Marketing flow site intelligent screening method based on big data analysis

Abstract

The application discloses a marketing flow site intelligent screening method based on big data analysis, and relates to the technical field of site intelligent screening, comprising the following steps: collecting multi-source data of candidate sites and constructing feature vectors; based on the feature vectors, quantifying the individualized response of each candidate site under a preset marketing strategy through a heterogeneous causal effect estimation method to obtain an initial causal effect score; constructing a graph structure, adopting a graph neural network method to fuse the influence of neighborhood facilities, generating a site-neighborhood fusion vector, and correcting the initial causal effect score; obtaining risk measurement indexes, combining the corrected causal effect score, and generating a final optimal score of each candidate site; sorting all candidate sites based on the final optimal score, and selecting the best optimal site; the application solves the problems that marketing site selection depends on single data, causal effect evaluation is inaccurate, and neighborhood influence and risk control are lacking.

Description

Technical Field

[0001] This invention relates to the field of intelligent venue screening technology, and more specifically, to a method for intelligent screening of marketing traffic venues based on big data analysis. Background Technology

[0002] With the rapid development of digital marketing, businesses are increasingly emphasizing marketing strategy optimization based on precise data analysis. The effectiveness of marketing campaigns is closely related to the traffic and potential customer characteristics of the selected marketing venues. Currently, businesses primarily rely on manual experience, historical data statistics, or single-indicator evaluations when selecting marketing venues, such as screening based on indicators like pedestrian traffic, customer density, or shop size. These methods have the following limitations in practical applications: Traditional site selection methods often rely on single data sources, such as historical sales data or pedestrian traffic statistics, lacking multi-dimensional data integration of the site and its surrounding environment. For example, they fail to adequately consider multi-source information such as neighborhood facilities, transportation convenience, and consumer characteristics, resulting in insufficient scientific rigor and accuracy in site selection decisions. Furthermore, existing technologies often employ correlation analysis or simple regression models to evaluate marketing effectiveness, but these methods struggle to effectively distinguish the true causal effects of marketing strategies on site traffic and sales performance. Due to significant differences in the characteristics of candidate sites, simple average effect evaluations may introduce biases, failing to achieve personalized and accurate site recommendations. To address these issues, this invention proposes a solution. Summary of the Invention

[0003] To overcome the aforementioned deficiencies of the prior art, embodiments of the present invention provide an intelligent screening method for marketing traffic venues based on big data analysis. By constructing feature vectors based on multi-source data, employing heterogeneous causal effect estimation combined with graph neural networks, and performing multi-objective optimization, the method solves the problems of marketing venue selection relying on single data, inaccurate causal effect assessment, and lack of neighborhood influence and risk control.

[0004] To achieve the above objectives, the present invention provides the following technical solution: This application provides a method for intelligent screening of marketing traffic venues based on big data analysis. The method includes: collecting multi-source data of candidate venues and constructing feature vectors; based on the feature vectors, quantifying the personalized response of each candidate venue under a preset marketing strategy using a heterogeneous causal effect estimation method to obtain an initial causal effect score; constructing a graph structure of the candidate venues and their neighboring facilities, using a graph neural network method to fuse the influence of neighboring facilities, generating a venue-neighborhood fusion vector, and correcting the initial causal effect score using the venue-neighborhood fusion vector to obtain a corrected causal effect score; obtaining a risk measurement index characterizing the marketing risk of the venue, and combining it with the corrected causal effect score to generate a final optimal score for each candidate venue; ranking all candidate venues based on the final optimal score, and selecting the best optimal venue.

[0005] In one embodiment, multi-source data of candidate sites are collected and feature vectors are constructed. Specifically, the multi-source data is preprocessed; preliminary features are constructed based on the preprocessed data, and the correlation coefficient between each preliminary feature and the target indicator is obtained; preliminary features with correlation coefficients higher than a preset correlation threshold are retained to obtain a highly correlated feature set; the highly correlated feature set is dimensionality reduced to form the final feature vector of the candidate sites.

[0006] In one embodiment, based on the feature vector, a heterogeneous causal effect estimation method is used to quantify the personalized response of each candidate site under a preset marketing strategy to obtain an initial causal effect score. Specifically, the candidate sites are initially divided into a treatment group and a control group based on whether they have received historical marketing strategy interventions. The final feature vector is used to balance and correct the treatment group and the control group. Based on the heterogeneous causal effect estimation method, the balanced and corrected treatment group and the control group are trained to construct a heterogeneous causal effect model and output the initial causal effect score of each candidate site. Several resampled subsets are generated based on the initial causal effect score and the corresponding training dataset. The model is retrained based on each resampled subset to obtain the predicted initial causal effect scores under multiple resampled subsets and obtain confidence intervals. The initial causal effect score and its confidence interval for each candidate site are output.

