Sales data analysis method based on big data and cloud computing

Abstract

The invention provides a sales data analysis method based on big data and cloud computing, and the method comprises the steps: cleaning product sales data from multiple channels, and carrying out the partition storage processing according to dimensions; constructing a hypergraph model according to the processed data; designing a hypergraph convolutional network to learn the hypergraph model and optimize the synergistic effect between product nodes; constructing a dynamic strategy generation model according to the synergistic effect score, and outputting an optimal strategy; and applying the optimal strategy to a user front end to collect feedback data, optimizing the prediction capability of the hypergraph model, and designing a time decay regularization mechanism to control the dependence of the hypergraph model on historical data, thereby realizing complete feedback closed-loop optimization. The method not only optimizes the formulation process of the sales strategy, but also can improve the coping ability of an enterprise in an environment where the market changes rapidly, and finally improves the overall market competitiveness.

Description

Technical Field

[0001] The present invention belongs to the technical field of big data and cloud computing, and in particular to a sales data analysis method based on big data and cloud computing. Background Art

[0002] In modern enterprise management, sales data analysis is an important part of decision-making. With the globalization and digitalization of corporate business, the scale and complexity of sales data are growing rapidly. Enterprises not only need to obtain sales data from traditional sales channels (such as physical stores, e-commerce platforms, etc.), but also need to integrate external data from multiple dimensions such as social media, user reviews, and climate conditions. These data come from a wide range of sources and have diverse structures, including multiple dimensions such as time, region, customer, and product type. Therefore, how to efficiently process and analyze these heterogeneous data has become a core challenge facing enterprises.

[0003] Existing sales data analysis methods mainly rely on traditional time series analysis, statistical regression models, and machine learning models. These methods perform well when processing single-dimensional or small-scale sales data, but when faced with multi-dimensional and large-scale sales data, they gradually expose the following major problems:

[0004] Inefficiency of data management and storage: Traditional sales data processing systems mostly rely on relational databases or centralized storage systems. With the rapid growth of data volume, this centralized storage architecture often cannot provide sufficient scalability and flexibility, resulting in slow data processing, especially when processing real-time data. In addition, data of different dimensions (such as product, region, time) often need to be stored and processed separately, and there is a lack of efficient unified management methods, resulting in low efficiency in query and analysis.

[0005] Neglect of inter-product synergies: Sales analysis usually models each product independently, ignoring the potential synergies between different products. For example, the promotion of a popular product may drive sales of related products, but existing analysis methods (such as regression-based prediction models) are difficult to effectively capture such complex cross-product synergies. Especially in the sales environment of multi-category products, the associations between products may be nonlinear, and traditional analysis methods cannot deeply explore these complex sales patterns.

[0006] Insufficient real-time and elastic computing: Traditional sales analysis tools are mostly based on batch processing systems, and the results of data analysis often have large delays, making it difficult to provide real-time feedback to corporate decision makers. In the actual sales environment, the real-time nature of data is crucial, especially in the case of promotional activities, seasonal fluctuations, or sudden market changes. Companies need to obtain sales analysis results in a timely manner and quickly adjust sales strategies. In addition, traditional systems often experience performance bottlenecks during peak periods (such as e-commerce promotions), and are unable to flexibly expand computing resources, resulting in system crashes or extremely slow response speeds.

[0007] In response to the above problems, although distributed architecture and big data processing technologies based on cloud computing have been applied to a certain extent in recent years, these systems are mostly focused on improving data storage and computing efficiency, and have failed to effectively solve the problems of synergy analysis between products and lack of real-time performance. Summary of the invention

[0008] The purpose of the present invention is to design a sales data analysis method based on big data and cloud computing, which overcomes the shortcomings of the prior art in terms of low data storage efficiency, inability to analyze product synergy effects, and lack of real-time performance, and provides enterprises with a new, efficient, and intelligent sales data analysis method. This not only optimizes the sales strategy formulation process, but also improves the company's ability to respond in a rapidly changing market environment, and ultimately improves the overall market competitiveness.

