Intelligent evaluation system for advertisement putting effect monitoring

Abstract

The invention discloses an intelligent evaluation system for advertisement putting effect monitoring, and relates to the technical field of advertisement putting, and the system comprises the steps: loading a bidirectional pre-training language model and a multi-modal fusion model, and synchronizing advertisement metadata to generate a structured feature matrix; collecting user behaviors, biological characteristics and environment data in real time, generating a four-channel environment map, splicing the four-channel environment map with the structured characteristic matrix, and inputting the spliced four-channel environment map and the structured characteristic matrix into a Transform encoder to form a joint characteristic vector; matching a high-density high-emotion region according to the candidate budget allocation scheme, and generating a cognitive load index in combination with the facial micro-expression of the user; user interaction data cognition load scenes are analyzed, core advertisement group weights are evaluated and optimized through a regret value mechanism and a Shapley value, and a final dynamic putting scheme is generated; the response difference of the advertisement group to the cognitive load is quantified through the sensitivity coefficient, and the marginal contribution of the advertisement group is evaluated in combination with the Shapley value, so that the accurate identification and budget inclination of the core advertisement group are realized.

Description

Technical Field

[0001] The present invention relates to the technical field of advertising placement, and particularly to an intelligent evaluation system for monitoring the effect of advertising placement. Background Art

[0002] In recent years, the technology for monitoring the effect of advertising placement has undergone a paradigm shift from traditional statistical models to machine learning-driven ones. Early methods relied on simple static metrics such as exposure and click-through rate, combined with linear regression or random forest models for effect prediction, but it was difficult to handle the complex correlations of multimodal data. With the development of deep learning, models based on Transformer and multimodal fusion have been introduced, but their real-time adaptability in dynamic environments is still insufficient. For example, in the prior art, the spatio-temporal alignment of multimodal data depends on artificial rules, the non-linear relationship between environmental parameters (such as temperature, density) and advertising effects is roughly modeled, and parameter updates are mostly batch processed, unable to respond in real time to changes in user cognitive load. In addition, budget allocation strategies are mostly based on static historical data and lack the ability to adaptively adjust to dynamic scenarios (such as high-density crowd areas, user mood fluctuations).

[0003] The deficiencies of the prior art are mainly reflected in three aspects: (1) The initialization of Bayesian network parameters depends on artificial prior assumptions and does not dynamically adjust in combination with multimodal environmental data, resulting in insufficient sensitivity of the model to environmental changes; (2) The budget allocation strategy does not combine user cognitive load with the sensitivity of advertising groups, resulting in limited accuracy and dynamics of resource allocation. Summary of the Invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides an intelligent evaluation system for monitoring the effect of advertising placement to solve the problems of insufficient parameter self-adaptability and low accuracy and resource utilization efficiency caused by the decoupling of budget allocation and user cognitive load.

[0006] To solve the above technical problems, the present invention provides the following technical solutions:

[0007] The present invention provides an intelligent evaluation system for monitoring the effect of advertising placement, which includes,

[0008] A model loading module that loads a bidirectional pre-trained language model and a multimodal fusion model, synchronizes advertising metadata to generate a structured feature matrix, and initializes Bayesian network parameters;

[0009] A feature fusion module that real-time collects user behavior, biometric, and environmental data, generates a four-channel environmental map, and splices it with the structured feature matrix to input into a Transformer encoder to form a joint feature vector;

[0010] The click-through rate prediction module calculates the immediate reward based on the user's click feedback, updates the Bayesian network parameters through variational inference, combines the joint feature vector, and predicts the click-through rate;

[0011] The budget allocation module generates an initial candidate budget plan using Thompson sampling based on the predicted click-through rate, constructs an objective function, and generates the final candidate budget allocation plan through a dynamic exploration-exploitation strategy;

[0012] The cognitive load module matches high-density and high-emotion regions according to the candidate budget allocation plan, adjusts the multi-modal stimulus parameters, and generates a cognitive load index in combination with the user's facial micro-expressions;

[0013] The sensitivity analysis module analyzes the cognitive load scenarios of user interaction data, optimizes the weights of the core advertisement group through the regret value mechanism and the Shapley value evaluation, and generates the final dynamic placement plan.

[0014] As a preferred solution of the intelligent evaluation system for advertising placement effect monitoring described in the present invention, wherein: the generation of the structured feature matrix refers to loading the bidirectional pre-trained language model and the multi-modal fusion model into the inference optimization engine, and using the loaded bidirectional pre-trained language model and multi-modal fusion model to perform data cleaning and verification on the historical advertisement metadata to generate the structured feature matrix.