[0007] In one embodiment, the final feature vector is used to balance the treatment group and the control group. Specifically, the final feature vector of each candidate site is used as the input feature to obtain the treatment indicator variable and the observation effect variable for each site. Based on the feature vectors of the treatment group and the control group, a propensity score model is constructed and the propensity score of each candidate site is output. For each site in the treatment group, the site with the smallest propensity score error is selected from the control group for matching. After matching, a balanced subset of the treatment group and the control group is generated and adjusted until all features meet the balance requirements. The balanced datasets of the treatment group and the control group are output.

[0008] In one embodiment, a heterogeneous causal effect model is constructed by training balanced and corrected treatment and control groups using a heterogeneous causal effect estimation method. Specifically, the treatment and control datasets are used as training sets, which include the final feature vector, treatment indicator variable, and observed effect variable for each candidate site. The final feature vector is used as input features, the treatment indicator variable as an intervention marker, and the observed effect variable as a dependent variable. A potential control outcome prediction model is trained on the treatment group data to estimate the potential effect of the treatment group sites without intervention. A potential treatment outcome prediction model is trained on the control group data to estimate the potential effect of the control group sites with intervention. The residual causal effect of the treatment group is calculated for the treatment group sites based on the output of the potential control outcome prediction model and the dependent variable of the treatment group sites. The residual causal effect of the control group sites is calculated for the control group sites based on the output of the potential treatment outcome prediction model and the dependent variable of the control group sites. Based on the residual causal effects of the treatment group and the control group, a heterogeneous causal effect model is constructed, outputting the initial causal effect score for each candidate site under a preset marketing strategy.

[0009] In one embodiment, a graph structure is constructed with candidate sites as nodes and related facilities in their neighborhood as auxiliary nodes. The attribute and relationship information of the neighborhood facilities are fused to generate a site-neighborhood fusion representation. The initial causal effect score is then corrected to obtain a corrected causal effect score. Specifically, the following steps are taken: facility information within the neighborhood of the candidate site is collected; based on the facility information between the candidate site and its neighboring facilities, a graph structure is constructed, where the candidate site is a node in the graph and the neighboring facilities are auxiliary nodes; the graph structure is used to interact and fuse the features between the candidate site node and the neighboring facility node using the message passing mechanism of a graph neural network, to obtain a site-neighborhood fusion vector. Based on the site-neighborhood fusion vector, the initial causal effect score of each candidate site is corrected to obtain the corrected causal effect score.

[0010] In one embodiment, the initial causal effect score of each candidate site is corrected based on the site-neighborhood fusion vector to obtain the corrected causal effect score. Specifically, the initial causal effect score is aligned with the site-neighborhood fusion vector; based on the alignment result, the initial causal effect score and the site-neighborhood fusion vector are weighted and fused to obtain the fused causal effect score of the candidate site; and the fused causal effect score is nonlinearly transformed to obtain the corrected causal effect score.

[0011] In one embodiment, a risk metric is obtained and combined with the corrected causal effect score to generate a final optimal score for each candidate site. Specifically, the risk metric is collected and standardized to obtain a standardized risk metric. A multi-objective optimization problem is constructed based on the corrected causal effect score and the standardized risk metric, which includes maximizing the corrected causal effect score and minimizing the standardized risk metric. A Pareto front search model is constructed based on the multi-objective optimization problem. The weights of the Pareto front search model are obtained through an adaptive weight adjustment mechanism, and an optimization algorithm is used to iteratively solve the Pareto front search model. The Pareto optimal solution set for the candidate sites is output to obtain the final optimal score for each candidate site.

[0012] In one embodiment, the weights of the Pareto frontier search model are obtained through an adaptive weight adjustment mechanism. Specifically, this involves: obtaining a preset initial weight vector for the Pareto frontier search model; calculating the comprehensive score of each candidate site in the current iteration based on the initial weight vector; obtaining the contribution percentages of the corrected causal effect score and the standardized risk metric to the comprehensive score; comparing the contribution percentages of the corrected causal effect score and the standardized risk metric to the comprehensive score with the preset target weight vector to obtain the deviation; and adjusting the initial weight vector based on the deviation using a gradient adjustment method to output the weights of the Pareto frontier search model.

[0013] In one embodiment, all candidate sites are sorted based on the final selection score, and the best preferred site is selected. Specifically, the final selection scores are sorted in descending order to generate a candidate site ranking list; according to the preset filtering conditions of the ranking list, the candidate sites at the top are selected as preferred sites, the filtering conditions include a preset number of top candidate sites and sites that reach a preset score threshold; and a set of preferred sites is output.

