Mobile advertisement effect evaluation system based on user interaction behavior

Abstract

The invention relates to the technical field of advertisement data analysis, and provides a user interaction behavior-based mobile advertisement effect evaluation system, which comprises an acquisition and judgment module for performing judgment after advertisement exposure, simultaneously sensing three types of signals in a preset window, calculating a door domain consistency score and acquiring an interaction event, generating a local statistical vector and an interaction strength score; the service evaluation module is used for performing causal inference by adopting a random control experiment, decomposing each contact contribution based on a multi-contact attribution model, obtaining an advertisement increment effect and a confidence interval thereof, and generating an evaluation result according to three levels of creativity, advertisement and crowd; the cooperative control module is used for monitoring the state, performing online adjustment on the sampling rate, the feature dimension and the return frequency, controlling the evaluation variance not to exceed a preset threshold value, and dynamically balancing the endurance and the evaluation precision; and the privacy convergence block is used for protecting data by adopting a differential privacy noise injection and federated learning mode, and is also responsible for recording and managing agreement states and withdrawal requests of the users.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of advertising data analysis, and particularly relates to a mobile advertising effect evaluation system based on user interaction behavior. BACKGROUND

[0002] Mobile advertising is a paid promotion content displayed on mobile devices such as smart phones, tablet computers and the like, and is a marketing method using the characteristics of mobile devices and mobile Internet, including display advertising, in-application advertising, short message advertising and the like. Mobile advertising has precision, immediacy and interactivity, and has become an important bridge connecting users and goods or services.

[0003] In China, application No. CN202510084836.2 discloses a mobile advertising effect analysis method and system based on user behavior. The modules include: through the SDA interface and API interface of the mobile advertising platform, real-time grabbing of detailed data of user behavior and advertising placement, through calculation of activity parameters, advertising acceptance parameters and consumption level parameters, the user is divided into different activity, acceptance and consumption level user groups, accurate characterization and classification of user behavior are realized, which helps advertisers to more accurately target users, can mine potential user demand, improve the accuracy and conversion rate of advertising, and at the same time reduce advertising cost; through calculation of advertising revenue index, advertising adaptation index and shopping willingness index and other key indicators, the effect of advertising placement is comprehensively evaluated, which provides a scientific decision basis for advertisers, helps to optimize advertising placement strategy and improve advertising effect.

[0004] The existing local life placement still simply pursues single click and exposure count, lacks a way to accurately distinguish between casual browsing and real store preparation, and behaviors occurring at different times are not assigned different values. On the other hand, offline transactions often occur outside the platform, resulting in scattered data sources and repeated calculations, making it difficult to evaluate the effectiveness and payback period of the placement in a unified manner.

[0005] To solve the above problems, the present application provides a mobile advertising effect evaluation system based on user interaction behavior. The system senses different collection signals on the user device side, makes a high-confidence determination on the action of entering the store, combines the determination result with common interactions after the advertisement, obtains a more stable store credibility score, and reduces the effect of misjudgment and repeated counting without touching personal privacy information. SUMMARY

[0006] The present application aims to solve the problem of distinguishing between casual browsing and real store preparation in local life placement, and the problem of scattered data sources caused by offline transactions, and provides a mobile advertising effect evaluation system based on user interaction behavior.

[0007] The application solves the above technical problems by adopting the technical scheme of a mobile advertisement effect evaluation system based on user interaction behavior, comprising:

[0008] A collection and determination module determines exposure and false touch after advertisement exposure, simultaneously senses three types of signals, i.e., short-time deceleration, sudden change in geomagnetic intensity, and environmental light jump within a preset window, calculates domain consistency score, collects interaction events, generates local statistical vector and interaction intensity score, and performs exposure and false touch determination after advertisement exposure.

[0009] A service evaluation module receives the local statistical vector and interaction intensity score, performs caliber uniformity and calibration, performs causal inference by using a randomized controlled experiment, decomposes the contribution of each touch point based on a multi-touch point attribution model, obtains advertisement incremental effect and its confidence interval, and generates evaluation results in terms of three levels of creativity, advertisement, and crowd.