[0009] In order to achieve the above object, the present invention provides a sales data analysis method based on big data and cloud computing, the method comprising:

[0010] S1. Product sales data from multiple channels are cleaned and processed by dimension partition storage, wherein the dimension partition storage processing is partition storage and constructing a synergy matrix S in the partition storage processing to represent the synergy relationship between products, and outputting the processed data;

[0011] S2. Construct a hypergraph model based on the processed data. Each node of the hypergraph model represents a product, and the hyperedge represents the collaborative relationship between multiple products. Each hyperedge has a corresponding weight. Then, a regularization term is designed to optimize the hypergraph. Finally, the optimized hypergraph model H is input, H = (V, ε), where V is a node set, representing the set of all products; ε is a hyperedge set, representing the collaborative relationship between products, and each hyperedge e k With weight w(e k );

[0012] S3. Design a hypergraph convolutional network to learn the hypergraph model and optimize the synergy between product nodes. The final feature representation of the prediction obtained after training the hypergraph convolutional network is mapped to a synergy score.

[0013] S4. constructing a dynamic strategy generation model according to the synergy effect score, wherein the dynamic strategy generation model generates real-time product recommendations and promotion strategies according to the real-time product sales data flowing in, and outputs the optimal strategy;

[0014] S5. Apply the optimal strategy to the user front end to collect feedback data, and design feedback factors based on the feedback data. The feedback factors are the differences between the actual click-through rate and purchase conversion rate and the synergy score predicted by the hypergraph model. Finally, the feedback factors are integrated into the new process of the hypergraph convolutional network to optimize the prediction ability of the hypergraph model. At the same time, a time decay regularization mechanism is designed to control the hypergraph model’s dependence on historical data, so as to achieve a complete feedback closed-loop optimization.

[0015] Furthermore, the dimension partition storage process includes the following steps:

[0016] The cleaned data is partitioned and stored according to the time, product and region dimensions, defined as D partition , the specific operations are as follows:

[0017] The time dimension is divided into days, weeks, and months;

[0018] The product dimension is segmented by brand and category;

[0019] The regional dimension is divided by city and region;

[0020] Construct the synergy matrix S to represent the synergy relationship between products, where the synergy s ij Calculated by the following formula:

[0021]

[0022] Among them, s ij represents the synergy effect between products i and j; represents the sales data of product i at time t and region g; represents the sales data of product j at time t and region g; ∈ represents a small positive number used to prevent division by zero; the synergy matrix S is used for compression and index optimization, and t represents time.

[0023] Furthermore, the structure of the hypergraph model H is expressed as follows:

[0024] Each product i It is a node in the hypergraph. The attributes of the node are the sales data of the corresponding product in different periods and regions. To define;

[0025] A super edge kConnecting multiple products indicates that these products have significant synergy in the same space and time. The construction of hyperedges is based on the synergy matrix S, whose elements s ij Representative product i and p j If the synergy between multiple products reaches the set threshold, they are connected on the same hyperedge;

[0026] At the same time, design the synergistic effect threshold θ s , only when s ij >θ s Only when , products i and j are connected to the same hyperedge.

[0027] Furthermore, the design regularization term optimizes the hypergraph, specifically:

[0028] In order to control the size of the hyperedge and avoid excessive connection of irrelevant products, a regularization term R(e k ), adjust the complexity of the hyperedge, expressed as follows

[0029]

[0030] Among them, R(e k ) represents the hyperedge e k The regularization term is used to balance the synergy strength and hyperedge complexity; α represents the weight for controlling the synergy strength; β represents the penalty coefficient for adjusting the number of nodes in the hyperedge to limit the excessive complexity of the hyperedge; k represents the hyperedge e k The number of nodes included; s ij represents the synergy between nodes i and j;

[0031] At the same time, in order to deal with the possible noise and abnormal synergy effects in large-scale sales data, a special regularization term L is designed special , which is used to further constrain the synergy matrix S, is expressed as follows:

[0032]

[0033] Among them, L special represents a special regularization term, which is used to smooth out excessive synergy effect values ​​and prevent abnormal growth of synergy effects between certain products; γ represents the intensity of controlling the regularization term; exp(-s ij ) is used to suppress extreme values, making abnormally large s ij is smoothed.