[0014] As can be seen from the above technical solutions, the embodiments of this application have the following advantages: By integrating multi-source data to construct candidate site feature vectors, and combining heterogeneous causal effect estimation to accurately quantify the marginal improvement effect of each site under a preset marketing strategy, this approach utilizes graph neural networks to fuse the influence of neighboring facilities for causal score correction. Finally, multi-objective optimization and risk measurement are combined to generate a final optimal score, achieving scientific ranking and selection of candidate sites. Compared to traditional methods relying on single data or experience-based judgments, this approach fully leverages multi-dimensional information, eliminates confounding biases, captures individualized response differences and neighborhood effects, and balances benefits and risks. It not only significantly improves the accuracy and robustness of marketing site selection but also optimizes resource allocation and decision-making efficiency, providing a reliable scientific basis for precision marketing and market expansion. Attached Figure Description

[0015] Figure 1 This is a schematic diagram of a marketing traffic venue intelligent screening method based on big data analysis, provided for an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0017] Reference Figure 1 As shown in the diagram, this invention provides a flowchart of a marketing traffic venue intelligent screening method based on big data analysis, which includes the following steps: S1, collect multi-source data of candidate sites, preprocess the multi-source data and construct feature vectors.

[0018] In this embodiment, multi-source data of candidate sites is collected, and the multi-source data is preprocessed and features are constructed, specifically as follows: Multi-source data of candidate sites is collected from multiple data sources, including mobile terminal location data, transaction and consumption data, geographic information data and surrounding points of interest (POI) data, transportation data, historical marketing campaign and effect data, social media and event calendar data; The multi-source data is preprocessed, including denoising, deduplication, timestamp and spatial coordinate alignment, and outlier removal, to ensure the comparability of each source data in both time and space dimensions. Preliminary features are constructed based on the preprocessed data, and the correlation coefficients between each preliminary feature and the target indicator are obtained; Preliminary features with correlation coefficients higher than a preset correlation threshold are retained to obtain a highly correlated feature set; The dimensionality of the highly relevant feature set is reduced using factor analysis to form the final feature vector of the candidate sites.

[0019] It should be noted that preliminary features refer to quantitative indicators generated from preprocessed multi-source data according to business rules or statistical methods to characterize the attributes of candidate sites. Specifically, these may include pedestrian flow characteristics (such as average daily / hourly pedestrian flow, peak and off-peak pedestrian flow differences), consumption characteristics (such as average transaction value, total transaction amount, repeat purchase rate, sales growth rate), spatial characteristics (such as surrounding POI density, competitor density, and transportation accessibility indicators), time series characteristics (such as holiday consumption growth, pedestrian flow changes during historical marketing campaigns), and social / event characteristics (such as social media mentions, and attention during hot events within a time window).

[0020] Target metrics are core outcome variables used to measure the effectiveness of marketing strategies or the value of a venue. They serve as reference standards when selecting features and include conversion rate metrics (such as the proportion of visitors to the total potential customer flow during the marketing campaign, the number of registrations / applications) and consumption metrics (such as average order value, total transaction amount, sales growth rate, and ROI (return on investment)).

[0021] After multi-source data collection, preprocessing, preliminary feature construction, correlation screening, and factor analysis dimensionality reduction, the generated candidate site final feature vector not only retains the most informative features that are most explanatory to marketing effectiveness, but also reduces redundancy and computational complexity through dimensionality reduction and compression. At the same time, it integrates multi-source information such as pedestrian flow, consumption, space, time series, and social events, making the features highly interpretable and scalable, and providing a stable and reliable input basis for subsequent heterogeneous causal effect modeling, thereby enabling more accurate evaluation and optimization of the marketing potential of candidate sites.

[0022] S2, based on the feature vector, the personalized response of each candidate site under the preset marketing strategy is quantified by the heterogeneous causal effect estimation method to obtain the initial causal effect score.

[0023] In this embodiment, based on the feature vector, a heterogeneous causal effect estimation method is used to quantify the personalized response of each candidate site under the preset marketing strategy, and an initial causal effect score is obtained, specifically as follows: Based on whether each candidate site had received historical marketing strategy interventions, they were initially divided into a treatment group (received intervention) and a control group (not received intervention). The final eigenvectors are used to balance and correct the treatment and control groups to ensure that the distributions of the two groups are similar in the feature space, thereby reducing selection bias and improving the accuracy of heterogeneous causal effect estimation. Based on the heterogeneous causal effect estimation method, the treatment group and the control group after balance correction are trained to construct a heterogeneous causal effect model; The initial causal effect score for each candidate site is obtained by outputting the heterogeneous causal effect model. Obtain the initial causal effect score and corresponding training dataset for each candidate site. Perform sampling with replacement on the training dataset to generate several resampled subsets, where the number of samples in each resampled subset is the same as that in the original training dataset. For each resampled subset, the model is retrained using a heterogeneous causal effect model to obtain the initial causal effect scores predicted for multiple resampled subsets. Confidence intervals are then obtained using the percentile method. The initial causal effect scores predicted by the resampled subsets are sorted in ascending order, and the [number]th [presumably a specific score] is taken. quantiles and quantiles as The confidence interval boundary, where, This is the complement of the confidence level, with values ​​between 0 and 1. This represents the quantile position corresponding to the lower limit of the confidence interval. This represents the quantile position corresponding to the upper limit of the confidence interval. This represents the overall confidence level of the confidence interval; Output the initial causal effect score and its confidence interval for each candidate site.