[0010] A collaborative control module monitors terminal power, temperature, and network state, and performs online adjustment on sampling rate, feature dimension, and backhaul frequency to control the evaluation variance to be less than a preset threshold, and dynamically balances the endurance and evaluation accuracy.

[0011] A privacy integration module implements data minimization during collection and reporting, protects data by using differential privacy noise injection and federated learning, records and manages the consent status and withdrawal request of users, stores data according to the predetermined retention period and regional requirements, and generates auditable operation logs.

[0012] As a preferred embodiment, the specific steps of the collection and determination module are as follows:

[0013] Within a preset time window after advertisement exposure, the original signal is collected according to the sampling rate issued by the collaborative control module, the original signal is denoised, a low-pass filter is used to remove high-frequency components higher than a preset cutoff frequency, for short-time abnormal points, a sliding window median absolute deviation detection is used, when the absolute deviation of the sample and the window median is greater than a preset multiple, the sample is replaced with the window median, the preprocessed signal is resampled and time-aligned.

[0014] The short-time deceleration confidence is calculated, the speed is reduced and normalized within a preset time window, and the formula is as follows:

[0015] ,

[0016] wherein represents the short-time deceleration confidence, represents the normalization function, represents the preset speed reduction threshold, max represents the larger value function, represents the observed speed reduction within the time window.

[0017] The geomagnetic intensity mutation confidence is calculated, the magnetic field difference amplitude is calculated in a preset time window and is normalized, and a formula is as follows:

[0018] ,

[0019] wherein represents the geomagnetic intensity mutation confidence, represents a geomagnetic intensity change amount, represents a geomagnetic mutation determination threshold, and min represents a minimum value function;

[0020] The ambient light jump confidence is calculated, the light relative change rate is calculated in a preset time window and is normalized, and a formula is as follows:

[0021] ,

[0022] wherein represents the ambient light jump confidence, represents the ambient light relative change rate, represents a light jump threshold;

[0023] The domain consistency score is calculated by weighted fusion, the interactive events of a user and an advertisement are collected in an event-driven manner, a plurality of types of interactive events are set, the types, time stamps and duration of event triggering are recorded according to a page life cycle callback trigger, exposure determination is performed according to the collected event data and original signals, the interactive intensity score is obtained by linear combination of normalized features according to weights, and a formula is as follows:

[0024] ,

[0025] wherein IS represents the interactive intensity score, V represents a viewing time length, S represents a scroll depth, and C represents a click number, is a weight, i is a weight index, and clip represents a clipping function, is a time length weight, is a depth weight, is a number weight.

[0026] As a preferred embodiment, the specific steps of the service evaluation module are as follows:

[0027] The service side receives a report package from the terminal side, checks the report package integrity and parameter version number, performs caliber conversion according to a version mapping table when the version is inconsistent, performs inverse quantization on the bucket and digest in the statistical vector according to the version table in the report package, performs cohort-based aggregation and deduplication on the count items, removes obviously abnormal packages, and corrects sample deviation caused by device distribution difference by using a reweighting method;

[0028] Logistic regression is used as the scoring model, and the calibrated features are used as input to construct the in-store credibility score. The formula is as follows:

[0029] ,

[0030] in Represents logarithmic probability. Represents the intercept term. Represents the gate domain consistency score, Indicates the gate region weight. Indicates the interaction intensity score. This represents the interaction strength weight, and j represents the index of the additional covariate. This indicates the corresponding covariate. The weight coefficient sequence, Represents the sequence of additional covariates. This indicates the credibility score for in-store visits;

[0031] In scenarios supporting randomized distribution, a randomized controlled trial is used to directly estimate the increment. Subjects are randomly assigned to a distribution group and a control group, and the average results for both groups are calculated. The increment is derived from the difference between the two group means. The significance is assessed using standard error and confidence intervals. In scenarios that are not completely randomized, propensity score is first fitted using observed covariates. Then, each sample is assigned a weight according to its processing instructions, and the difference between the weighted means of the two groups is calculated on the weighted samples as the causal increment estimate of the processing.

