Commodity selection method based on prepolymerization storage table
Through distributed data collection and storage, multi-dimensional product value evaluation and dynamic matching of customer group characteristics, combined with mixed integer planning model, the problems of incomplete data collection, inaccurate evaluation and unreasonable resource allocation in product selection are solved, precise product selection, personalized push and efficient resource utilization are achieved, and customer satisfaction and marketing effects are improved.
Patent Information
- Application Number
- CN202510440409.3
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-09
- Publication Date
- 2025-07-18
AI Technical Summary
The existing technology has a single data collection and storage method in product selection, which makes it difficult to obtain multi-source heterogeneous data in comprehensively and in real time. The product value evaluation and customer group matching lack accuracy, personalized recommendations cannot be achieved, and the resource allocation is unreasonable under constraints, resulting in poor product selection results.
Distributed data acquisition and storage are adopted, and data cleaning and multi-dimensional product value evaluation are carried out through pre-aggregated storage tables, combined with dynamic matching of customer characteristics and mixed integer planning models, a product selection method for maximum constraint optimization value is established, and Kafka message queue, collaborative filtering algorithm and reinforcement learning model are used for dynamic optimization.
It has achieved accurate product selection and personalized push, which improves customer purchase conversion rate and satisfaction, has high resource utilization efficiency, is safe and reliable data, is compliant and legal, and maximizes marketing effects.
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of data processing, and in particular to a product selection method based on a pre-aggregated storage table. Background Art
[0002] In today's digital business environment, data-driven decision-making is becoming increasingly crucial for enterprise operations, and product selection, as an important part of business activities, is no exception. With the rapid development of Internet technology, enterprises have accumulated a vast amount of customer behavior data, such as purchase records, interaction logs, etc. At the same time, there is a wide variety of products in the market, and the competition is fierce. How to accurately select products that meet the needs of the target customer group and have high value potential from a large number of candidate products has become the key for enterprises to enhance their competitiveness. At the same time, the continuous progress of data processing technologies, such as the development of distributed computing, big data storage, and machine learning algorithms, has provided more powerful tools and methods for product selection, enabling enterprises to have the opportunity to optimize product selection strategies in a more scientific way, improve operational efficiency and economic benefits.
[0003] At present, there are many limitations in the traditional product lottery selection method. It mainly designs different prizes manually based on experience, which is usually highly subjective and cannot maximize the value of the lottery activity. At the same time, when using data processing technology for the lottery selection method, on the one hand, the data collection and storage methods are relatively single, making it difficult to comprehensively and real-time obtain and process multi-source heterogeneous data, resulting in insufficient and time-sensitive information for product selection, and unable to keep up with market changes in a timely manner. On the other hand, there is a lack of accuracy in product value evaluation and customer group matching. Common evaluation methods often only consider a single attribute of the product or simple market data, ignoring the diversity and dynamic changes of customer needs; the customer group classification is not detailed enough, making it difficult to achieve personalized product recommendations and marketing. In addition, the existing technology lacks a systematic optimization method when dealing with constraints such as budget and inventory, which is prone to resource waste or unreasonable configuration. These defects make the product selection effect of enterprises poor, difficult to meet market demands, and affect the profitability and market competitiveness of enterprises. It is necessary to design a product selection method based on a pre-aggregated storage table to solve the above-mentioned problems. Summary of the Invention
[0004] The purpose of the present invention is to solve the deficiencies existing in the prior art and propose a product selection method based on a pre-aggregated storage table to solve the problems in the above technical solutions.
[0005] To achieve the above purpose, the present invention is realized through the following technical solutions: A product selection method based on a pre-aggregated storage table includes the following selection methods: S1. Distributed data collection and storage: Through a distributed log collection system, real-time obtain customer historical purchase records, lottery interaction logs, and candidate lottery product data, and after cleaning, store them in a pre-aggregated storage table; the pre-aggregated storage table adopts a columnar storage structure divided by time windows and asynchronously incrementally updates customer behavior data; S2. Multi-dimensional product value evaluation: Based on the customer historical purchase product value distribution, historical lottery product value curve, and customer prize acquisition records in the pre-aggregated storage table, calculate the comprehensive valuation of candidate products through a weighted probability model. The valuation includes basic value, customer expected value deviation, and inventory pressure coefficient; S3. Dynamic matching of customer group characteristics: According to the customer group labels in the pre-aggregated storage table, construct a customer group - product demand matrix, and use a collaborative filtering algorithm to calculate the matching degree between each customer group and lottery products, generating a differentiated push sequence; customer group classification is implemented based on the RFM model and purchase behavior clustering algorithm; S4. Constrained optimization for maximum value: Establish a mixed integer programming model, with product valuation and matching degree as revenue variables, and combine budget constraints, inventory constraints, and exposure equilibrium conditions to solve for the optimal product combination and push strategy. The model output includes a batch-by-batch delivery plan and a real-time replacement plan.
