend and the counterfactual control sequence as a causal effect, thereby achieving accurate attribution of the impact of strategy changes.

[0062] In response to the marketing performance anomaly warning signal, the roles to be analyzed are first determined. The warning signal carries the target marketing channel identifier and the warning trigger time; this target marketing channel is designated as the unit to be analyzed. Simultaneously, the remaining marketing channels that did not trigger the warning signal within the same time period are grouped into a donor pool. The logic behind the donor pool is that channels that did not trigger the warning were in normal operation before and after the warning trigger time, without experiencing significant strategy changes. Therefore, the marketing conversion rate trends of these channels can reflect the influence of common market environment factors and can serve as the basis for constructing a counterfactual comparison. The number of channels in the donor pool is denoted as... , are positive integers and In practical applications The typical range is 3 to 15. Too few channels in the donor pool will result in insufficient flexibility in fitting the synthesis results, while too many may introduce channels with business characteristics that differ too much from the target marketing channels, thereby reducing the synthesis quality. In an optional implementation, a pre-screening step can be added before forming the donor pool to calculate the similarity between each candidate channel and the target marketing channel in dimensions such as business type, user profile overlap, and geographic coverage, and only channels with similarity higher than a preset threshold will be retained in the donor pool.

[0063] After completing the role classification, data for calculating the composite coefficients needs to be extracted. A preset matching interval should be set before the alert trigger time. The length of the preset matching interval needs to cover a sufficient time span to ensure the reliability of the composite fit. A typical value for the preset matching interval length is 30 to 90 sampling points before the alert trigger time, with the specific value depending on the sampling frequency and business cycle characteristics of the marketing conversion rate time series. If the sampling frequency is one data point per day, the preset matching interval is 30 to 90 days before the alert trigger time; if the sampling frequency is one data point per hour, it corresponds to 30 to 90 hours. The preset matching interval should not be too short, otherwise the composite results are easily affected by short-term fluctuations and become unstable; it should also not be too long, because data from too far back may include early strategy changes or market structural changes, which could introduce interference.

[0064] The marketing conversion rate time series of each channel in the analysis unit and the donor pool is extracted within a preset matching interval. The marketing conversion rate is defined as the ratio of the number of users who completed a conversion to the total number of users reached within the time period corresponding to each sampling moment. The preset matching interval contains... At each sampling time, the marketing conversion rate time series of the unit to be analyzed within the preset matching interval is denoted as a sequence of length [length missing]. vector ,in Each component corresponds to the marketing conversion rate at one sampling time. The first component in the donor pool... The time series of marketing conversion rates for each channel within a preset matching interval is denoted as (length ). vector , Take 1 to Integers between [a certain range].

[0065] The core of the synthetic control method lies in solving a set of synthetic coefficients so that the weighted combination of the marketing conversion rate time series of each channel in the donor pool can reproduce the marketing conversion rate trend of the unit to be analyzed as accurately as possible within a preset matching interval. The synthetic coefficients are denoted as... They correspond to the 1st to the 1st in the donor pool, respectively. One channel. The composite coefficients are solved using constrained quadratic programming. The optimization objective and constraints of the constrained quadratic programming are as follows.

[0066] The optimization objective is to minimize the sum of squared time-by-time deviations between the composite sequence (obtained by linearly combining the marketing conversion rate time series of each channel in the donor pool according to the composite coefficients) and the marketing conversion rate time series of the unit to be analyzed within a preset matching interval. This optimization objective can be expressed as: In the above formula, For the unit to be analyzed and processed, within the preset matching interval, the first... Marketing conversion rate at each sampling point For the first in the donor pool The first channel within the preset matching range Marketing conversion rate at each sampling point For the first in the donor pool The synthesis coefficient corresponding to each channel This represents the total number of sampling times within the preset matching interval.