[0015] As a preferred solution of the intelligent evaluation system for advertising placement effect monitoring described in the present invention, wherein: the initialization of the Bayesian network parameters refers to setting the Bayesian hierarchical network parameters, constructing a Bayesian neural network using the PyTorch framework, initializing the weights with a uniform distribution, performing forward propagation on the structured feature matrix, and conducting batch inference tests on the constructed Bayesian neural network.

[0016] As a preferred solution of the intelligent evaluation system for advertising placement effect monitoring described in the present invention, wherein: the formation of the joint feature vector includes the following steps,

[0017] Collect the user's movement trajectory, user location, user physiological signals, and three-dimensional population distribution;

[0018] Obtain temperature data to generate a two-dimensional temperature heat map, and superimpose the three-dimensional population distribution to form a four-channel environmental map;

[0019] Extract the historical click-through rate mean, budget allocation ratio, and logarithm of the exposure volume in the structured feature matrix, input them into the Transformer encoder, and dynamically allocate weights through the multi-head attention mechanism to form the joint feature vector.

[0020] As a preferred solution of the intelligent evaluation system for advertising delivery effect monitoring according to the present invention, wherein: the updating of the Bayesian network parameters includes the following steps,

[0021] Obtain user click feedback in real time and calculate the immediate reward;

[0022] Using the variational inference framework, set the initialized Bayesian network parameters as the prior distribution, and perform likelihood function modeling. Adjust the Bayesian network parameters by establishing the posterior optimization objective;

[0023] Use the Adam optimizer to perform gradient descent iterative update to obtain the updated Bayesian network parameters.

[0024] As a preferred solution of the intelligent evaluation system for advertising delivery effect monitoring according to the present invention, wherein: the predicted click-through rate refers to inputting the updated Bayesian network parameters and the joint feature vector into the Bayesian neural network for forward propagation, and generating the predicted click-through rate by calculating the conditional expected value of the advertisement group at the current moment.

[0025] As a preferred solution of the intelligent evaluation system for advertising delivery effect monitoring according to the present invention, wherein: the generation of the final candidate budget allocation plan includes the following steps,

[0026] Based on the predicted click-through rate, use the Thompson sampling algorithm to generate an initial candidate budget plan for each advertisement group;

[0027] Construct an objective optimization function to maximize the weighted revenue, and add an L2 regularization term to suppress the drastic fluctuation of the budget;

[0028] Adjust the initial candidate budget plan through gradient descent to generate the final candidate budget allocation plan.

[0029] As a preferred solution of the intelligent evaluation system for advertising delivery effect monitoring according to the present invention, wherein: the adjustment of the multi-modal stimulus parameters includes the following steps,

[0030] Based on the final candidate budget allocation plan, combine the total budget amount and set a high-priority threshold to generate a dynamic delivery strategy;

[0031] Based on the dynamic delivery strategy and the three-dimensional population density heat map, minimize the path cost and preferentially cover high-density areas to generate an optimal path coordinate sequence and LED parameter instructions;

[0032] According to the LED parameter instructions and the user emotion activity index, adjust the multi-modal stimulus parameters in real time.

[0033] As a preferred solution of the intelligent evaluation system for advertising placement effect monitoring according to the present invention, wherein: the generation of the cognitive load index refers to adjusting the results of multi-modal stimuli, analyzing the pupil position and fixation duration using OpenCV, calculating the attention concentration index, and performing weighted fusion in combination with the EDA signal.

[0034] As a preferred solution of the intelligent evaluation system for advertising placement effect monitoring according to the present invention, wherein: the generation of the final dynamic placement plan refers to dividing the user interaction scenario according to the cognitive load index, constructing an advertisement group effect-cognitive load matrix, calculating the sensitivity coefficient and performing sensitivity ranking, combining the dynamic placement strategy, historical optimal data and regret value mechanism, evaluating the contribution degree of the advertisement group using the Shapley value, and dynamically adjusting the budget allocation to optimize the weight of the core advertisement group.