[0006] Further, in the S1 step, the distributed log collection system uses a Kafka message queue to achieve multi-source data streaming access; the pre-aggregated storage table includes the purchase frequency and lottery participation rate aggregated by the customer dimension, and the inventory turnover rate and historical winning distribution aggregated by the product dimension; the data cleaning process includes outlier removal and standardization processing, and the Z-score algorithm is used to detect outliers in purchase records.
[0007] 1. Further, in the S2 step, the weighted probability model satisfies V = (α * V_base) + (β * ΔV_expect) - (γ * C_inventory); where V_base is the moving average of the historical transaction price of the product; ΔV_expect is regression predicted through the premium rate of the customer's historical winning products; C_inventory = log(current inventory / safety inventory) * timeliness decay factor; the weights α, β, γ are dynamically adjusted through a random forest model, and the features include market heat trend and promotion activity intensity.
[0008] Further, in the S3 step, the customer group classification uses an improved K-means++ algorithm, and the feature vector includes purchase cycle, price sensitivity, and lottery participation depth; an attention mechanism is introduced when constructing the demand matrix to assign higher weights to recent interaction behaviors; the collaborative filtering algorithm combines item similarity and customer group similarity to fill sparse data through matrix factorization.
[0009] 2. Further, in the step S4, the objective function of the mixed-integer programming model is max Σ(V_i * x_i) + Σ(S_j * y_j); where V_i is the commodity valuation; S_j is the customer group matching degree bonus; The constraint conditions include diversity constraints: Σ(c_i * x_i) ≤ B_total (total budget constraint) Σ(x_i) ≥ N_min * category coverage; customer group coverage constraint: y_j ≥ θ * proportion of customer group size; The solution uses the branch and bound algorithm for parallel processing of large-scale variable scenarios.
[0010] Further, the click-through rate, conversion rate, and prize abandonment rate of the lottery commodities by customers are collected in real time through the data logging monitoring system. When the deviation between the actual data and the prediction exceeds the threshold, dynamic re-optimization is triggered, and the reinforcement learning model is used to adjust the commodity valuation weights, quickly replace inefficient commodities based on the Hungarian algorithm, and increase the exposure weights of high-potential customer groups through flexible budget reallocation.
[0011] Further, homomorphic encryption technology is used for data storage, and customer sensitive information is anonymized during the pre-aggregation stage; a compliance filtering layer is introduced for the selection of lottery commodities to automatically block categories prohibited by law and high-risk commodities; the output results of the value maximization model are subject to anti-fraud detection to identify abnormal benefit transfer patterns.
[0012] In summary, the present invention provides a commodity selection method based on a pre-aggregation storage table, which has the following beneficial effects: 1. Through distributed data collection and storage, multi-dimensional commodity value evaluation, and dynamic matching of customer group characteristics, the effects of accurate commodity selection and personalized push are achieved. In the distributed data collection and storage link, the Kafka message queue is used to obtain multi-source data in real time, which is cleaned and stored in the pre-aggregation storage table, providing an accurate basis for subsequent analysis. At the same time, the multi-dimensional commodity value evaluation comprehensively considers various factors to calculate the comprehensive valuation, ensuring that the selected commodities have high value potential. The dynamic matching of customer group characteristics classifies the customer groups based on scientific algorithms, constructs a demand matrix and calculates the matching degree. The three work together to accurately locate the commodities that meet the needs of different customer groups, realize personalized push, and improve the customer purchase conversion rate and satisfaction.