[0067] The constraints include two terms: the first term requires that all composite coefficients be non-negative, i.e. For all Valid; Item 2 requires the sum of all composite coefficients to equal 1, i.e. The purpose of imposing a non-negativity constraint is to avoid negative synthesis coefficients leading to extrapolation. If negative coefficients are allowed, the synthesis result may exceed the range of values ​​for all donor pool channels, thus losing economic interpretability. Imposing a constraint that the sum of the coefficients equals 1 ensures that the synthesis result always falls within the convex hull of the values ​​for each channel in the donor pool, allowing the counterfactual control sequence to be reasonably interpreted as a "weighted average of the trends of several real channels".

[0068] The aforementioned constrained quadratic programming problem is a convex optimization problem, and the existence and uniqueness of the global optimal solution are theoretically guaranteed. In practical implementation, the effective set method or the interior point method can be used. In an optional implementation, when the number of donor pool channels... Larger and preset matching interval length When the time interval is relatively short, the optimization problem may exhibit overfitting, meaning that the synthesized sequence closely matches the unit to be analyzed within the preset matching interval, but its extrapolation ability is poor outside the preset matching interval. To mitigate this problem, a regularization term can be introduced into the optimization objective. For example, a penalty term proportional to the sum of squares of the time-by-time deviations can be added after the sum of squares of the time-by-time deviations. The proportionality coefficient of the penalty term typically ranges from 0.01 to 0.1.

[0069] refer to Figure 4The horizontal axis of the graph represents time in days, ranging from 1 to 100; the vertical axis represents marketing conversion rate, ranging from approximately 0.06 to 0.11. The graph shows the actual marketing conversion rate time series of the target marketing channels, the counterfactual control series constructed from the donor pool channels, and the original marketing conversion rate time series of each channel in the donor pool.

[0070] At the 60th day on the horizontal axis, the graph uses a vertical line to mark the warning trigger time, dividing the entire time interval into two parts. The left side of the vertical line marks the preset matching interval, corresponding to day 1 to day 60; the right side of the vertical line marks the period after the warning trigger time, corresponding to day 61 to day 100.

[0071] Within the preset matching interval, the actual marketing conversion rate time series of the target marketing channel is represented by a solid line, fluctuating between 0.085 and 0.105. Within the same interval, the counterfactual control sequence is represented by a dashed line. It can be seen that the counterfactual control sequence closely matches the actual trend of the target marketing channel within the preset matching interval; the two curves almost overlap, with only a very small fitting residual. This high degree of fit verifies that the composite coefficients obtained through constrained quadratic programming can accurately reproduce the trend characteristics of the target marketing channel within the preset matching interval by weighting the marketing conversion rate time series of each channel in the donor pool. The composite coefficients corresponding to each channel in the donor pool are marked in the upper right corner of the figure. =0.42, =0.28, =0.22, =0.08, all composite coefficients are non-negative and their sum is equal to 1, which satisfies the constraints of the constrained quadratic programming.

[0072] After the warning was triggered, the actual marketing conversion rate of the target marketing channel showed a significant downward trend, gradually decreasing from approximately 0.095 before the warning to approximately 0.07 by day 100. Meanwhile, the counterfactual control sequence continued to fluctuate within the range of 0.085 to 0.095 after the warning was triggered, without exhibiting a similar downward trend. The counterfactual control sequence characterizes the expected marketing conversion rate trend of the target marketing channel under the assumption of no strategy change. Since no strategy changes occurred in any channel in the donor pool before or after the warning was triggered, linearly combining the marketing conversion rates of the donor pool channels after the warning using the same coefficient is equivalent to projecting the influence of the common market environment onto the target marketing channel. In the figure, a double-headed arrow marks the vertical distance between the actual trend and the counterfactual control sequence near day 85, labeled as a causal effect. The vertical spacing represents the value of the causal effect sequence at the corresponding time point, reflecting the net causal impact of the marketing strategy change on the marketing conversion rate of the target marketing channel. The graph shows that the absolute value of the causal effect gradually increases over time after the warning trigger point, indicating that the negative impact of the strategy change on marketing effectiveness is not a one-time shock but rather a continuous, accumulating effect. The graph also shows four thin lines with different trends, corresponding to the marketing conversion rate time series of the four channels in the donor pool. These channels maintained their normal fluctuation patterns before and after the warning trigger point, without exhibiting a sustained decline similar to that of the target marketing channel. This further confirms that the decline in the marketing conversion rate of the target marketing channel was not caused by common market environment factors, but rather stemmed from the strategy change that occurred within that channel itself.