[0032] As a preferred embodiment, the specific steps of the collaborative control module are as follows:

[0033] During terminal operation, continuously monitor operational metrics and calculate the sampling rate and reporting strategy at the current time point based on the operational metrics and the received server precision constraint parameters.

[0034] The solution is obtained by solving the constrained optimization problem with Lagrange multipliers. , where t represents the current time, minimize the expected power consumption while ensuring that the estimated variance does not exceed the upper limit of accuracy. When computing resources are limited, a pre-set mapping table is used as an approximate solution and the version number is recorded. The interaction intensity score is mapped to the current power consumption and network quality as the reporting probability. The real-time reporting is determined by combining the local cache queue length and the allowed delay budget. When the conditions for real-time reporting are not met, the reporting is delayed and merged into batches to reduce energy consumption and network usage.

[0035] Based on the calculated strategy, the sampling frequency of the sensor, the features to be collected, and the local aggregation window are adjusted in real time, and the reporting timing and frequency are determined according to the set strategy. When a restricted operating state is detected, the sampling rate is reduced first and the local aggregation period is extended until the conditions are restored. When the terminal resources are sufficient, sampling is gradually restored and real-time reporting is allowed. For each decision, the parameter version, operating indicators, and energy consumption estimation log are recorded, and the strategy update is sent to the server according to the agreement to maintain the consistency of the end-cloud caliber.

[0036] As a preferred implementation, the specific steps of the privacy compliance module are as follows:

[0037] Record and manage user consent status and withdrawal requests; determine data collection and retention permissions in real time based on user decisions; perform data summarization on the device side; for reported statistics, use differential privacy noise injection and noise-enhancing updates before reporting according to a preset privacy budget; support federated learning framework and adopt secure aggregation mechanism to avoid plaintext information leakage from individual devices; implement encryption protection and access control for transmission channels and static storage; store raw data in designated physical partitions according to localization policies and automatically isolate data copies according to compliance requirements; delete relevant local and server records after user withdrawal of consent and record the deletion operation in the audit log.

[0038] Beneficial effects

[0039] Compared with the prior art, the present invention has the following advantages:

[0040] 1. By combining sensor data with user interaction behavior, it accurately distinguishes between casual browsing and genuine in-store intent, avoiding the single evaluation method of relying solely on click and impression counts in traditional approaches, thus improving the accuracy of advertising effectiveness.

[0041] 2. By integrating and deduplicating cross-platform and cross-channel data through the service evaluation module, the problem of scattered and duplicate calculations of offline transaction data is solved, ensuring the accuracy and consistency of evaluation results. It provides a complete analysis of the effectiveness of the campaign by combining online advertising results with offline in-store performance. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of this application, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of this application and should not be regarded as a limitation on the scope of this application.

[0043] Figure 1 This is a block diagram of the present invention;

[0044] Figure 2This is a comparison diagram of the effects of the present invention and the prior art, where gray bars represent the prior art and black bars represent the present invention. Detailed Implementation

[0045] To make the technical means, creative features, objectives, and effects of this invention easier to understand, the invention is further described below with reference to specific embodiments. However, the following embodiments are merely preferred embodiments of this invention and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments described herein without creative effort are all within the protection scope of this invention.

[0046] Example 1:

[0047] To achieve the above objectives, please refer to Figure 1 This invention provides a mobile advertising effectiveness evaluation system based on user interaction behavior, the system comprising the following modules:

[0048] The data collection and judgment module performs exposure and accidental touch judgment after the advertisement is exposed, and simultaneously senses three types of signals within a preset window: short-term deceleration, sudden change in geomagnetic intensity, and sudden change in ambient light. It calculates the gate consistency score, collects interactive events, and generates local statistical vectors and interaction intensity scores.

[0049] The service evaluation module receives local statistical vectors and interaction intensity scores, first unifies and calibrates the standards, then uses randomized controlled experiments to make causal inferences, decomposes the contribution of each touchpoint based on a multi-touchpoint attribution model, obtains the incremental effect of advertising and its confidence interval, and generates evaluation results according to three levels: creative, advertising and audience.

[0050] The collaborative control module monitors the terminal's battery level, temperature, and network status, and adjusts the sampling rate, feature dimensions, and feedback frequency online to control the evaluation variance to not exceed a preset threshold, dynamically balancing battery life and evaluation accuracy.