[0013] 2. Through multi-dimensional commodity value evaluation, constraint optimization value maximization and dynamic re-optimization mechanism, the effects of efficient resource utilization and value maximization are achieved. Multi-dimensional commodity value evaluation comprehensively considers commodity value and provides a scientific basis for product selection. Constrained optimization value maximization establishes a mixed integer programming model to solve the optimal commodity combination and push strategy under constraints such as budget and inventory, and reasonably allocate resources. When the deviation between actual data and prediction exceeds the threshold, the dynamic re-optimization mechanism is activated, and with the help of adjustment strategies such as reinforcement learning models, it ensures that resources can be fully utilized and waste can be avoided during the product selection and marketing process, so as to maximize commodity value and marketing effects and improve economic benefits.
[0014] 3. Through data cleaning, homomorphic encryption technology and compliance filtering layer, the data security and reliability and legal and compliant operation are achieved. In the data cleaning process, the Z-score algorithm is used to remove outliers and standardize the processing to ensure data quality. Dynamic encryption technology is used to ensure the security of customer sensitive information during data storage. Anonymization processing further protects customer privacy. The selection of lottery products introduces a compliance filtering layer to automatically block legally prohibited categories and high-risk products. In the process of data processing and product selection, it ensures that the data is secure and reliable, and the operation is fully in compliance with laws and regulations.
[0015] 4. Through customer group classification, collaborative filtering algorithm and attention mechanism, the effect of accurately grasping customer group needs and improving marketing accuracy is achieved. At the same time, customer group classification based on RFM model and purchase behavior clustering algorithm can accurately divide customer groups with different characteristics; collaborative filtering algorithm integrates the similarity of items and customer groups, fills sparse data through matrix decomposition, accurately calculates the matching degree, introduces attention mechanism when constructing the demand matrix, gives higher weight to recent interactive behaviors, and is more in line with the current needs of customers. The combination of the three can deeply understand the needs of each customer group, accurately push products, improve the pertinence and effectiveness of marketing activities, reduce marketing costs, and improve the input-output ratio of marketing resources. DETAILED DESCRIPTION
[0016] The following will be combined with the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.
[0017] Embodiment 1: The present invention provides a technical solution: a commodity selection method based on a pre-aggregated storage table, including the following selection method: S1. Distributed data collection and storage: Through a distributed log collection system, obtain historical customer purchase records, lottery interaction logs, and candidate lottery product data in real time. After cleaning, store them in a pre-aggregated storage table. The pre-aggregated storage table uses a columnar storage structure divided by time windows and asynchronously updates customer behavior data incrementally. Using a distributed log collection system can efficiently collect a large amount of scattered customer data, ensuring the comprehensiveness and real-time nature of the data. At the same time, the division by time windows and the columnar storage structure facilitate quick query and analysis of the data, and asynchronous incremental updates reduce the burden of data processing and improve the efficiency of data updates, enabling the information for product selection to closely follow market dynamics and laying a solid foundation for subsequent accurate product selection; S2. Multi-dimensional product value evaluation: Based on the value distribution of historical customer-purchased products, the value curve of historical lottery products, and customer prize acquisition records in the pre-aggregated storage table, calculate the comprehensive valuation of candidate products through a weighted probability model. The valuation includes the basic value, the deviation of customer expected value, and the inventory pressure coefficient. Through the multi-dimensional product value evaluation method, various factors are comprehensively considered, making the product valuation more comprehensive and accurate. The basic value reflects the value level of the product itself, the deviation of customer expected value can better meet customer needs, and the inventory pressure coefficient helps to reasonably control inventory costs. This comprehensive evaluation method can screen out products with higher cost performance and market potential, improving the quality and return of product selection; S3. Dynamic matching of customer group characteristics: According to the customer group labels in the pre-aggregated storage table, construct a customer group - product demand matrix, and use a collaborative filtering algorithm to calculate the matching degree between each customer group and lottery products, generating a differentiated push sequence. Customer group classification is achieved based on the RFM model and purchase behavior clustering algorithm. The dynamic matching of customer group characteristics, based on a scientific customer group model and algorithm, can deeply explore the demand characteristics of different customer groups. By constructing a demand matrix and calculating the matching degree through a collaborative filtering algorithm, differentiated push for different customer groups is realized. This not only improves customers' interest and participation in lottery products but also enhances the marketing effect, improving customer satisfaction and loyalty; S4. Maximizing value through constraint optimization: Establish a mixed-integer programming model with product valuation and matching degree as revenue variables, combined with budget constraints, inventory constraints, and exposure equilibrium conditions, to solve for the optimal product combination and push strategy. The model output includes a batch-by-batch delivery plan and a real-time replacement plan. Maximizing value through constraint optimization finds the optimal solution under various constraint conditions by establishing a mixed-integer programming model. The batch-by-batch delivery plan can reasonably arrange resources, avoiding the competitive pressure and resource waste caused by concentrated delivery; the real-time replacement plan enhances the flexibility to respond to market changes. This method ensures the maximization of product value and marketing effect with limited resources.