[0073] Solving for the composition coefficients Afterwards, it is necessary to verify the quality of the synthetic fit. A common verification method is to calculate the root mean square deviation between the synthetic sequence and the marketing conversion rate time series of the unit to be analyzed within the preset matching interval. If the root mean square deviation exceeds 5% to 10% of the average marketing conversion rate of the unit to be analyzed within the preset matching interval, the synthetic fit is considered to be of unsatisfactory quality. In this case, it is advisable to expand the donor pool range, extend the preset matching interval, or adjust the regularization parameters and solve again.

[0074] After the quality of the composite fit is verified, the obtained composite coefficients are applied to the marketing conversion rate time series of each channel in the donor pool after the warning trigger time, and a linear combination is performed to obtain the counterfactual control series. Assume that a total of [number missing] data were collected after the warning trigger time. The data at the sampling time point, the counterfactual control sequence at the ... The values ​​at each sampling time point are: In the above formula, For the counterfactual comparison sequence in the 1st The value at each sampling time, For the first in the donor pool The first channel after the warning is triggered. Actual marketing conversion rate at each sampling point These are the composite coefficients obtained from the previous solution. The counterfactual comparison sequence characterizes the expected marketing conversion rate trend of the target marketing channel under the assumption that no strategy change has occurred. The underlying logic is that the channels in the donor pool did not undergo strategy changes before and after the warning trigger time. Therefore, the trend changes of these channels after the warning trigger time only reflect the influence of the common market environment. Combining these channels with the same composite coefficients as those in the preset matching interval is equivalent to projecting the influence of the common market environment according to the target marketing channel's previous response pattern to the market environment, thereby obtaining the trend that the target marketing channel should have shown when it was not affected by strategy changes.

[0075] The causal effect sequence is obtained by subtracting the actual marketing conversion rate time series of the unit to be analyzed after the warning trigger time from the counterfactual control sequence time step by time. The values ​​at each sampling time point are: In the above formula, For the first The causal effect value at each sampling time point. This represents the actual marketing conversion rate of the unit to be analyzed at that moment. This represents the value of the counterfactual contrast sequence at that moment. When A negative value indicates that the actual marketing conversion rate of the target marketing channel is lower than the expected level assuming no strategy change, meaning the strategy change has had a negative causal impact on marketing effectiveness; when A positive value indicates that the strategy change actually led to a positive improvement in performance. The causal effect sequence is used as the output of the attribution results for the deviation in performance of the target marketing channel due to the marketing strategy change.

[0076] To further evaluate the statistical significance of the causal effect sequences and avoid misinterpreting normal random fluctuations as true causal effects, a permutation test can be performed. Specifically, each channel in the donor pool is hypothetically considered as the unit to be analyzed, and the remaining channels are treated as donor pools. The causal effect sequences for each channel in each donor pool are calculated using the same synthetic control method. A placebo causal effect sequence. Since the channels in the donor pool have not actually undergone strategic changes, their placebo causal effect sequence should theoretically fluctuate around 0. The causal effect sequence of the target marketing channel is compared with that of all... The attribution result is considered statistically significant if the causal effect of the target marketing channel is consistently greater in absolute value than the vast majority of placebo causal effects. A specific criterion is to calculate the root mean square (RMS) value of each causal effect sequence after the warning trigger time; if the RMS value of the causal effect sequence of the target marketing channel is significantly greater than the placebo causal effect in all... If a sequence ranks within the top 10% (i.e., ranked 1st or 2nd, depending on the size of the donor pool), the causal effect is considered statistically significant. In a scenario where the donor pool includes 10 channels, this is equivalent to an equivalent p-value of approximately 0.09, and the significance threshold can be adjusted according to the stringency of business requirements.