[0024] It's important to note that by obtaining the initial causal effect score for each candidate venue under the preset marketing strategy, the actual marginal improvement that different venues might experience when receiving the same marketing strategy can be quantified, effectively distinguishing which venues have greater response potential. Compared to site selection methods that rely on experience or simple traffic indicators, this causal effect-driven screening avoids misjudgments caused by spurious correlations or data noise, ensuring that the selected high-quality venues are those that truly generate significant effects due to the marketing strategy. This not only improves resource allocation efficiency and reduces the cost of ineffective advertising but also provides a scientific basis for subsequent marketing strategy optimization, achieving more precise and efficient market expansion.

[0025] Furthermore, the final feature vector is used to perform a balance correction between the treatment group and the control group, specifically as follows: The final feature vector of each candidate site is used as the input feature to obtain the processing indicator variable (1 indicates acceptance of the marketing strategy, 0 indicates non-acceptance) and the observed effect variable (such as the number of visitors, transaction amount or conversion rate) for each site. Based on the feature vectors of the treatment group and the control group, a propensity scoring model is constructed to predict the probability of each candidate site accepting marketing intervention, and the propensity score of each candidate site is output. Among them, the propensity score model can use logistic regression, gradient boosting tree, random forest or neural network, and use cross-validation and regularization strategies during training to prevent overfitting; For each treatment group's site, one or more sites with the smallest tendency score error were selected from the control group for matching; After matching is completed, a balanced subset of the treatment group and the control group is generated. The standardized mean difference of each feature is calculated for the matched treatment group and the control group. If the standardized mean difference exceeds the preset standard threshold, the matching strategy is adjusted or rematching is performed until all features meet the balance requirements. Output balanced datasets for the treatment and control groups, including site unique identifiers, final feature vectors, treatment indicator variables, and observation effect variables.

[0026] It's important to note that the core purpose of balancing the treatment and control groups is to eliminate confounding biases caused by differences in the distribution of candidate site characteristics, ensuring that the causal effect estimation results are more reliable and interpretable. By using methods such as propensity score matching, the distributions of key features in the treatment and control groups are made more consistent, thus ensuring that the differences between the two groups mainly stem from the marketing strategy itself, rather than the external environment or inherent site attributes. This not only improves the accuracy of causal effect estimation and reduces spurious correlation interference, but also enhances the model's generalization ability to different scenarios, enabling the final causal effect score to more realistically reflect the actual marginal improvement effect of the marketing strategy on the site.

[0027] Furthermore, based on the heterogeneous causal effect estimation method, the balanced corrected treatment group and control group are trained to construct a heterogeneous causal effect model, specifically: The balanced datasets of the treatment and control groups were obtained and used as the training set, including the final feature vector, treatment indicator variable, and observation effect variable for each candidate site; The final feature vector is used as the input feature. In this study, indicator variables were treated as intervention markers, and observed effect variables were treated as dependent variables. A potential control outcome prediction model is trained on the treatment group data to estimate the potential effects of the treatment group sites without intervention, and the output is the predicted effect without intervention. The training objects for the potential control outcome prediction model are the final feature vectors and observed effect variables of the treatment group sites (which have accepted the marketing strategy); A potential treatment outcome prediction model is trained on control group data to estimate the potential effects of intervention on the control group sites, and the predicted intervention effect is output. The training objects for the potential treatment outcome prediction model are the final feature vectors and observed effect variables of the control group site (which did not receive intervention); Among them, the potential control outcome prediction model and the potential treatment outcome prediction model are both trained using regression trees. During the training process, cross-validation is used to select hyperparameters and regularization strategies to prevent overfitting. At the same time, ensemble methods can be combined to improve the prediction robustness. The residual causal effect of the treatment group was calculated by combining the output of the prediction model based on the potential control results with the dependent variable of the treatment group site. The residual causality of the treatment group The specific calculation formula is as follows:

[0028] In the formula, The dependent variable for the treatment group site is the observed effect variable for the treatment group site.