[0051] The privacy compliance module minimizes data during the collection and reporting process, protects data by using differential privacy noise injection and federated learning, records and manages user consent status and withdrawal requests, stores data according to predetermined retention periods and geographical requirements, and generates auditable operation logs.

[0052] Example 2:

[0053] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the data acquisition and determination module:

[0054] Within a preset time window after the advertisement is displayed, such as 30 seconds, the original signal is collected according to the sampling rate issued by the collaborative control module, including: terminal speed change sequence, terminal acceleration sequence, terminal magnetic field strength sequence, ambient light intensity sequence, and touch interaction event sequence; the original signal is denoised by using a low-pass filter to remove high-frequency components higher than the preset cutoff frequency, and for short-term anomalies, the absolute deviation of the median of the sliding window is used for detection. When the absolute deviation of the sample from the median of the window exceeds 3 times, the sample is replaced with the median of the window, and the preprocessed signal is resampled and time-aligned.

[0055] Calculate the short-term deceleration confidence score, and then calculate the velocity dimensionality reduction and normalization within a preset time window, using the following formula:

[0056] ,

[0057] in Indicates the confidence level of short-term deceleration. Represents the normalization function. This represents the preset speed reduction threshold, and max represents a larger value function. This indicates the observed velocity decrease within the time window;

[0058] To calculate the confidence level of geomagnetic intensity abrupt changes, the magnetic field difference amplitude is calculated and normalized within a preset time window using the following formula:

[0059] ,

[0060] in Indicates the confidence level of a sudden change in geomagnetic intensity. This represents the change in geomagnetic intensity. This represents the threshold for determining geomagnetic abrupt changes, and min represents the smaller value function.

[0061] To calculate the confidence level of ambient light abrupt changes, the relative rate of change of light within a preset time window is calculated and then normalized. The formula is as follows:

[0062] ,

[0063] in This indicates the confidence level of changes in ambient light intensity. Indicates the relative rate of change of ambient light intensity. Indicates the threshold for light transition;

[0064] The consistency score is calculated using a weighted fusion method. Event-driven data collection is performed on user-ad interaction events, including clicks, long presses, scrolling, and switching. These events are triggered by callbacks throughout the page's lifecycle, recording the event type, timestamp, and duration. Exposure is determined based on the collected event data and raw signals. For example, if the visible area ratio is greater than 50%, the continuous visible duration is greater than 1 second, and the average scroll rate threshold is less than 0.5 screen heights per second, the interaction intensity score is obtained through a linear combination of normalized features with weights. The formula is as follows:

[0065] ,

[0066] Where IS represents the interaction intensity score, V represents the viewing time, S represents the scrolling depth, and C represents the number of clicks. Here, is the weight, i is the weight index, and clip represents the truncation function. Weighted by duration, For depth weights, The weight is the number of times.

[0067] Example 3:

[0068] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the service evaluation module:

[0069] The server receives reporting packets from the client, which include statistical vectors, interaction strength scores, gate consistency scores, and judgment results. The server verifies the integrity of the reporting packets and the parameter version numbers. If the versions do not match, the server performs a conversion according to the version mapping table. Based on the version table in the reporting packets, the server performs inverse quantization on the buckets and summaries in the statistical vectors, performs cohort-based aggregation and deduplication on the count items, removes obviously abnormal packets, and uses a reweighting method to correct the sample offset caused by differences in device distribution.

[0070] Logistic regression is used as the scoring model, and the calibrated features are used as input to construct the in-store credibility score. The formula is as follows:

[0071] ,

[0072] in Represents logarithmic probability. Represents the intercept term. Represents the gate domain consistency score, Indicates the gate region weight. Indicates the interaction intensity score. This represents the interaction strength weight, and j represents the index of the additional covariate. This indicates the corresponding covariate. The weight coefficient sequence, Represents the sequence of additional covariates. This indicates the credibility score for in-store visits;

[0073] In scenarios supporting randomized delivery, such as in-app pop-ups, coupons, and in-store promotions, a randomized controlled trial is used to directly estimate the increment. Subjects are randomly assigned to either a delivery group or a control group, and the average results for both groups are calculated. The increment is derived from the difference between the two group means, and its significance is assessed using standard error and confidence intervals. In scenarios that are not completely randomized, such as targeted delivery to users or users actively claiming coupons, propensity score is first fitted using observed covariates. Then, assign a weight to each sample according to its processing instructions; for example, the weight of the processing group is... The control group has a weight of The difference between the weighted means of the two groups is calculated on the weighted sample as an estimate of the causal increment of the treatment.