[0034] Furthermore, the update rule of the hypergraph convolutional network is as follows:

[0035] For each node p i,The goal of the hypergraph convolution operation is to aggregate features from its neighboring nodes and propagate features through the convolution formula, which is expressed as follows:

[0036]

[0037] in, Represents node p i Feature representation at convolutional layer l+1; Represents the hyperedge e k Other nodes p j The features in the convolutional layer l; w(e k ) represents the hyperedge e k The weight is used to measure the synergy between nodes; Represents the hyperedge e k The number of connected nodes, used as a normalization factor; W (l) represents the learnable weight matrix of the convolutional layer l, which is used for linear transformation of features; σ represents the activation function, which is used to introduce nonlinear operations;

[0038] Then perform multi-layer convolution:

[0039] Initial Input For node p i The eigenvector x i , after multiple layers of hypergraph convolution, it is updated to That is, the final feature representation in the convolutional layer L, where each layer of convolution updates the feature representation of the node.

[0040] Furthermore, the loss function of the hypergraph model includes node classification loss and synergy effect regularization loss;

[0041] The node classification loss is that some nodes have known synergy effect labels in historical data, and the cross entropy loss is used to train the hypergraph model to classify the known synergy effect labels, which is expressed as follows:

[0042]

[0043] Among them, y i is node p i The real label, is the label predicted by the model;

[0044] At the same time, design the regularization L reg The hyperedge weight is limited to be too large, and the model training is guided by combining the known synergy matrix S, which is expressed as follows:

[0045]

[0046] Among them, w(e k ) is the synergy effect weight;

[0047] A special regularization term L is also introduced for some extreme cases special,extreme To suppress the occurrence of abnormal situations, the regular term is used to punish node p i Connected by too many hyperedges, represented as follows:

[0048]

[0049] Among them, |ε(p i )| indicates connection to node p i The number of hyperedges in the hypergraph, N is the total number of hyperedges in the hypergraph;

[0050] The total loss function is expressed as follows:

[0051] L total =L cls +λ 1 L reg +λ 2 L special,extreme

[0052] Among them, λ i and λ 2 It is a hyperparameter that weighs various losses and is used to control the strength of regularization.

[0053] Furthermore, the real-time streaming product sales data includes order data, user behavior data and inventory data.

[0054] Furthermore, the steps of constructing the dynamic strategy generation model include:

[0055] Based on real-time order data order Calculate each node p i , that is, the real-time sales growth rate r of each product i (t), represents the sales growth rate of the product in time t, expressed as follows:

[0056]

[0057] Among them, S i (t) represents product p i Total sales at time t, S i (tT) represents the sales volume of the previous time window, ∈ is a smoothing term to avoid division by zero error; r i (t) used to reflect real-time sales trends of products;

[0058] Score based on product synergy Real-time sales growth rate i (t) and inventory status I i Design objective function G i(t), expressed as follows:

[0059]

[0060] Among them, G i (t) represents the product p at time t i The strategy score is used to evaluate the product recommendation priority; κ, μ, ν represent the weight coefficients of controlling the influence of each factor; Indicates product p i The synergy score indicates the potential of the product to be sold together with other products; i Indicates product p i Inventory status. When the inventory is low, the recommendation weight will be weakened;

[0061] In the process of strategy optimization, a special regularization term L is designed special-hot , balances the diversity of strategy generation and avoids over-recommending certain popular products, which results in the neglect of sales of other products. This regularization term penalizes products with high concentration in strategy scores. The formula is as follows:

[0062]

[0063] Among them, δ represents the regularization term weight.

[0064] Furthermore, the feedback factor is designed as follows:

[0065]

[0066] Among them, Δy i represents the feedback factor, represents the product p i Synergy score of actual performance and model prediction The difference between them; ρ represents the weight of click feedback, which is used to control the weight of click feedback in adjusting the synergy score; ω represents the weight of purchase feedback, which is used to control the weight of purchase feedback in adjusting the synergy score; F click (p i ) indicates product p i The click feedback indicates the frequency of users clicking on the recommended products; purchase (p i ) indicates product p i The purchase feedback indicates the actual purchase conversion rate of users for recommended products.