[0077] In an optional implementation, in addition to outputting the time-by-time causal effect sequence, all effects after the warning trigger time can also be calculated. The arithmetic mean of the causal effect values ​​at each sampling time point is used as the cumulative average causal effect, or the sum of the causal effect values ​​at each time point is calculated as the cumulative causal effect curve, providing decision-makers with attribution perspectives of different granularities. The cumulative average causal effect is suitable for assessing the overall average impact level caused by policy changes, while the cumulative causal effect curve is convenient for observing whether the impact of policy changes amplifies or gradually diminishes over time.

[0078] In another optional implementation, when multiple marketing channels trigger abnormal marketing performance warning signals simultaneously, the above-mentioned synthetic control attribution process can be executed independently for each target marketing channel that triggers the warning, and the causal effect sequences of each channel can be summarized and compared to identify which channel is most affected by the strategy change and which channel recovers from the performance deviation the fastest, thereby providing a more comprehensive decision-making basis for the refined adjustment of marketing strategies.

[0079] This invention is not limited to the specific embodiments described above. The invention extends to any new feature or combination disclosed in this specification, as well as any new method or process step or combination disclosed herein.

Claims

1. A multi-channel data monitoring, early warning, and attribution method based on a time-series prediction model, characterized in that, Includes the following steps: Step 1: Based on the multidimensional Hawkes self-excitation point process, jointly model the marketing event flow received from multiple marketing channels. Construct an event arrival intensity function for each marketing channel, which is composed of the basic arrival intensity component and the cross-channel mutual excitation intensity component. Estimate the parameters of the multidimensional Hawkes self-excitation point process. At each sampling time, perform intensity component decomposition for each marketing channel. Determine the mutual excitation contribution rate according to the proportion of the cross-channel mutual excitation intensity component in the total value of the event arrival intensity function. Arrange the mutual excitation contribution rates at each sampling time in chronological order to generate a time series of mutual excitation contribution rates for each marketing channel. Step 2: The SPOT streaming extreme value detection method is used to process the time series of mutual stimulation contribution rates of each marketing channel. For data points below the screening threshold, a generalized Pareto distribution is fitted to the lower limit deviation. The tail distribution parameters of the generalized Pareto distribution are dynamically updated as new data points arrive. Based on the updated generalized Pareto distribution, a dynamic warning threshold is determined at a preset false alarm rate level. When the mutual stimulation contribution rate data points are lower than the dynamic warning threshold, an abnormal marketing effect warning signal carrying the target marketing channel identifier and the warning trigger time is generated. Step 3: In response to the marketing performance abnormality warning signal, the target marketing channel is designated as the unit to be analyzed and processed based on the synthetic control method. The remaining marketing channels that have not triggered the marketing performance abnormality warning signal are formed into a donor pool. The time series of marketing conversion rates of each channel in the donor pool are used to solve the synthetic coefficient and construct the counterfactual control sequence of the unit to be analyzed and processed. The actual marketing conversion rate time series of the unit to be analyzed and processed after the warning trigger time is compared with the counterfactual control sequence time by time to obtain the causal effect sequence and output it as the effect deviation attribution result.

2. The method as described in claim 1, characterized in that, Each marketing event in the marketing event stream carries the time of the event and the channel identifier to which it belongs; the basic arrival strength component is the constant rate at which each marketing channel spontaneously generates marketing events; The cross-channel mutual stimulation intensity component is the sum of the stimulation effects exerted by each marketing event that has occurred in other channels on each marketing channel. The stimulation effect of a single marketing event decreases exponentially over time from the moment the marketing event occurs.