[0074] The server allocates the estimated increment to each touchpoint based on creative, ad unit, and audience package. The attribution method decomposes the increment using marginal contribution approximation to obtain the relative contribution of each delivery unit. Confidence intervals are calculated for the increment estimate and significance tests are performed. The confidence intervals are constructed from the approximate standard error based on the sample variance. The significance test provides probability values ​​to judge the reliability of the estimate. The in-store credibility, increment estimate, confidence interval, and sample size indicators given for creative, ad, and audience are summarized to generate stratified reports, time series, and funnel views.

[0075] Example 4:

[0076] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the collaborative control module:

[0077] During terminal operation, continuously monitor operating indicators such as battery level, device temperature, foreground / background status, network type, and processor load. Based on the operating indicators and the received server precision constraint parameters, calculate the sampling rate and reporting strategy for the current time point.

[0078] Specifically, the calculation of the sampling rate and reporting strategy at the current time point is achieved by solving a constrained optimization problem with Lagrange multipliers, within the allowed sampling rate range, based on real-time power consumption, temperature, processor load, network quality, and the upper limit of accuracy issued by the server. , where t represents the current time, minimize the expected power consumption while ensuring that the estimated variance does not exceed the upper limit of accuracy. When computing resources are limited, a pre-set mapping table is used as an approximate solution and the version number is recorded. The interaction intensity score is mapped to the current power consumption and network quality as the reporting probability. The real-time reporting is determined by combining the local cache queue length and the allowed delay budget. When the conditions for real-time reporting are not met, the reporting is delayed and merged into batches to reduce energy consumption and network usage.

[0079] Based on the calculated strategy, the sampling frequency of the sensor, the features to be collected, and the local aggregation window are adjusted in real time, and the reporting timing and frequency are determined according to the set strategy. When insufficient power, overheating of the device, entering the background, or deterioration of network quality are detected, the sampling rate is reduced first and the local aggregation cycle is extended until the conditions are restored. When sufficient power, idle device, and good network are detected, sampling is gradually restored and real-time reporting is allowed. For each decision, the parameter version, operating indicators, and energy consumption estimation log are recorded, and the strategy update is sent to the server according to the agreement to maintain the consistency of the end-cloud.

[0080] Example 5:

[0081] This embodiment is an explanation of Embodiment 1, specifically outlining the implementation steps of the privacy compliance module:

[0082] Record and manage user consent status and withdrawal requests; determine data collection and retention permissions in real time based on user decisions; perform data minimization and summarization on the device side; report bucketed summaries when necessary, such as for small samples or rare events; for reported statistics, use differential privacy noise injection and noise-enhancing updates according to a preset privacy budget before reporting; support federated learning frameworks and employ secure aggregation mechanisms to prevent plaintext information leakage from individual devices; implement encryption protection and access control for transmission channels and static storage; store raw data in designated physical partitions according to a regional strategy and automatically isolate data copies according to compliance requirements; delete relevant local and server records after a user withdraws consent, and record the deletion operation in the audit log.

[0083] like Figure 2 This is a performance comparison chart of a mobile advertising effectiveness evaluation system based on user interaction behavior. The horizontal axis lists key performance indicators, and the vertical axis represents exemplified performance scores, ranging from 0 to 5, with higher values ​​indicating better performance. The aim is to intuitively demonstrate the expected improvement of the present invention in key capabilities compared to typical existing technologies, showcasing the differences between existing technologies and the present invention in performance evaluation. Through weighting, time-based weighting, and data fusion techniques, the chart demonstrates the accuracy and intelligence of the present invention in advertising effectiveness evaluation, helping advertisers optimize their campaign strategies, improve the accuracy of payback period predictions, and effectively identify users' true intent to visit stores.