[0035] The beneficial effects of the present invention are as follows: through the combination of the distributed file system and Kafka streaming processing, the efficient loading of the multi-modal model and the automation of historical data cleaning are realized. By setting the adaptive mechanism of the global benchmark reward mean, specific variance matrix and noise standard deviation, and combining the variational inference framework to update the parameters of the real-time click-through rate feedback, the limitation of relying on artificial prior assumptions is solved; the dynamic learning rate decay strategy balances the model convergence speed and the overfitting risk; through the spatio-temporal alignment of the four-channel environmental map and user behavior data, combined with the multi-head attention mechanism of the Transformer, the non-linear correlation modeling of environmental parameters and advertisement features is realized, and the objective function of Thompson sampling and L2 regularization enables the budget allocation strategy to dynamically respond to the cognitive load scenario; by quantifying the response difference of the advertisement group to the cognitive load through the sensitivity coefficient and combining the Shapley value to evaluate the marginal contribution of the advertisement group, the accurate identification of the core advertisement group and budget tilt are realized. BRIEF DESCRIPTION OF THE DRAWINGS

[0036] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings required for the description of the embodiments will be briefly introduced below. Obviously, the drawings in the following description are only some embodiments of the present invention, and those of ordinary skill in the art can obtain other drawings without creative efforts based on these drawings.

[0037] Figure 1 It is a system diagram of the intelligent evaluation system for advertising placement effect monitoring in Embodiment 1.

[0038] Figure 2 It is a schematic diagram of feature fusion and click-through rate prediction in Embodiment 1.

[0039] Figure 3 It is a schematic diagram of the dynamic adjustment of budget allocation in Embodiment 1.

[0040] Figure 4 Schematic diagram of cognitive load analysis and sensitivity assessment in Embodiment 1 Detailed implementation manners

[0041] To make the above objects, features and advantages of the present invention more obvious and understandable, the following detailed description of the specific implementation manners of the present invention will be given in conjunction with the accompanying drawings of the specification

[0042] In the following description, many specific details are set forth to facilitate a thorough understanding of the present invention. However, the present invention may be implemented in other ways different from those described herein. Those skilled in the art may make similar extensions without departing from the connotation of the present invention. Therefore, the present invention is not limited by the specific embodiments disclosed below

[0043] Secondly, the so-called "one embodiment" or "embodiment" herein refers to a specific feature, structure or characteristic that may be included in at least one implementation manner of the present invention. The "in one embodiment" that appears in different places in this specification does not all refer to the same embodiment, nor is it a separate or alternative embodiment that excludes other embodiments

[0044] Embodiment 1, referring to Figures 1 to 4 , which is the first embodiment of the present invention. This embodiment provides an intelligent evaluation system for monitoring the effect of advertising placement, including the following steps

[0045] A model loading module that loads a bidirectional pre-trained language model and a multi-modal fusion model, synchronizes advertisement metadata to generate a structured feature matrix, and initializes Bayesian network parameters, including the following steps

[0046] Pull the bidirectional pre-trained language model and the multi-modal fusion model from the distributed file system, and use the inference optimization engine to convert the bidirectional pre-trained language model and the multi-modal fusion model into an optimized format; load the converted bidirectional pre-trained language model and the multi-modal fusion model into the inference optimization engine, set the maximum sequence length of the bidirectional pre-trained language model to 512 characters, and the number of input channels of the multi-modal fusion model to 3 (RGB + temperature + density), and perform a performance test to obtain the successfully loaded bidirectional pre-trained language model and multi-modal fusion model; pull the historical data of the past 30 days from the advertisement metadata topic through a Kafka (distributed message queue system) consumer, and perform data cleaning and verification through the successfully loaded bidirectional pre-trained language model and multi-modal fusion model to generate a structured feature matrix

[0047] Further explanation, the historical data includes advertisement ID, mean and standard deviation of click-through rate, budget allocation ratio, exposure volume, audience demographics such as age, gender, geographical distribution, etc

[0048] Data cleaning includes invalid data filtering and numerical field standardization.

[0049] Specifically, invalid data filtering refers to deleting the ad group records with an exposure volume < 1000 times, because small sample data cannot reflect the real effect, checking the effectiveness of click-through rate, excluding records with undefined click-through rate or a standard deviation greater than 0.1 to avoid the interference of outliers, and verifying the rationality of budget allocation to ensure that the budget ratio is between 0% - 100%, and correcting outliers (such as 120%) to the nearest neighbor mean (such as the average allocation ratio of 15% of adjacent ad groups); numerical field standardization refers to normalizing the budget allocation ratio after invalid data filtering to the interval [0, 1], performing logarithmic transformation on the exposure volume to narrow the numerical range, and encoding the categorical fields of the audience demographics.

