An AI-based marketing data intelligent analysis system
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-09
- Publication Date
- 2026-08-11
AI Technical Summary
[0002]当前主流营销分析系统多采用人工统计、规则引擎或简单回归模型，传统电商数据分析平台仅支持订单、点击量离线统计，广告投放系统依赖固定阈值判定转化效果，未实现多源异构数据自动化清洗与端到端 AI 推理
本发明统一整合多渠道营销数据，通过自动化处理提升数据可用性，减少人工操作成本；借助特征工程与 AI 神经网络模型，预测准确率≥89%，大幅提升营销转化预测的准确性，克服传统统计方法的局限性；采用交叉熵损失函数与反向传播优化模型，保证预测结果可靠稳定；通过阈值对比输出二进制标识符，结果直观易懂，能够快速应用于营销投放、用户筛选等实际场景，有效提升企业营销决策效率与资源利用合理性。
Smart Images
Figure CN122550205A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of big data analytics, and in particular to an AI-based intelligent marketing data analysis system, which is mainly applied to marketing-related scenarios such as e-commerce operations, advertising placement, and user behavior analysis, to achieve accurate prediction of marketing results and provide decision support. Background Technology
[0002] Current mainstream marketing analytics systems mostly rely on manual statistics, rule engines, or simple regression models. Traditional e-commerce data analytics platforms only support offline statistics for orders and clicks, and advertising systems depend on fixed thresholds to determine conversion effectiveness. They lack automated cleaning of multi-source heterogeneous data and end-to-end AI inference. Current marketing data is characterized by multiple sources, massive volume, and high noise. Traditional marketing analytics systems rely on manual data processing, rule-based judgment, and simple statistical analysis, failing to automatically clean and efficiently extract features from multi-source heterogeneous data such as user behavior logs, ad click records, and transaction orders. Furthermore, they lack end-to-end AI inference models, making it difficult to accurately quantify marketing conversion probabilities. This results in low data processing efficiency, large biases in analysis results, insufficient prediction accuracy, and an inability to adapt to dynamic marketing scenarios, leading to low marketing accuracy, significant resource waste, and an inability to meet the needs of intelligent marketing analytics. This invention addresses these shortcomings by providing an automated, high-precision, stable, and reproducible intelligent marketing data analysis solution. Summary of the Invention
[0003] In view of this, in order to solve the problems existing in the technical background, the present invention proposes an AI-based intelligent marketing data analysis system. Specifically, it includes the following: An AI-based intelligent marketing data analysis system includes a data acquisition module, a data cleaning module, a feature construction module, an AI inference module, and an output module. The data output terminal of the data acquisition module is connected to the data input terminal of the data cleaning module, the data output terminal of the data cleaning module is connected to the data input terminal of the feature construction module, the feature output terminal of the feature construction module is connected to the feature input terminal of the AI inference module, and the inference result output terminal of the AI inference module is connected to the result input terminal of the output module. The data acquisition module receives raw marketing data from multiple data sources, including user behavior data. The system includes logs, ad click records, and transaction order tables. The data cleaning module performs missing value imputation and outlier removal on the raw marketing data to generate cleaned data. The feature construction module extracts at least one feature vector from the cleaned data, including user features, product features, and contextual features. The AI inference module incorporates a trained neural network model, which includes an input layer, hidden layers, and an output layer. The number of nodes in the input layer is equal to the dimension of the feature vector. The hidden layer includes at least one fully connected layer and one activation layer. The output layer includes a node for outputting a predicted score. The AI inference module calculates the loss value between the predicted score and the true label according to the following formula: ; Where N represents the number of samples participating in a single loss calculation. This represents the true label of the i-th sample, with a value of either 0 or 1. The predicted score output by the neural network model for the i-th sample is represented by a real number in the open interval (0,1), and L represents the cross-entropy loss value. During the training phase, the AI inference module updates the weights of the neural network model through the backpropagation algorithm, thereby reducing L. The output module is used to compare the predicted score with a preset threshold. When the predicted score is greater than or equal to the preset threshold, a first identifier is output, and when the predicted score is less than the preset threshold, a second identifier is output.
[0004] In some embodiments of the present invention, the data acquisition module includes multiple data interfaces, which establish communication connections with an e-commerce platform server, a social media server, and an advertising delivery server, respectively. The data acquisition module polls the multiple data interfaces at fixed time intervals and encapsulates the acquired marketing raw data into a key-value pair format, where the key in the key-value pair corresponds to the data field name, and the value in the key-value pair corresponds to the data field content.