3. The method as described in claim 1, characterized in that, The parameters of the multidimensional Hawkes self-excitation point process are estimated using the expectation-maximization algorithm, and the expectation phase and the maximization phase are executed in each iteration. In the expectation phase, for each marketing event in the marketing event stream, the posterior attribution probability of each marketing event triggered by the stimuli effect of each source channel is calculated. The posterior attribution probability is determined based on the share of the stimuli effect contributed by each source channel to the channel to which the current marketing event belongs at the time of the current marketing event's occurrence relative to the total stimuli effect of all sources.

4. The method as described in claim 3, characterized in that, In the maximization phase, the basic arrival intensity components of each channel and the excitation decay rate between each channel pair are re-estimated based on the posterior attribution probability. The updated parameters are then substituted into the expectation phase of the next iteration until the parameter change between two consecutive iterations is less than the preset convergence threshold, at which point the iteration stops.

5. The method as described in claim 1, characterized in that, The intensity component decomposition includes: at each sampling time, calculating the cross-channel mutual stimulation intensity component and the total event arrival intensity function of the current marketing channel, and dividing the cross-channel mutual stimulation intensity component by the total event arrival intensity function to obtain the mutual stimulation contribution rate of the current marketing channel at the current sampling time.

6. The method as described in claim 1, characterized in that, The SPOT streaming extreme value detection method includes an initialization phase: a preset length initialization window is extracted from the beginning segment of the mutual excitation contribution rate time series; all mutual excitation contribution rate data points below the preset initial screening threshold are selected within the initialization window; the difference between each selected mutual excitation contribution rate data point and the preset initial screening threshold is recorded as the lower limit deviation of the corresponding data point; a generalized Pareto distribution is fitted using the lower limit deviations of all selected mutual excitation contribution rate data points to obtain the initial tail shape parameter value and tail scale parameter value.

7. The method as described in claim 6, characterized in that, After the initialization window is completed, the point-by-point streaming detection phase begins. Upon receiving each newly arrived mutual-stimulation contribution rate data point, the following steps are performed: First, determine if the newly arrived mutual-stimulation contribution rate data point is below the current filtering threshold. If it is, calculate the lower limit deviation of the newly arrived mutual-stimulation contribution rate data point and include it in the existing lower limit deviation set. Then, refit the generalized Pareto distribution to update the tail shape parameter and tail scale parameter values. Based on the updated generalized Pareto distribution, determine the dynamic warning threshold at the current time, under the warning false alarm rate level. Next, determine if the newly arrived mutual-stimulation contribution rate data point is below the dynamic warning threshold. If it is, determine that the cross-channel mutual-stimulation effect of the corresponding marketing channel has abnormally decayed, and generate a marketing effect abnormality warning signal carrying the target marketing channel identifier and the warning trigger time.

8. The method as claimed in claim 1, characterized in that, In step 3, a preset matching interval is set before the warning is triggered, and the marketing conversion rate time series of each channel in the unit to be analyzed and the donor pool are extracted within the preset matching interval for solving the synthesis coefficient.

9. The method as described in claim 8, characterized in that, The composite coefficients are solved by constrained quadratic programming. The constrained quadratic programming requires that all composite coefficients are non-negative and that the sum of all composite coefficients equals 1. The optimization objective is to minimize the sum of squared deviations between the composite sequence obtained by linearly combining the marketing conversion rate time series of each channel in the donor pool according to the composite coefficients and the marketing conversion rate time series of the unit to be analyzed within the preset matching interval.

10. The method as described in claim 9, characterized in that, The obtained composite coefficients are applied to the marketing conversion rate time series of each channel in the donor pool after the warning trigger time, and linear combination is performed to obtain the counterfactual control sequence. The counterfactual control sequence represents the expected marketing conversion rate trend of the target marketing channel under the assumption that no strategy change has occurred.

Images