[0050] Extract the historical click-through rate mean, budget allocation standard deviation, logarithmic exposure volume, and categorical field encoding of the audience demographics from all the cleaned data to generate a structured feature matrix; based on the structured feature matrix, set the Bayesian hierarchical network parameters, and use the PyTorch (open-source machine learning library) framework to construct a Bayesian neural network and output the initialized Bayesian network parameters.

[0051] Further explanation, setting the Bayesian hierarchical network parameters includes setting the global baseline reward mean, specific variance matrix, and noise standard deviation.

[0052] Specifically, setting the global baseline reward mean refers to setting the average click-through rate of the initial baseline according to the statistics of historical ad data; setting the specific variance matrix refers to initializing the diagonal elements and non-diagonal elements for each ad group to represent the initial assumption that the effects of each ad group are independent and have small differences; setting the noise standard deviation is to control the sensitivity of the Bayesian hierarchical model to the observed noise. If the click-through rate prediction fluctuates too much in the subsequent observation, the noise standard deviation can be adjusted dynamically.

[0053] Using the PyTorch framework to construct a Bayesian neural network means setting a neural network composed of an input layer, fully connected layers, and an output layer, initializing the weights using the Xavier uniform distribution, setting the bias term of the output layer to 0, and setting the bias term of the input layer to the minimum value; performing forward propagation on the structured feature matrix, conducting batch inference tests on the constructed Bayesian neural network to check for consistency, and outputting the initialized Bayesian network parameters.

[0054] The feature fusion module, which collects user behavior, biometric, and environmental data in real time, generates a four-channel environmental map, and splices it with the structured feature matrix and inputs it into the Transformer encoder to form a joint feature vector, includes the following steps

[0055] Collect the mobile trajectories of user MAC addresses through WiFi probes, and generate user stay heatmaps through AP association algorithms; deploy Bluetooth beacons, calculate user locations through received signal strength, and generate dynamic crowd density heatmaps (density value = number of users per unit area); collect users' facial micro-expressions and skin conductivity signals, filter them through LabVIEW (graphical programming tool), and generate user emotion activity indexes; use lidar to scan the three-dimensional space of the mall, identify the dynamic crowd distribution through the DBSCAN clustering algorithm (density-based spatial clustering algorithm), and generate voxel-level three-dimensional crowd density heatmaps; align all the data in the user stay heatmaps, dynamic crowd density heatmaps, user emotion activity indexes, and three-dimensional crowd density heatmaps according to timestamps; obtain the real-time temperature distribution from the mall temperature control system and generate two-dimensional temperature heatmaps; overlay the three-dimensional crowd density heatmaps and two-dimensional temperature heatmaps to generate four-channel environmental maps.

[0056] Extract the historical click-through rate mean, budget allocation ratio, and logarithm of exposure volume related to the current ad group from the structured feature matrix output by the initialized Bayesian network, and inject the four-channel environmental map, user stay heatmap, dynamic crowd density heatmap, and user emotion activity index to form a spliced feature vector; input the spliced feature vector into a Transformer encoder (a deep learning model based on self-attention mechanism), and through the multi-head attention mechanism, calculate the dynamic weight allocation (such as assigning higher attention weights to periods with high user emotions) to generate a joint feature vector.

[0057] The click-through rate prediction module calculates the immediate reward based on user click feedback, updates the Bayesian network parameters through variational inference, and combines the joint feature vector to predict the click-through rate, including the following steps

[0058] When the user clicks on an advertisement, the system obtains in real time the feedback records of clicks or impressions without clicks through the advertisement server interface. Each feedback record contains the advertisement ID, the exposure timestamp, and the user ID, which are used to track the correlation between user behavior and advertisement effectiveness. Combining the feedback records of the real-time exposure volume and click volume of the advertisement group, the click-through rate at that moment is calculated. Multiplying the click-through rate by the click volume to obtain the immediate reward. Using the variational inference framework, the initialized Bayesian network parameters are set as the prior distribution, assuming that the immediate reward follows a normal distribution for likelihood function modeling. By establishing a posterior optimization objective, the Bayesian network parameters are adjusted to maximize the evidence lower bound and balance the difference between the likelihood function and the prior distribution. Using the Adam optimizer (with an initial learning rate of 0.01) for gradient descent iterative updates, to prevent overfitting, the learning rate is dynamically decayed after each update of the Bayesian network parameters, obtaining the updated Bayesian network parameters. Inputting the updated Bayesian network parameters and the joint feature vector into the Bayesian neural network for forward propagation, and generating the predicted click-through rate by calculating the conditional expectation value of the advertisement group at the current moment. For example, after the update, the baseline rewards of a certain advertisement group are all 0.08, combined with the dynamic weight allocation of the feature vector, and the final predicted click-through rate is 8%.