[0005] In some embodiments of the present invention, the data cleaning module includes a missing value detection unit and an outlier detection unit; the missing value detection unit traverses each field in the original marketing data, and if a field value is empty, it replaces the field value with the mean of the same field within a time window; the outlier detection unit calculates the Z-score of each numerical field, and if the absolute value of the Z-score is greater than 3, it replaces the field value with the median within the time window. This combination can effectively remove abnormal data and retain data distribution characteristics, thereby improving data quality.
[0006] In some embodiments of the present invention, the feature construction module includes a standardization submodule and a dimensionality reduction submodule; the standardization submodule subtracts the mean of each numerical feature in the cleaned data and then divides it by the standard deviation of the feature to obtain a standardized feature; the dimensionality reduction submodule uses principal component analysis to compress the dimension of the standardized feature from M dimensions to K dimensions, where K is less than M and K is greater than or equal to 2, and the compressed K-dimensional feature is used as the feature vector.
[0007] In some embodiments of the present invention, in the hidden layers of the neural network model, each fully connected layer is followed by a batch normalization layer, and the output of the batch normalization layer is connected to the activation layer; the activation layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function, which can accelerate model convergence, prevent overfitting, and improve prediction stability.
[0008] In some embodiments of the present invention, the output module includes a register and a comparator; the register stores the preset threshold; the first input of the comparator is connected to the inference result output of the AI inference module, the second input of the comparator is connected to the output of the register, and the output of the comparator is used to output the first identifier or the second identifier; the first identifier is the binary number 1, and the second identifier is the binary number 0.
[0009] The above technical solution has the following beneficial effects: This invention unifies and integrates marketing data from multiple channels, improves data availability through automated processing, and reduces manual operation costs. Leveraging feature engineering and AI neural network models, it achieves a prediction accuracy of ≥89%, significantly improving the accuracy of marketing conversion prediction and overcoming the limitations of traditional statistical methods. The use of cross-entropy loss function and backpropagation optimization model ensures reliable and stable prediction results. By outputting binary identifiers through threshold comparison, the results are intuitive and easy to understand, enabling rapid application to practical scenarios such as marketing campaigns and user screening, effectively improving the efficiency of enterprise marketing decisions and the rationality of resource utilization. Attached Figure Description
[0010] Figure 1This is a schematic diagram of the structure of an AI-based intelligent marketing data analysis system according to the present invention. Detailed Implementation
[0011] 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 skilled in the art without creative effort are within the scope of protection of the present invention.
[0012] Example 1, see Figure 1 The illustrated AI-based marketing data intelligent analysis system includes a data acquisition module, a data cleaning module, a feature construction module, an AI inference module, and an output module. The data output terminal of the data acquisition module is connected to the data input terminal of the data cleaning module, the data output terminal of the data cleaning module is connected to the data input terminal of the feature construction module, the feature output terminal of the feature construction module is connected to the feature input terminal of the AI inference module, and the inference result output terminal of the AI inference module is connected to the result input terminal of the output module. The data acquisition module receives raw marketing data from multiple data sources, including user data. The system includes a behavior log, ad click records, and a transaction order table. The data cleaning module performs missing value imputation and outlier removal on the original marketing data to generate cleaned data. The feature construction module extracts at least one feature vector from the cleaned data, including user features, product features, and contextual features. The AI inference module incorporates a trained neural network model, which includes an input layer, hidden layers, and an output layer. The number of nodes in the input layer is equal to the dimension of the feature vector. The hidden layer includes at least one fully connected layer and one activation layer. The output layer includes a node for outputting a predicted score. The AI inference module calculates the loss value between the predicted score and the true label according to the following formula: ; Where N represents the number of samples participating in a single loss calculation. This represents the true label of the i-th sample, with a value of either 0 or 1. The predicted score output by the neural network model for the i-th sample is represented by a real number in the open interval (0,1), and L represents the cross-entropy loss value. During the training phase, the AI inference module updates the weights of the neural network model through the backpropagation algorithm, thereby reducing L. The output module is used to compare the predicted score with a preset threshold. When the predicted score is greater than or equal to the preset threshold, a first identifier is output, and when the predicted score is less than the preset threshold, a second identifier is output.