[0029] The residual causal effect of the control group site was calculated by combining the output of the potential treatment outcome prediction model with the dependent variable of the control group site. The residual causality of the control group The specific calculation formula is as follows:

[0030] In the formula, The dependent variable is the control group site, which is the observed effect variable of the control group site.

[0031] Based on the residual causal effects of the treatment group and the control group, a heterogeneous causal effect model is constructed to output the initial causal effect score (expected marginal improvement value) of each candidate site under the preset marketing strategy, and the final feature vector of the candidate site is used as input for training.

[0032] The specific calculation formula for the heterogeneous causal effect model is as follows:

[0033] In the formula, The initial causal effect score is given by w, which is a weighting coefficient used to balance the contribution of the residual causal effect between the treatment group and the control group. It can be determined by weighting according to the sample ratio of the treatment group and the control group.

[0034] It's important to note that the advantage of obtaining the initial causal effect score using the above method lies in its ability to not only utilize the actual observed effects of the treatment and control groups but also supplement the estimation of unobserved scenarios through a potential outcome prediction model, thereby minimizing bias caused by a single data source. By employing residual causal effects, the individualized response differences of venues under the marketing strategy can be captured, and a weighted mechanism can balance the contributions of different group samples, making the final output initial causal effect score more objective, robust, and interpretable. This method can accurately quantify the marginal improvement effect of each venue under a given marketing strategy, providing a scientific basis and reliable support for subsequent selection of high-quality venues.

[0035] S3. The candidate site and its neighboring facilities are constructed as a graph structure. The influence of the neighboring facilities is fused using a graph neural network method to generate a site-neighborhood fusion vector. The initial causal effect score is then corrected using the site-neighborhood fusion vector to obtain the corrected causal effect score.

[0036] In this embodiment, the candidate site and its neighboring facilities are constructed as a graph structure. A graph neural network method is used to fuse the influence of neighboring facilities to generate a site-neighborhood fusion vector. The initial causal effect score is then corrected using the site-neighborhood fusion vector to obtain the corrected causal effect score, specifically: Collect facility information within the neighborhood of the candidate site, including the proximity of the facility information to the candidate site in terms of geospatial or functional association; Based on the facility information between the candidate site and its neighboring facilities, a graph structure is constructed, where the candidate site is a node in the graph, the neighboring facilities are auxiliary nodes, and the proximity relationship between the candidate site and the facilities is an edge. The graph structure utilizes the message passing mechanism between nodes in a graph neural network to interact and fuse the features between candidate site nodes and neighboring facility nodes, thereby obtaining a site-neighborhood fusion vector. The graph neural network is a neural network with multiple message passing mechanisms. Each layer weights and fuses the features of the auxiliary nodes of the neighborhood facilities with the features of their corresponding candidate site nodes, thereby enhancing the neighborhood representation capability of the candidate site nodes layer by layer and obtaining the site-neighborhood fusion vector. Here, the features refer to the multi-dimensional data vector carried by each node in the graph structure (including candidate site nodes and auxiliary nodes of neighborhood facilities) to describe its attributes and state, such as the type, size, and historical passenger flow of the candidate site, as well as the quantified or encoded information such as the type, function, distance, and passenger attraction of the neighborhood facilities.

[0037] Based on the site-neighborhood fusion vector, the initial causal effect score of each candidate site is corrected to obtain the corrected causal effect score.

[0038] It should be noted that by constructing candidate sites and their surrounding facilities into a graph structure and employing a graph neural network with a multi-layer message passing mechanism for feature interaction and fusion, the complex geospatial and functional relationships between candidate sites and their surrounding facilities can be explicitly and deeply modeled. This process not only achieves the structured integration of multi-source neighborhood information but also adaptively captures multi-hop, nonlinear interactions between facilities through iterative message passing, thereby generating a high-order, condensed "site-neighborhood fusion vector." This site-neighborhood fusion vector effectively characterizes the comprehensive effect of the neighborhood environment on the site's marketing potential. This allows subsequent corrections to the initial causal effect score to no longer be based on simple rules or statistical correlations but on a deep semantic understanding of complex spatial dependencies, significantly improving the accuracy, interpretability, and refinement of marketing effect assessment and site selection decisions.

[0039] Furthermore, based on the site-neighborhood fusion vector, the initial causal effect score for each candidate site is corrected to obtain the corrected causal effect score, specifically: The initial causal effect score is aligned with the site-neighborhood fusion vector. The alignment process includes projecting the site-neighborhood fusion vector onto the same representation space as the initial causal effect score using a pre-trained projection matrix. Based on the alignment results, the initial causal effect score is weighted and fused with the site-neighborhood fusion vector to obtain the fusion causal effect score of the candidate site. The weighted fusion is a linear weighted summation, and the weights are learned by the neural network based on the final feature vector of the candidate site to reflect the differences in the importance of different neighborhood influences. A nonlinear transformation is performed on the fused causal effect score to obtain a corrected causal effect score. The nonlinear transformation adjusts the contribution of neighborhood influence to the final score through a sigmoid activation function.