[0059] The budget allocation module, based on the predicted click-through rate, uses Thompson sampling (Thompson sampling algorithm) to generate an initial candidate budget plan, constructs an objective function, and generates the final candidate budget allocation plan through a dynamic exploration-exploitation strategy, including the following steps,

[0060] Based on the predicted click-through rate, combined with the historical budget allocation ratio (such as the current ratio of 15%), the standard deviations of historical exposure and click-through volume, a distribution model is constructed for each ad group through the Thompson sampling algorithm, including the number of successes and failures. Set the number of samplings according to the number of ad groups, set the sampling upper and lower limits based on the historical budget ratio, perform N samplings on each ad group, generate a budget ratio candidate value for each sampling, add noise and truncate it to the sampling upper and lower limits; summarize the budget ratio candidate values of all ad groups to generate an initial candidate budget plan; construct an objective optimization function based on the initial candidate budget plan, combined with the predicted click-through rate, the ratio of the conversion times to the click times of each historical ad group, and the recommended willingness score feedback by users, maximize the weighted revenue, and add an L2 regularization term to suppress the drastic fluctuation of the budget, adjust the initial candidate budget plan through gradient descent, and preferentially improve the high-potential ad groups (such as the groups with a large click-through rate improvement space but a low current ratio) to obtain an optimized candidate budget plan; encode the optimized candidate budget plan into a multi-armed action (such as [0.4, 0.3, 0.3] indicating allocating 40% to ad group A); through the dynamic exploration and exploitation balance, in the initial high-exploration stage (such as time t < 500 seconds), calculate the distribution standard deviation of the budget ratio candidate value of each ad group, set a standard deviation threshold, when the standard deviation is greater than the standard deviation threshold, mark it as low confidence, and preferentially select the optimized candidate budget plan with low confidence (such as the 10%-20% with a large fluctuation range), in the later high-exploitation stage (such as time t ≥ 500 seconds), select the stable plan with the highest revenue (such as the 15% with excellent long-term performance) to generate the final candidate budget allocation plan (such as ad group A = 40%, B = 30%, C = 30%, and the remaining 10% is reserved as emergency funds).

[0061] The cognitive load module, according to the candidate budget allocation plan, matches the high-density and high-emotion areas, adjusts the multi-modal stimulus parameters, and combines the user's facial micro-expressions to generate the cognitive load index, including the following steps,

[0062] Based on the final candidate budget allocation plan, combined with the total budget amount (such as 100,000 yuan per day) and the high-priority threshold (such as the click-through rate prediction > 8%), generate a dynamic delivery strategy.

[0063] Further explanation, the logic of generating the dynamic delivery strategy is as follows:

[0064] Sort the ad groups in descending order of the predicted click-through rate, preferentially allocate the budget to the high-potential groups (such as ad group A), and combine the four-channel environment map (crowd density, temperature, emotion index) to match the high-budget ad groups to the high-density and emotionally active areas (such as the mall entrance), output the dynamic strategy including the ad group ID, delivery area, budget ratio, and display frequency, and reserve 10% of the emergency funds for sudden traffic scenarios.

[0065] It should be noted that the high - priority threshold is determined by analyzing the distribution of click - through rates in historical advertising data and using statistical quantiles (such as the 75th or 80th percentile).

[0066] Based on the dynamic placement strategy and the three - dimensional population density heat map, the A* algorithm is used to plan the moving path of the advertising terminal, minimizing the path cost (distance + time penalty) and preferentially covering high - density areas. At the same time, a turning - radius constraint (increasing the cost when less than 1.5 meters) is set for obstacle avoidance; the path is updated every 5 seconds according to the real - time population distribution. If the density in a certain area drops by 30%, it is automatically re - planned, and the optimal path coordinate sequence and LED parameter instructions (such as brightness, color temperature) are output to guide the terminal device to perform dynamic display; according to the LED parameter instructions and the user's emotional activity index (0 - 1), the multi - modal stimulation parameters are adjusted in real - time, as follows:

[0067] When the emotional index is greater than 0.8, the brightness is adjusted to 1000 cd / m 2 (to attract attention), and when it is less than 0.3, it is reduced to 500 cd / m 2 (to reduce interference);

[0068] When the emotion is high, the color temperature is switched to the warm color tone (4000K, creating a warm feeling), and when the emotion is low, the color temperature is switched to the cold color tone (6500K, enhancing the sense of technology);

[0069] When the emotional index is greater than 0.6, the brand theme song is played (volume 70 dB), and when it is in a low - emotion period, promotional information is played (volume 30 dB).