[0018] In step S1, the distributed log collection system uses a Kafka message queue to achieve multi-source data streaming access; the pre-aggregated storage table contains the purchase frequency aggregated by customer dimension, the lottery participation rate, the inventory turnover rate aggregated by product dimension, and the historical winning distribution; the data cleaning process includes outlier removal and standardization processing. The Z-score algorithm is used to detect outliers in purchase records, and multi-source data streaming access is achieved through the Kafka message queue, ensuring the efficiency and stability of data transmission. The data is multi-dimensionally aggregated through the pre-aggregated storage table, facilitating the quick acquisition of key information. By using the Z-score algorithm to remove outliers and perform standardization processing during the data cleaning process, the quality and reliability of the data are improved, making subsequent analysis and decision-making more accurate and reducing product selection mistakes caused by incorrect data.
[0019] In step S2, the weighted probability model satisfies V = (α * V_base) + (β * ΔV_expect) - (γ * C_inventory); where V_base is the moving average of the historical transaction price of the product; ΔV_expect is regression predicted through the premium rate of the customer's historical winning products; C_inventory = log (current inventory / safety inventory) * timeliness decay factor; the weights α, β, and γ are dynamically adjusted through a random forest model, and the features include the market heat trend and the intensity of promotional activities. The calculation methods of the parameters in the weighted probability model are scientific and reasonable, and the weights can be dynamically adjusted according to the market heat and the intensity of promotional activities. This can more flexibly adapt to market changes, accurately evaluate the value of products, adjust the weights according to the activity intensity during promotional activities, highlight the promotional value of products, select more suitable products for the activity, and improve the activity effect and sales performance.
[0020] In step S3, the customer group classification uses an improved K-means++ algorithm, and the feature vector includes the purchase cycle, price sensitivity, and lottery participation depth; an attention mechanism is introduced when constructing the demand matrix to assign higher weights to recent interaction behaviors; the collaborative filtering algorithm combines item similarity and customer group similarity, and fills in sparse data through matrix factorization. The improved K-means++ algorithm combined with multi-dimensional feature vectors can more accurately classify customer groups. Introducing the attention mechanism highlights recent interaction behaviors, making product selection more in line with the current needs of customers. The collaborative filtering algorithm combines multiple similarities and fills in sparse data, improving the accuracy of match degree calculation, making the pushed products more in line with the actual needs of customer groups, and increasing customer participation and purchase willingness.
[0021] In step S4, the objective function of the mixed integer programming model is maxΣ(V_i * x_i) + Σ(S_j * y_j); where V_i is the product valuation; S_j is the customer group matching degree bonus; The constraints include diversity constraints: Σ(c_i * x_i) ≤ B_total (total budget constraint), Σ(x_i) ≥ N_min * category coverage; customer group coverage constraint: y_j ≥ θ * proportion of customer group size; The branch and bound algorithm is used for solving, parallel processing is applied to large-scale variable scenarios, and the objective function of the mixed integer programming model comprehensively considers commodity valuation and customer group matching degree, achieving optimal allocation of resources under various constraints. The branch and bound algorithm combined with parallel processing improves the solving efficiency and can quickly obtain the optimal commodity combination and push strategy, which helps to reasonably allocate the budget, cover more customer groups, and improve the utilization efficiency of marketing resources in large-scale business scenarios.
[0022] The click-through rate, conversion rate, and award abandonment rate of customers for lottery products are collected in real time through the data logging monitoring system. When the deviation between the actual data and the prediction exceeds the threshold, dynamic re-optimization is triggered. The reinforcement learning model is used to adjust the commodity valuation weights, the Hungarian algorithm is used to quickly replace inefficient products, and the exposure weights of high-potential customer groups are increased through flexible budget reallocation. The data logging monitoring system enables real-time monitoring of customer behavior. The dynamic re-optimization mechanism can respond to market changes in a timely manner. By adjusting weights with the reinforcement learning model, replacing products with the Hungarian algorithm, and reallocating flexible budgets, the product selection and push strategies can be continuously optimized according to the actual effects, improving the effectiveness and return on investment of marketing activities, and maintaining the adaptability and competitiveness of the product selection method.