[0040] It should be noted that by introducing a graph neural network, candidate sites and their neighboring facilities are modeled as a graph structure. During message passing and feature fusion, this approach can fully capture the potential impact of the neighborhood environment on candidate sites, achieving an organic combination of the initial causal effect score and the neighborhood representation. Compared to causal effect estimation that relies solely on the site's own attributes, this method can dynamically identify and quantify the importance of different neighborhood factors. After alignment, weighted fusion, and nonlinear transformation, the final corrected causal effect score more accurately reflects the site's performance in the actual environment, thereby improving the accuracy, robustness, and decision value of causal effect estimation.

[0041] S4. Obtain the risk metric and combine it with the corrected causal effect score to generate the final optimal score for each candidate site.

[0042] In this embodiment, risk measurement indicators are obtained and combined with the corrected causal effect score to generate the final optimal score for each candidate site, specifically: Risk metrics are collected and standardized risk metrics are obtained through Z-score standardization. These risk metrics include environmental risk, operational risk, and social impact risk. Among them, environmental risks are air / water pollution indices, operational risks are equipment failure rate and number of downtime days, and social impact risks are the number of community complaints and the probability of accidents. A multi-objective optimization problem is constructed based on the corrected causal effect score and the standardized risk metric. The first objective of the multi-objective optimization problem is to maximize the corrected causal effect score, and the second objective is to minimize the standardized risk metric. Based on the multi-objective optimization problem, a Pareto front search model is constructed. The Pareto front search model transforms the multi-objective optimization problem into searching for the Pareto optimal solution set by setting weights. The weights of the Pareto front search model are obtained through an adaptive weight adjustment mechanism, and the Pareto front search model is solved iteratively using an optimization algorithm. During the iteration process, the optimization is judged to be completed based on the convergence conditions, which include the change in the comprehensive score being less than a preset threshold or the number of iterations reaching the upper limit. Output the Pareto optimal solution set for the candidate sites, and obtain the final optimal score for each candidate site.

[0043] The Pareto front search model is an optimization modeling method for comprehensively evaluating and selecting candidate sites under multi-objective constraints. Its core idea is to systematically characterize the trade-off between "maximizing benefits" and "minimizing risks" without simply merging multiple objectives linearly. Specifically, the model uses the corrected causal effect score as the benefit objective and the Z-score-standardized environmental risk, operational risk, and social impact risk as cost objectives. First, by introducing weight vectors, the multi-objective problem is mapped into an iteratively solvable scalar objective function space, thus exploring the non-dominated solution regions among the objectives under different weight combinations. During the search process, an adaptive weight adjustment mechanism is adopted, dynamically adjusting the weights according to the distribution of the current solution set in the objective space, enabling the algorithm to cover a complete and uniform Pareto front, avoiding bias towards a single objective or getting trapped in local optima. Simultaneously, by setting a comprehensive score change threshold and a maximum number of iterations as convergence criteria, a balance is ensured between computational efficiency and solution stability during the search process. Ultimately, the Pareto optimal solution output by the model represents the optimal trade-off between causal benefit and multidimensional risk control for each candidate site, assuming that "one site is superior to another site in all objectives". The corresponding comprehensive evaluation result is used as the final selection score, thus providing a quantitative basis with sufficient interpretability and technical integrity for site decision-making.

[0044] Furthermore, the weights of the Pareto frontier search model are obtained through an adaptive weight adjustment mechanism, specifically: Obtain the initial weight vector preset by the Pareto frontier search model. The initial weight vector can be preset based on experience or obtained through historical data statistics to reflect the relative importance of causal effects and risks in the initial stage. Based on the initial weight vector, the comprehensive score of each candidate site in the current iteration is calculated. The comprehensive score is obtained by normalizing the corrected causal effect score and the standardized risk metric, and then linearly fusing them according to the initial weight. The comprehensive score is calculated using the following formula:

[0045] In the formula, For comprehensive scoring, The initial weights for the corrected causal effect scores, To correct the causal effect score, To standardize the initial weights of risk metrics, To standardize risk measurement indicators.

[0046] The contribution percentages of the corrected causal effect score to the overall score and the contribution percentages of the standardized risk measurement index to the overall score were obtained respectively. The percentage of the contribution of the corrected causal effect score to the overall score. The specific calculation formula is as follows:

[0047] The contribution percentage of the standardized risk metric to the overall score The specific calculation formula is as follows:

[0048] The deviation is obtained by comparing the contribution percentage of the corrected causal effect score to the comprehensive score and the contribution percentage of the standardized risk metric to the comprehensive score with the preset target weight vector. The deviation The specific calculation formula is as follows:

[0049] In the formula, To determine the expected target weights for the corrected causal effect scores, The expected target weights for standardized risk measurement indicators.