[0070] According to the results of multi - modal stimulation adjustment, OpenCV (open - source computer vision library) is used to analyze the pupil position and fixation duration, and the attention concentration index is calculated (attention concentration index = fixation duration / exposure time × 100%); combined with the EDA signal (skin conductance change) and the MFN index (micro - expression complexity) for weighted fusion to generate the cognitive load index.

[0071] The sensitivity analysis module analyzes the cognitive load scenarios of user interaction data, optimizes the weights of the core advertising group through the regret - value mechanism and the Shapley value evaluation, and generates the final dynamic placement plan, including the following steps

[0072] According to the cognitive load index, the user's real - time interaction data (click - through rate, dwell time, user ID) is divided into three types of scenarios: high cognitive load, medium cognitive load, and low cognitive load, and a causal graph model is constructed, as follows

[0073] Based on historical user interaction data, set the high cognitive load threshold to 0.7 and the low cognitive load threshold to 0.3. When the cognitive load index is greater than 0.7, it represents high cognitive load, indicating that the user is in a high-pressure state (such as during the peak of a mall promotion). When the cognitive load index is between 0.3 and 0.7 (including 0.3 and 0.7), it represents medium cognitive load, indicating that the user is in a normal browsing environment. When the cognitive load index is less than 0.3, it represents low cognitive load, indicating that the user is in a relaxed state (such as in a rest area); Attach a scenario label (such as "high load") to each user interaction record and bind it to the ad group ID to generate a dataset with scenario markers; Define the direct effect variables as the immediate click-through rate and conversion rate after ad exposure, and the indirect effect variables as brand search volume and social media mentions; Based on historical user data, establish a direct causal path for ad group exposure, click-through rate, and conversion rate, and add an indirect path for ad exposure, increased brand search volume, and long-term conversion rate growth; Group by the ad group ID and the scenario labels in the dataset with scenario markers, and statistically calculate the growth rate of brand search volume triggered by the average click-through rate and conversion rate in each scenario. Through Granger causality testing, quantify the contribution ratio of direct effects and indirect effects, and construct a two-dimensional matrix of the ad group effect-cognitive load matrix, where the horizontal axis is the ad group ID and the vertical axis is the cognitive load index interval (high / medium / low), and the corresponding average click-through rate and conversion rate of the scenario are filled in the cells;

[0074] According to the two-dimensional matrix of cognitive load index - effect, quantify the sensitivity of the ad group to changes in the cognitive load index, calculate the sensitivity coefficient, and the expression is:

[0075]

[0076] where G represents the sensitivity coefficient, and CTR l represents the click-through rate in the low cognitive load scenario, and CTR h represents the click-through rate in the high cognitive load scenario.

[0077] Arrange the ad groups in descending order of the sensitivity coefficient to generate a cognitive load sensitivity ranking (such as ad group A > D > B > C), and the specific grading rules are as follows:

[0078] High-sensitivity group (sensitivity > 70%) needs to be optimized first (such as adjusting the placement strategy); Medium-sensitivity group (40% - 70%) for routine monitoring; Low-sensitivity group (< 40%) can ignore the impact of the cognitive load index.

[0079] Integrate the cognitive load sensitivity ranking, dynamic delivery strategy, and historical optimal strategy data (such as the historical best click-through rate being 12%), complete data alignment and parameter initialization, generate a structured strategy dataset, and set the initial global optimal benchmark click-through rate equal to 12% and the total cumulative regret value equal to 0; synchronize the real-time click-through rate and the structured strategy dataset, calculate the current revenue and the global optimal revenue, and generate a daily revenue comparison table.

[0080] Furthermore, the expression for the current revenue is:

[0081] μ(a t ) = C cur ×B bud ;

[0082] where μ(a t ) represents the revenue of ad group a at time t, a represents the ad group identifier, t represents the time dimension, C cur represents the current click-through rate of the ad group, and B bud represents the current budget proportion of the ad group in the total budget;

[0083] maxμ(a) = C opt ×B bud ;

[0084] where maxμ(a) represents the global optimal revenue of ad group a among all ad groups, and C opt represents the optimal click-through rate of the ad group.