[0084] The embodiments of the present invention described above are subject to modification and change of method by those skilled in the art without departing from the embodiments and broader aspects of the present invention. The appended claims are intended to include all such modifications and changes of method that do not depart from the present invention.

Claims

1. A mobile advertisement effectiveness evaluation system based on user interaction behavior, characterized by, The application comprises: The collection determination module determines exposure and mis-touch after exposure of an advertisement, simultaneously senses three types of signals, i.e., short-time deceleration, sudden change in geomagnetic intensity and sudden change in ambient light, within a preset window, calculates the gate domain consistency score, collects interactive events, generates a local statistical vector and an interactive intensity score; The service evaluation module receives the local statistical vector and the interactive intensity score, performs caliber unification and calibration, performs causal inference by using a randomized controlled experiment, decomposes the contribution of each touch point based on a multi-touch attribution model, obtains an advertisement incremental effect and a confidence interval thereof, and generates evaluation results in three levels of creative, advertisement and crowd; The collaborative control module monitors terminal power, temperature and network state, and adjusts sampling rate, feature dimension and return frequency online to control the evaluation variance to be less than a preset threshold, and dynamically balances endurance and evaluation accuracy; The privacy integration module implements data minimization in the collection and reporting process, protects data by using differential privacy noise injection and federated learning, records and manages the consent status and withdrawal request of users, stores data according to a predetermined retention period and regional requirements, and generates operation logs. 2.The mobile advertisement effectiveness evaluation system based on user interaction behavior according to claim 1, wherein: The specific steps of the collection determination module are as follows: Within a preset time window after exposure of an advertisement, the original signal is collected according to the sampling rate issued by the collaborative control module, the original signal is denoised, a low-pass filter is used to remove high-frequency components higher than a preset cutoff frequency, for short-time abnormal points, a sliding window median absolute deviation detection is used, when the absolute deviation of the sample and the window median is greater than a preset multiple, the sample is replaced with the window median, the preprocessed signal is resampled and time-aligned; The short-time deceleration confidence is calculated by reducing the dimension of the speed and then normalizing it within a preset time window, and the formula is as follows: ， wherein denotes a short-time deceleration confidence, denotes a normalization function, denotes a preset speed drop threshold, max denotes a maximum function, denotes a speed drop observed within a time window; The geomagnetic intensity mutation confidence is calculated by calculating the differential amplitude of the magnetic field and then normalizing it within a preset time window, and the formula is as follows: ， wherein denotes a geomagnetic intensity mutation confidence, denotes a geomagnetic intensity change amount, denotes a geomagnetic mutation determination threshold, and min denotes a minimum function. The ambient light jump confidence is calculated by calculating the relative change rate of the light and then normalizing it within a preset time window, and the formula is as follows: ， wherein represents an ambient light illumination jump confidence, represents a relative rate of change of ambient light illumination, represents an illumination jump threshold. 3.The mobile advertisement effectiveness evaluation system based on user interaction behavior according to claim 2, wherein: The specific steps of the collection determination module further comprise: The gate domain consistency score is calculated by weighted fusion, the interactive events of the user and the advertisement are collected in an event-driven manner, multiple types of interactive events are set, the type, timestamp and duration of the event trigger are recorded according to the page life cycle callback trigger, exposure is determined according to the collected event data and the original signal, the interactive intensity score is obtained by linear combination of the normalized features according to the weight, and the formula is as follows: ， where IS represents the interaction strength score, V represents the watch time, S represents the scroll depth, C represents the click count, is the weight, i is the weight index, and clip represents the clipping function, is the time weight, is the depth weight, is the count weight.

4. The system of claim 1, wherein: The specific steps of the service evaluation module are as follows: The server receives the reporting package from the terminal side, checks the integrity and parameter version number of the reporting package, performs caliber conversion according to the version mapping table if the versions are inconsistent, dequantizes the bucket and digest in the statistical vector according to the version table in the reporting package, performs cohort-based aggregation and deduplication on the count items, removes obviously abnormal packages, and corrects the sample offset caused by the difference in device distribution by using re-weighting.