[0050] Based on the deviation, the initial weight vector is adjusted by the gradient adjustment method so that the weights more accurately reflect the actual impact of causal effects and risks in the comprehensive optimization in the next iteration. The iteration is repeated until the weight vector meets the convergence condition, which includes the weight change amplitude being lower than a preset change amplitude threshold. The gradient adjustment method for adjusting the initial weight vector can be understood as follows: using the comprehensive score in the current iteration, it is represented as the result of fusing the causal effect score and the risk index according to the current weight vector. Then, the gradient of the comprehensive score relative to the weight vector is obtained. Based on the gradient direction, the weight vector is adjusted so that the comprehensive score in the next iteration tends to strengthen the target with insufficient contribution. For example, if the contribution of the causal effect is lower than expected, it is updated in the direction of increasing the weight of the causal effect, and vice versa. By repeating this process through multiple iterations, the weight vector gradually approaches the optimal allocation that balances the causal effect and the risk target, thereby enabling the multi-objective optimization model to achieve reasonable optimization of the comprehensive score while maintaining the stability of the Pareto front solution. Output the weights of the Pareto frontier search model.

[0051] S5. Sort all candidate venues based on the final selection score and select the preferred venue.

[0052] In this embodiment, all candidate sites are sorted based on the final selection score, and the preferred site is selected, specifically as follows: The final selection scores are sorted in descending order to generate a candidate site ranking list; According to the preset filtering conditions of the candidate site ranking list, the candidate sites at the top are selected as preferred sites. The filtering conditions include a preset number of top candidate sites and sites that reach a preset score threshold. Output the best set of preferred sites to support subsequent site decisions or deployments.

[0053] The above formulas are all dimensionless calculations. The formulas are derived from software simulations based on a large amount of collected data to obtain the most recent real-world results. The preset parameters in the formulas are set by those skilled in the art according to the actual situation.

[0054] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented using software, the above embodiments can be implemented, in whole or in part, in the form of a computer program product.

[0055] Those skilled in the art will recognize that the modules and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0056] In addition, the functional modules in the various embodiments of this application can be integrated into one processing module, or each module can exist physically separately, or two or more modules can be integrated into one module.

[0057] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0058] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for intelligently filtering marketing traffic venues based on big data analysis, characterized in that, Includes the following steps: Collect multi-source data of candidate sites and construct feature vectors; Based on the feature vector, the heterogeneous causal effect estimation method is used to quantify the personalized response of each candidate site under the preset marketing strategy and obtain the initial causal effect score. Candidate sites and their neighboring facilities are constructed as a graph structure. The influence of neighboring facilities is fused using a graph neural network method to generate a site-neighborhood fusion vector. The initial causal effect score is then corrected using the site-neighborhood fusion vector to obtain the corrected causal effect score. Obtain risk metrics that characterize the marketing risk of a venue, and combine them with the adjusted causal effect score to generate the final optimal score for each candidate venue; All candidate venues are ranked based on the final selection score, and the best venue is selected.

2. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 1, characterized in that, The process of collecting multi-source data from candidate sites and constructing feature vectors involves: The multi-source data is preprocessed; Preliminary features are constructed based on the preprocessed data, and the correlation coefficients between each preliminary feature and the target indicator are obtained; Preliminary features with correlation coefficients higher than a preset correlation threshold are retained to obtain a highly correlated feature set; The dimensionality of the highly relevant feature set is reduced to form the final feature vector of the candidate sites.

3. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 1, characterized in that, Based on the feature vector, a heterogeneous causal effect estimation method is used to quantify the personalized response of each candidate site under the preset marketing strategy, and an initial causal effect score is obtained, specifically as follows: Based on whether each candidate site had received historical marketing strategy interventions, they were initially divided into a treatment group and a control group; The final feature vectors were used to perform balance correction between the treatment and control groups; Based on the heterogeneous causal effect estimation method, the treatment group and control group after balance correction are trained to construct a heterogeneous causal effect model and output the initial causal effect score for each candidate site. Several resampled subsets are generated based on the initial causal effect scores and the corresponding training dataset; The model is retrained based on each resampled subset to obtain the initial causal effect scores predicted under multiple resampled subsets, and the confidence interval is obtained. Output the initial causal effect score and its confidence interval for each candidate site.

4. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 3, characterized in that, The process of balancing the treatment and control groups using the final feature vectors specifically involves: The final feature vector of each candidate site is used as the input feature to obtain the processing indicator variable and observation effect variable for each site; Based on the feature vectors of the treatment group and the control group, a propensity scoring model is constructed, and the propensity score of each candidate site is output. For each treatment group, the site with the smallest tendency score error was selected from the control group for matching. After matching is completed, balanced subsets of the treatment group and the control group are generated and adjusted until all features meet the balance requirements. Output the balanced datasets for the treatment and control groups.

5. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 3, characterized in that, The heterogeneous causal effect estimation method is used to train the balanced correction treatment group and control group to construct a heterogeneous causal effect model, specifically as follows: The treatment group and control group datasets are used as training sets, which include the final feature vector, treatment indicator variable and observation effect variable for each candidate site; The final feature vector is used as the input feature, the indicator variable is processed as the intervention label, and the observed effect variable is used as the dependent variable. A potential control outcome prediction model was trained on the treatment group data to estimate the potential effects of the treatment group site without intervention. A potential treatment outcome prediction model was trained on control group data to estimate the potential effects of intervention on the control group site. The residual causal effect of the treatment group was calculated by combining the output of the prediction model based on the potential control results with the dependent variable of the treatment group site. The residual causal effect of the control group site was calculated by combining the output of the potential treatment outcome prediction model with the dependent variable of the control group site. Based on the residual causal effects of the treatment group and the control group, a heterogeneous causal effect model is constructed to output the initial causal effect score of each candidate site under the preset marketing strategy.

6. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 1, characterized in that, The process involves constructing a graph structure with candidate sites as nodes and related facilities within their neighborhoods as auxiliary nodes. This structure integrates the attribute and relationship information of the neighborhood facilities to generate a site-neighborhood fusion representation. The initial causal effect score is then corrected to obtain a corrected causal effect score. Specifically: Collect facility information within the vicinity of the candidate site; Based on the facility information between the candidate site and its neighboring facilities, a graph structure is constructed, where the candidate site is a node in the graph and the neighboring facilities are auxiliary nodes. By utilizing the message passing mechanism between nodes in a graph neural network, the graph structure interacts and fuses the features between candidate site nodes and neighboring facility nodes to obtain a site-neighborhood fusion vector. Based on the site-neighborhood fusion vector, the initial causal effect score of each candidate site is corrected to obtain the corrected causal effect score.

7. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 6, characterized in that, The initial causal effect score for each candidate site is corrected based on the site-neighborhood fusion vector to obtain the corrected causal effect score, specifically as follows: Align the initial causal effect score with the site-neighborhood fusion vector; Based on the alignment results, the initial causal effect score is weighted and fused with the site-neighborhood fusion vector to obtain the fusion causal effect score of the candidate site; A nonlinear transformation is performed on the fused causal effect score to obtain the corrected causal effect score.

8. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 1, characterized in that, The process of obtaining risk metrics and combining them with the corrected causal effect score to generate the final optimal score for each candidate site is as follows: Collect risk measurement indicators and obtain standardized risk measurement indicators through standardization processing; A multi-objective optimization problem is constructed based on the corrected causal effect score and the standardized risk metric, wherein the multi-objective optimization problem includes maximizing the corrected causal effect score and minimizing the standardized risk metric. Based on a multi-objective optimization problem, a Pareto frontier search model is constructed. The weights of the Pareto front search model are obtained through an adaptive weight adjustment mechanism, and the Pareto front search model is solved iteratively using an optimization algorithm. Output the Pareto optimal solution set for the candidate sites, and obtain the final optimal score for each candidate site.

9. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 8, characterized in that, The method of obtaining the weights of the Pareto frontier search model through an adaptive weight adjustment mechanism is as follows: Obtain the initial weight vector preset by the Pareto frontier search model; Based on the initial weight vector, calculate the comprehensive score for each candidate site in the current iteration; The contribution percentages of the corrected causal effect score to the overall score and the contribution percentages of the standardized risk measurement index to the overall score were obtained respectively. The deviation is obtained by comparing the contribution percentage of the corrected causal effect score to the comprehensive score and the contribution percentage of the standardized risk metric to the comprehensive score with the preset target weight vector. Based on the deviation, the initial weight vector is adjusted using the gradient adjustment method to output the weights of the Pareto front search model.

10. The intelligent screening method for marketing traffic venues based on big data analysis according to claim 1, characterized in that, The process of ranking all candidate sites based on the final selection score and selecting the best site is as follows: The final selection scores are sorted in descending order to generate a candidate site ranking list; According to the preset filtering conditions of the sorting list, the candidate venues at the top are selected as preferred venues. The filtering conditions include a preset number of top candidate venues and venues that reach a preset score threshold. Output the preferred set of sites.

Images