[0085] Based on the historical ad group performance data, analyze the normal fluctuation range, set the regret value threshold, and distinguish between strategy failure and normal fluctuation; according to the daily revenue comparison table, calculate the single-day regret value (the single-day regret value is equal to the global optimal revenue minus the current revenue). If the single-day regret value is greater than the regret value threshold, mark it as a strategy failure; accumulate the regret values according to the ad group ID, and combine the cognitive load sensitivity ranking to sort according to the comprehensive score (cumulative regret value × (1 - sensitivity coefficient)) to generate a regret value ranking.

[0086] Based on the core business objectives of ad delivery (such as revenue, conversion rate, user satisfaction), combine the user click-through rate, conversion rate, and net promoter score reflecting user satisfaction to define a comprehensive effect function, and the expression is:

[0087] v(S) = ∑ a∈S (CTR a ×w CTR +CVR a ×w CVR +NPS a ×w NPS );

[0088] Among them, v(S) represents the comprehensive effect value of the advertising group subset S, and CTR a represents the click-through rate of advertising group a, and w CTR represents the weight coefficient of the click-through rate, and CVR a represents the conversion rate of advertising group a, and w CVR represents the weight coefficient of the conversion rate, and NPS a represents the net promoter score of advertising group a, and w NPS represents the weight coefficient of the net promoter score.

[0089] Based on the comprehensive effect value of the advertising group subset S, traverse all advertising group subsets, distinguish the direct improvement of the click-through rate / conversion rate brought by the advertising group after joining the subset and the effects triggered by the advertising group through indirect paths such as brand search volume, and calculate the Shapley value of each advertising group. The expression is:

[0090]

[0091] Among them, φ m represents the Shapley value of advertising group m, v(S∪{m}) represents the comprehensive effect value when the advertising group subset S contains advertising group m, and (S∪{m})-v(S) represents the contribution value of advertising group m to the effect of the advertising group subset S. If the difference is positive, it means that m has a positive contribution to the effect. M represents the number of advertising groups.

[0092] Combine the Shapley value results with the regret value ranking for comprehensive sorting to identify the core advertising groups with high contributions; based on the core advertising group identification results, adjust the final candidate budget allocation. For high-contribution groups, increase the budget proportion, and for low-contribution groups, cut the budget proportion. At the same time, constraint checks should be carried out to ensure that the sum of the proportions of each group after adjustment is 100%, and generate the final dynamic delivery plan.

[0093] In summary, through the combination of the distributed file system and Kafka stream processing, the present invention realizes the efficient loading of the multi-modal model and the automation of historical data cleaning. By setting the adaptive mechanisms of the global benchmark reward mean, the specific variance matrix, and the noise standard deviation, and combining the variational inference framework to update the parameters of the real-time click-through rate feedback, the limitation of relying on artificial prior assumptions is solved; the dynamic learning rate decay strategy balances the model convergence speed and the overfitting risk; through the spatio-temporal alignment of the four-channel environmental map and the user behavior data, and combining the multi-head attention mechanism of Transformer, the non-linear correlation modeling of environmental parameters and advertising features is realized. The objective function of Thompson sampling and L2 regularization enables the budget allocation strategy to dynamically respond to the cognitive load scenario; by quantifying the response differences of advertising groups to cognitive load with the sensitivity coefficient and combining the Shapley value to evaluate the marginal contribution of advertising groups, the accurate identification of core advertising groups and budget tilt are realized.

[0094] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit them. Although the present invention has been described in detail with reference to the preferred embodiments, those of ordinary skill in the art should understand that the technical solutions of the present invention can be modified or equivalently replaced without departing from the spirit and scope of the technical solutions of the present invention, and they should all be covered within the scope of the claims of the present invention.