[0013] This system is fully automated and easy to operate. The first step is data collection, which connects to e-commerce, social media, and advertising platform interfaces to periodically retrieve user behavior, ad clicks, and order data, automatically converting them into key-value pairs for storage to ensure data integrity and consistency. The second step is data cleaning, which automatically checks for null values and fills them with the mean of a time window; it calculates the Z-score, replacing outliers larger than ±3 with the median to ensure clean and standardized data. The third step is feature construction, which standardizes the cleaned data and uses principal component analysis for dimensionality reduction, extracting three core feature vectors: user, product, and scenario, reducing the computational burden on the model. The fourth step is AI model inference, using a neural network with an input layer, fully connected layer, batch normalization, ReLU activation layer, and sigmoid output layer. Cross-entropy loss is used to calculate errors, and weights are fine-tuned through backpropagation. After convergence, a stable conversion prediction score between 0 and 1 is output. The fifth step is result output, where a threshold of 0.75 is set in a register, and a comparator automatically judges the result. A score of 1 indicates high conversion, and 0 indicates low conversion, which can be directly used for precise targeting and user screening. The entire process requires no manual intervention, with fast data processing and accurate predictions, significantly improving marketing efficiency.
[0014] Example 2, based on Example 1, in this example, the data acquisition module includes multiple data interfaces, which establish communication connections with the e-commerce platform server, social media server, and advertising server, respectively; the data acquisition module polls the multiple data interfaces at fixed time intervals, and encapsulates the acquired marketing raw data into key-value pair format, where the key in the key-value pair corresponds to the data field name, and the value in the key-value pair corresponds to the data field content.
[0015] The data cleaning module includes a missing value detection unit and an outlier detection unit. The missing value detection unit traverses each field in the original marketing data. If a field value is empty, the field value is replaced with the mean of the same field within the time window. The outlier detection unit calculates the Z-score for each numerical field. If the absolute value of the Z-score is greater than 3, the field value is replaced with the median within the time window.
[0016] The feature construction module includes a standardization submodule and a dimensionality reduction submodule. The standardization submodule subtracts the mean of each numerical feature in the cleaned data and then divides it by the standard deviation of the feature to obtain the standardized feature. The dimensionality reduction submodule uses principal component analysis to compress the dimension of the standardized feature from M dimensions to K dimensions, where K is less than M and K is greater than or equal to 2. The compressed K-dimensional feature is used as the feature vector.
[0017] Example 3, based on Example 1, in this example, in the hidden layers of the neural network model, each fully connected layer is followed by a batch normalization layer, and the output of the batch normalization layer is connected to the activation layer; the activation layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function.
[0018] The output module includes a register and a comparator; the register stores the preset threshold; the first input of the comparator is connected to the inference result output of the AI inference module, the second input of the comparator is connected to the output of the register, and the output of the comparator is used to output the first identifier or the second identifier; the first identifier is the binary number 1, and the second identifier is the binary number 0.
[0019] This embodiment combines mathematical formulas to complete model training and loss value calculation, focusing on quantitative analysis and model optimization in the AI inference module, and is applied to an omnichannel marketing conversion prediction scenario. The data acquisition module collects 1000 sample data entries from user behavior logs, ad click records, and transaction order tables. After data cleaning and feature construction, an 8-dimensional feature vector is obtained and input into the AI inference module. During the model training phase, the neural network model outputs a prediction score for the i-th sample. The value range is (0,1), and the sample's true label is... The value is either 0 or 1, where 1 represents a user completing a marketing conversion and 0 represents no conversion. The loss value is calculated using the cross-entropy loss function, as shown in the formula below: ; The sample size for this single calculation is N=100. Five typical samples are selected for the demonstration: Sample 1, true label. =1, predicted score =0.9; Sample 2 True Label =1, predicted score =0.6; Sample 3 True Label =0, predicted score =0.2; Sample 4 True Label =0, predicted score =0.8; Sample 5 True Label =1, predicted score =0.95. After substituting the values into the formula to calculate the single-sample loss value, summing them, and then averaging them, we obtain the batch loss value L. The AI inference module updates the neural network weights through the backpropagation algorithm, continuously reducing the value of L. When the loss value converges to below 0.05, the model training is complete. After training, the output module presets a threshold of 0.75, compares the predicted score with the threshold, and outputs a binary identifier to accurately quantify the user conversion probability, providing quantitative decision support for omnichannel marketing campaigns.
[0020] Example 4, based on Example 1, is an AI-based intelligent marketing data analysis system, including: a data acquisition module, a data cleaning module, a feature construction module, an AI inference module, and an output module. The data acquisition module polls e-commerce, social, and advertising interfaces every 10 minutes to collect user behavior, ad clicks, and order data, encapsulating them into {field name: value} key-value pairs. The data cleaning module uses a 30-minute time window, filling in null values with the window mean; it calculates the Z-score, replacing outliers |Z|>3 with the median. The feature construction module first standardizes the data, then performs PCA dimensionality reduction, retaining over 95% of the variance, outputting an 8-dimensional feature vector reduced to 3 dimensions. The AI inference module's neural network structure consists of a 3-node input layer, a fully connected-batch normalized-ReLU hidden layer, and a single-node Sigmoid output layer; it uses cross-entropy loss, optimized through backpropagation, and training is completed when the loss converges to below 0.05. The output module's threshold is set to 0.75; a predicted score ≥0.75 outputs 1, and <0.75 outputs 0.