5. The system for evaluating the effect of mobile advertising based on user interaction behavior according to claim 4, wherein: The specific steps of the service evaluation module further comprise: The logistic regression is used as a scoring model, and the calibrated features are used as input to construct the store visit credibility score, and the formula is as follows: ， wherein represents a log odds ratio, represents an intercept term, represents a gate domain consistency score, represents a gate domain weight, represents an interaction strength score, represents an interaction strength weight, j represents an index of an additional covariate, represents a sequence of weight coefficients corresponding to covariates , represents a sequence of additional covariates, represents a store visitability score.

6. The system for evaluating the effect of mobile advertising based on user interaction behavior according to claim 4, wherein: The specific steps of the service evaluation module further comprise: In scenarios supporting randomized distribution, a randomized controlled trial is used to directly estimate the increment. Subjects are randomly assigned to a distribution group and a control group, and the average results for both groups are calculated. The increment is derived from the difference between the two group means. The significance is assessed using standard error and confidence intervals. In scenarios that are not completely randomized, propensity score is first fitted using observed covariates. Then, each sample is assigned a weight according to its processing instructions, and the difference between the weighted means of the two groups is calculated on the weighted samples as the causal increment estimate of the processing.

7. The system of claim 4, wherein: The specific steps of the service evaluation module further include: The server allocates the estimated increment as the contribution value of each touch point according to the creative, ad unit and audience package, the attribution method is decomposed by marginal contribution approximation, the relative contribution of each delivery unit is obtained, the confidence interval is calculated for the increment estimation and the significance test is performed, the confidence interval is constructed by the approximate standard error based on the sample variance, the significance test provides a probability value to judge the reliability of the estimation, the indicators given according to the material, advertisement and audience are summarized, and the layered report, time series and funnel view are generated.

8. The system for evaluating the effect of mobile advertising based on user interaction behavior according to claim 1, wherein: The specific steps of the collaborative control module are: The running indicators are continuously monitored during the terminal running, and the sampling rate and reporting strategy at the current time point are calculated based on the running indicators and the received server precision constraint parameters; by solving a constrained optimization problem with Lagrange multipliers where t represents the current time, makes the expected power consumption minimum under the premise that the estimation variance does not exceed the accuracy upper limit, when the computing resources are limited, a preset mapping table is used as an approximate solution and a version number is recorded; the interaction intensity score, the current power and the network quality are mapped into a reporting probability, and the local cache queue length and the allowed delay budget are combined to determine real-time reporting, when the instant reporting condition is not met, the reporting is delayed and reported in batches to reduce energy consumption and network occupation. 9.The mobile advertisement effectiveness evaluation system based on user interaction behavior according to claim 8, wherein: The specific steps of the collaborative control module further include: According to the calculated strategy, the sampling frequency of the sensor, the feature items to be collected and the local aggregation window are adjusted in real time, and the reporting time and reporting frequency are determined according to the set strategy; when it is detected that the terminal is in a limited running state, the sampling rate is preferentially reduced and the local aggregation period is prolonged until the condition is restored, when it is detected that the terminal resource is sufficient, the sampling is gradually restored and immediate reporting is allowed, the parameter version, running indicator and energy consumption estimation log are recorded for each decision, and the strategy update is delivered to the server according to the agreement to maintain the consistency of the terminal and cloud caliber.

10. The system for evaluating the effect of mobile advertising based on user interaction behavior according to claim 1, wherein: The specific steps of the privacy compliance module are: The consent status and withdrawal request of the user are recorded and managed, the data collection and retention authority is decided in real time according to the user decision, data abstraction is performed on the terminal side, differential privacy noise injection is adopted for the reported statistics according to the preset privacy budget, and the update is reported after noise, federal learning framework is supported and security aggregation mechanism is adopted to avoid leakage of clear information of a single device, encryption protection and access control are implemented on the transmission channel and static storage, the original data is stored in the specified physical partition according to the regionalization strategy, and the data copy is automatically isolated according to the compliance requirements, the relevant local and server records are deleted after the user withdraws the consent, and the deletion operation is recorded in the audit log.

Images