Claims

1. An intelligent evaluation system for monitoring the effect of advertising placement, characterized in that: including a model loading module that loads a bidirectional pre-trained language model and a multi-modal fusion model, synchronizes advertisement metadata to generate a structured feature matrix, and initializes Bayesian network parameters; a feature fusion module that collects user behavior, biometric, and environmental data in real time, generates a four-channel environmental map, and splices it with the structured feature matrix and inputs it into a Transformer encoder to form a joint feature vector; a click-through rate prediction module that calculates an immediate reward based on user click feedback, updates the Bayesian network parameters through variational inference, and combines the joint feature vector to predict the click-through rate; a budget allocation module that generates an initial candidate budget plan using Thompson sampling based on the predicted click-through rate, constructs an objective function, and generates a final candidate budget allocation plan through a dynamic exploration-exploitation strategy; a cognitive load module that matches high-density and high-emotion regions according to the candidate budget allocation plan, adjusts multi-modal stimulus parameters, and combines the user's facial micro-expression to generate a cognitive load index; a sensitivity analysis module that analyzes the cognitive load scenario of user interaction data, optimizes the weights of the core advertisement group through a regret value mechanism and a Shapley value evaluation, and generates a final dynamic delivery plan.

2. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, wherein: The generation of the structured feature matrix refers to loading the bidirectional pre-trained language model and the multi-modal fusion model into an inference optimization engine, using the successfully loaded bidirectional pre-trained language model and multi-modal fusion model to perform data cleaning and verification on historical advertisement metadata, and generating a structured feature matrix.

3. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, characterized in that: The initialization of the Bayesian network parameters refers to setting the Bayesian hierarchical network parameters, constructing a Bayesian neural network using the PyTorch framework, initializing the weights with a uniform distribution, performing forward propagation on the structured feature matrix, and conducting batch inference tests on the constructed Bayesian neural network.

4. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, characterized in that: The formation of the joint feature vector includes the following steps collecting the user's movement trajectory, user location, user physiological signals, and three-dimensional population distribution; obtaining temperature data to generate a two-dimensional temperature heat map, and superimposing the three-dimensional population distribution to form a four-channel environmental map; extracting the historical click-through rate mean, budget allocation ratio, and logarithm of exposure volume in the structured feature matrix, inputting them into the Transformer encoder, and dynamically allocating weights through a multi-head attention mechanism to form a joint feature vector.

5. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, wherein: The update of the Bayesian network parameters includes the following steps obtaining user click feedback in real time and calculating the immediate reward; using the variational inference framework, setting the initialized Bayesian network parameters as the prior distribution, performing likelihood function modeling, and adjusting the Bayesian network parameters by establishing a posterior optimization objective; using an Adam optimizer to perform gradient descent iterative updates to obtain the updated Bayesian network parameters.

6. The intelligent evaluation system for advertising placement effect monitoring according to claim 1, wherein: The prediction of the click-through rate refers to inputting the updated Bayesian network parameters and the joint feature vector into the Bayesian neural network for forward propagation, and generating a predicted click-through rate by calculating the conditional expectation value of the advertisement group at the current moment.

7. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, characterized in that: The generation of the final candidate budget allocation plan includes the following steps generating an initial candidate budget plan for each advertisement group using the Thompson sampling algorithm based on the predicted click-through rate; Construct an objective optimization function to maximize the weighted revenue and add an L2 regularization term to suppress drastic fluctuations in the budget; Adjust the initial candidate budget plan through gradient descent to generate the final candidate budget allocation plan.

8. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, characterized in that: The adjustment of the multi-modal stimulation parameters includes the following steps Based on the final candidate budget allocation plan, combine the total budget amount and set a high-priority threshold to generate a dynamic delivery strategy; Based on the dynamic delivery strategy and the three-dimensional population density heat map, minimize the path cost and preferentially cover high-density areas to generate an optimal path coordinate sequence and LED parameter instructions; According to the LED parameter instructions and the user mood activity index, adjust the multi-modal stimulation parameters in real time.

9. The intelligent evaluation system for advertising delivery effect monitoring according to claim 1, characterized in that: The generation of the cognitive load index refers to analyzing the pupil position and fixation duration using OpenCV according to the adjustment results of multi-modal stimulation, calculating the attention concentration index, and performing weighted fusion in combination with the EDA signal.

10. The intelligent evaluation system for monitoring the advertising delivery effect according to claim 1, characterized in that: The generation of the final dynamic delivery plan refers to dividing the user interaction scenarios according to the cognitive load index and constructing a causal model to obtain the advertisement group effect-cognitive load matrix, calculating the sensitivity coefficient and ranking the sensitivity, combining the dynamic delivery strategy, historical optimal data and regret value mechanism, evaluating the contribution degree of the advertisement group using the Shapley value, and dynamically adjusting the budget allocation to optimize the weight of the core advertisement group.

Images