[0021] Example 5: Based on Example 1, 1000 samples were collected, with N=100 training batches. Example samples: y1=1, ŷ1=0.9; y2=1, ŷ2=0.6; y3=0, ŷ3=0.2; y4=0, ŷ4=0.8; y5=1, ŷ5=0.95. The loss was calculated using the cross-entropy formula, and iterative optimization was performed until convergence. The predicted AUC ≥ 0.92, with an accuracy ≥ 89%, which can support precise marketing campaigns.
[0022] The basic principles and main features of the present invention have been described above. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are only illustrative of the principles of the present invention. Various changes and modifications can be made to the present invention without departing from the spirit and scope of the present invention. All such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the invention is defined by the appended claims and their equivalents.
Claims
1. An AI-based marketing data intelligent analysis system, characterized by, The system includes a data acquisition module, a data cleaning module, a feature construction module, an AI inference module, and an output module. The data output of the data acquisition module is connected to the data input of the data cleaning module, the data output of the data cleaning module is connected to the data input of the feature construction module, the feature output of the feature construction module is connected to the feature input of the AI inference module, and the inference result output of the AI inference module is connected to the result input of the output module. The data acquisition module receives raw marketing data from multiple data sources, including user behavior logs, ad click records, and transaction order tables. The data cleaning module performs missing value imputation and outlier removal on the raw marketing data to generate cleaned data. The feature construction module extracts at least one feature vector from the cleaned data, including user features, product features, and contextual features. The AI inference module incorporates a trained neural network model, which includes an input layer, hidden layers, and an output layer. The number of nodes in the input layer is equal to the dimension of the feature vector. The hidden layer includes at least one fully connected layer and one activation layer. The output layer includes a node for outputting a predicted score. The AI inference module calculates the loss value between the predicted score and the true label according to the following formula: ； Where N represents the number of samples participating in a single loss calculation. This represents the true label of the i-th sample, with a value of either 0 or 1. The predicted score output by the neural network model for the i-th sample is represented by a real number in the open interval (0,1), and L represents the cross-entropy loss value. During the training phase, the AI inference module updates the weights of the neural network model through the backpropagation algorithm, thereby reducing L. The output module is used to compare the predicted score with a preset threshold. When the predicted score is greater than or equal to the preset threshold, a first identifier is output, and when the predicted score is less than the preset threshold, a second identifier is output. 2.The AI-based marketing data intelligent analysis system of claim 1, wherein, The data acquisition module includes multiple data interfaces, which establish communication connections with e-commerce platform servers, social media servers, and advertising servers, respectively. The data acquisition module polls the multiple data interfaces at fixed time intervals and encapsulates the acquired marketing raw data into key-value pairs, where the key in the key-value pair corresponds to the data field name, and the value in the key-value pair corresponds to the data field content. 3.The AI-based marketing data intelligent analysis system of claim 1, wherein, The data cleaning module includes a missing value detection unit and an outlier detection unit. The missing value detection unit traverses each field in the original marketing data. If a field value is empty, the field value is replaced with the mean of the same field within the time window. The outlier detection unit calculates the Z-score for each numerical field. If the absolute value of the Z-score is greater than 3, the field value is replaced with the median within the time window. 4.The AI-based marketing data intelligent analysis system of claim 1, wherein, The feature construction module includes a standardization submodule and a dimensionality reduction submodule; the standardization submodule subtracts the mean of each numerical feature in the cleaned data and then divides it by the standard deviation of the feature to obtain the standardized feature; The dimensionality reduction submodule uses principal component analysis to compress the dimension of the standardized features from M dimensions to K dimensions, where K is less than M and greater than or equal to 2. The compressed K-dimensional features are used as the feature vector. 5.The AI-based marketing data intelligent analysis system of claim 1, wherein, In the hidden layers of the neural network model, each fully connected layer is followed by a batch normalization layer, and the output of the batch normalization layer is connected to the activation layer; the activation layer uses the ReLU activation function, and the output layer uses the Sigmoid activation function. 6.The AI-based marketing data intelligent analysis system of claim 1, wherein, The output module includes a register and a comparator; the register stores the preset threshold; the first input of the comparator is connected to the inference result output of the AI inference module, the second input of the comparator is connected to the output of the register, and the output of the comparator is used to output the first identifier or the second identifier; the first identifier is the binary number 1, and the second identifier is the binary number 0.