[0023] Homomorphic encryption technology is used for data storage, and customer sensitive information is anonymized during the pre-aggregation stage; A compliance filtering layer is introduced for lottery product selection to automatically block categories prohibited by law and high-risk products; The output results of the value maximization model are subject to anti-fraud detection to identify abnormal benefit transfer patterns. Homomorphic encryption technology and anonymization processing ensure the security and privacy of customer data, enhance customer trust. The compliance filtering layer ensures that product selection is legal and compliant, avoiding potential legal risks. At the same time, anti-fraud detection identifies abnormal benefit transfer, maintains a fair and just market environment, and safeguards the legitimate rights and interests of enterprises and customers, making the product selection method more reliable and sustainable.
[0024] Embodiment 2: This embodiment is based on a product selection method based on a pre-aggregation storage table, and elaborates in detail on the weighted probability model: 1. The sliding average of the historical transaction price of V_base products: In specific calculations, a time window is set, the past 180 days. All transaction prices of products within this time window are collected, denoted as P1, P2, ⋯, Pn; The sliding average Vbase = ∑(n - i = 1)Pi / n. As new transaction data is generated, the average value is updated in a sliding window manner to ensure that it can reflect the current price level of the product.
[0025] 2. ΔV_expect is predicted through the regression of the premium rate of the customer's historical winning products: First, calculate the premium rate of the customer's historical winning products. For each winning product, obtain its market reference price M and the actual lottery winning price (or cost price) C. The premium rate r = (M - C) / n. Then, collect the premium rate data of multiple historical winning products and related features such as product category, promotion activity type, etc. as training data. Finally, use a linear regression model for training to establish the relationship between the premium rate and these features. When predicting the ΔV_expect of a new candidate product, input the features of this product into the trained regression model to obtain the predicted premium rate deviation value.
[0026] 3. C_inventory = log (current inventory / safety inventory) * timeliness decay factor: The current inventory refers to the actual inventory quantity of the product when making product selection. The safety inventory is a pre-set inventory threshold based on historical sales data and market demand fluctuations to ensure that there will be no out-of-stock situation within a certain period of time. The timeliness decay factor is determined according to factors such as the shelf life of the product and market trends. For products with fast-changing trends, the timeliness decay factor is dynamically adjusted according to its popularity change curve in the market. Assuming the timeliness decay factor is λ(t), where t represents time, then C_inventory = log (current inventory / safety inventory) * λ(t).
[0027] 4. The weights α, β, γ are dynamically adjusted through a random forest model. The features include the market popularity trend and the intensity of promotion activities: First, collect market popularity trend data. For example, measure the market popularity of the product category through indicators such as network search index and social media discussion popularity. The intensity of promotion activities can be quantified according to factors such as the discount rate of the promotion activity and the number of products participating in the activity. Then, use these features, the comprehensive valuation V of historical products, and actual sales data as training data and input them into the random forest model for training. Finally, after training is completed, when determining the weights of a new candidate product, input the features such as the market popularity trend and the intensity of promotion activities corresponding to this product into the trained random forest model, and the model outputs the corresponding weights α, β, γ.
[0028] On the one hand, through a weighted probability model for in-depth analysis, the comprehensibility and practicality of the technical solution are improved. The calculation methods of V_base, ΔV_expect, and C_inventory are detailed, providing clear operation guidelines for technical personnel. This enables the product selection method not to remain at the theoretical level but to be effectively applied to actual business scenarios, helping enterprises more accurately evaluate the value of products, optimize product selection decisions, and improve operational efficiency and economic benefits. On the other hand, for the dynamic adjustment of the weights α, β, and γ, through specific random forest model training examples, the model can better adapt to market changes, which helps to more precisely evaluate the value of products, enhance the scientificity and accuracy of product selection, thereby optimizing resource allocation and improving marketing effects and commercial benefits.
[0029] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention in any other form. Any technical personnel familiar with the relevant field may use the disclosed technical content to make changes or modifications into equivalent embodiments with equivalent changes and apply them to other fields. However, as long as they do not depart from the technical solution content of the present invention, any simple modifications, equivalent changes, and modifications made to the above embodiments based on the technical essence of the present invention still fall within the protection scope of the technical solution of the present invention.
Claims
1. A product selection method based on a pre-aggregated storage table, characterized in that: It includes the following product selection methods: S1. Distributed data collection and storage: Through a distributed log collection system, real-time obtain customer historical purchase records, lottery interaction logs, and candidate lottery product data. After cleaning, store them in a pre-aggregated storage table; The pre-aggregated storage table adopts a columnar storage structure divided by time windows and asynchronously incrementally updates customer behavior data; S2. Multi-dimensional product value evaluation: Based on the customer historical purchase product value distribution, historical lottery product value curve, and customer prize acquisition records in the pre-aggregated storage table, calculate the comprehensive valuation of candidate products through a weighted probability model. The valuation includes basic value, customer expected value deviation, and inventory pressure coefficient; S3. Dynamic matching of customer group characteristics: According to the customer group labels in the pre-aggregated storage table, construct a customer group - product demand matrix, and use a collaborative filtering algorithm to calculate the matching degree between each customer group and lottery products, generating a differentiated push sequence; Customer group classification is implemented based on the RFM model and purchase behavior clustering algorithm; S4. Constraint optimization to maximize value: Establish a mixed-integer programming model, with product valuation and matching degree as revenue variables, combined with budget constraints, inventory constraints, and exposure equilibrium conditions, to solve the optimal product combination and push strategy. The model output includes a batch-by-batch delivery plan and a real-time replacement plan.
2. The method for selecting products based on a pre-aggregated storage table according to claim 1, wherein: In the S1 step, the distributed log collection system uses a Kafka message queue to achieve multi-source data streaming access; the pre-aggregated storage table includes the purchase frequency and lottery participation rate aggregated by the customer dimension, and the inventory turnover rate and historical winning distribution aggregated by the product dimension; the data cleaning process includes outlier removal and standardization processing, and the Z-score algorithm is used to detect outliers in purchase records.
3. The merchandise selection method based on a pre-aggregated storage table according to claim 1, wherein: In the S2 step, the weighted probability model satisfies V = (α * V_base) + (β * ΔV_expect) - (γ * C_inventory); Where V_base is the moving average of the historical transaction price of the product; ΔV_expect is predicted by regression of the premium rate of the customer's historical winning products; C_inventory = log (current inventory / safety inventory) * timeliness decay factor; the weights α, β, γ are dynamically adjusted through a random forest model, and the features include market heat trend and promotion activity intensity.
4. A method for selecting products based on a pre-aggregated storage table according to claim 1, characterized in that: In the S3 step, the customer group classification uses an improved K-means++ algorithm, and the feature vector includes purchase cycle, price sensitivity, and lottery participation depth; an attention mechanism is introduced when constructing the demand matrix to assign higher weights to recent interaction behaviors; the collaborative filtering algorithm combines item similarity and customer group similarity to fill sparse data through matrix factorization.
5. A method for selecting products based on a pre-aggregated storage table according to claim 1, wherein: In the S4 step, the objective function of the mixed-integer programming model is max Σ (V_i * x_i) + Σ (S_j * y_j); Where V_i is the product valuation; S_j is the customer group matching degree bonus; The constraints include diversity constraints: Σ(c_i*x_i) ≤ B_total (total budget constraint), Σ(x_i) ≥ N_min * category coverage; customer group coverage constraint: y_j ≥ θ * proportion of customer group size; The branch and bound algorithm is used for solving, and parallel processing is adopted for large-scale variable scenarios.
6. The commodity selection method based on a pre-aggregated storage table according to claim 1, characterized in that: Through the buried point monitoring system, the click-through rate, conversion rate, and prize abandonment rate of customers for the lottery products are collected in real time. When the deviation between the actual data and the prediction exceeds the threshold, dynamic re-optimization is triggered. The reinforcement learning model is used to adjust the product valuation weights, quickly replace inefficient products based on the Hungarian algorithm, and increase the exposure weights of high-potential customer groups through flexible budget reallocation.
7. A method for selecting products based on a pre-aggregated storage table according to claim 1, characterized in that: Homomorphic encryption technology is used for data storage, and customer sensitive information is anonymized during the pre-aggregation stage; A compliance filtering layer is introduced for the selection of lottery products to automatically block categories prohibited by law and high-risk products; The output results of the value maximization model are subject to anti-fraud detection to identify abnormal benefit transfer patterns.