[0067] Furthermore, the feedback factor is introduced into the loss function of the hypergraph convolutional network to serve as a guide for updating model parameters in future training processes, as shown below:

[0068]

[0069] Among them, L feedback represents the feedback adjustment loss, which is used to minimize the predicted synergy score of the product and the corrected score y after actual feedback i The difference between

[0070] The time decay regularization mechanism is expressed as follows:

[0071]

[0072] Among them, L time represents the time decay regularization term, which limits the drastic changes of hyperedge weights in different time periods; w(e k ,t) represents the hyperedge e k The weight at the current time t; w(e k ,tT) represents the hyperedge e k The weight in the previous time window tT; λ represents the time decay coefficient, which is used to control the degree of dependence of the hypergraph model on historical synergy effects.

[0073] The beneficial technical effects of the present invention are at least as follows:

[0074] (1) The present invention proposes a sales data partition storage and management method based on cloud native technology, which can perform hierarchical partitioning according to the multi-dimensional characteristics of sales data (such as time, product, region, etc.). The data of different partitions are managed through a distributed storage system (such as HDFS, S3), thereby achieving efficient access to data. At the same time, with the help of containerization technology (such as Kubernetes) and real-time stream processing frameworks (such as Apache Flink), the present invention can dynamically schedule computing resources according to changes in data traffic to achieve elastic expansion and real-time processing. This innovation solves the problems of low storage efficiency and inability to dynamically expand computing resources in traditional centralized storage systems when facing large-scale sales data, ensuring the efficiency and stability of the system under high-load environments.

[0075] (2) The present invention proposes a cross-product collaborative sales analysis method based on hypergraph learning. Different from the traditional independent product analysis method, this system models the sales relationship between multiple products as hyperedges through a hypergraph model, capturing the impact of multiple products on a certain promotional activity or market event at the same time. The hypergraph convolutional network (HGCN) can learn the complex nonlinear collaborative relationships between products and provide more accurate sales forecasts and promotion strategy optimization suggestions based on these relationships. Compared with traditional machine learning models, hypergraph learning can better handle the complex interactions between multiple products, especially in multi-category or cross-regional sales scenarios, and realizes in-depth mining of cross-product synergy effects.

[0076] (3) The present invention realizes real-time processing of sales data through a stream processing framework. The system can analyze data instantly when it flows in and provide instant sales strategy suggestions based on real-time sales data. Especially during e-commerce promotions or sudden market changes, the system can automatically adjust promotion plans or recommend cross-product collaborative strategies based on real-time data analysis results, thereby greatly improving the flexibility and timeliness of sales analysis. The data delay problem of traditional batch processing systems has been effectively solved in the present invention, ensuring the timeliness of analysis results and efficient feedback. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] The present invention is further described using the accompanying drawings, but the embodiments in the accompanying drawings do not constitute any limitation to the present invention. A person skilled in the art can obtain other drawings based on the following drawings without creative work.

[0078] Figure 1 This is a flow chart of the sales data analysis method based on big data and cloud computing of the present invention. DETAILED DESCRIPTION

[0079] Embodiments of the present invention are described in detail below, examples of which are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present invention, and cannot be understood as limiting the present invention.

[0080] In one or more embodiments, Figure 1 As shown, a sales data analysis method based on big data and cloud computing is disclosed, comprising the following steps S1-S5:

[0081] S1. Product sales data from multiple channels are cleaned and stored by dimension partitions, wherein the dimension partition storage processing is partition storage and constructing a synergy matrix S in the partition storage processing to represent the synergy relationship between products, and the processed data is output.

[0082] Specifically, collect sales data from multiple channels raw , including sources such as e-commerce platforms, physical retail, and social media. The data dimensions include product information, sales time, region, customer type, price, etc.

[0083] For outliers in sales data, an adaptive outlier correction formula is used. i (such as sales of a single product or customer spending amount), the modified formula for calculating standardization is as follows:

[0084]

[0085] Among them, x i represents the original data, such as the sales of a single product or the amount of money spent by a customer; λ represents the adaptive adjustment parameter, which is used to dynamically adjust the correction strength; Represents x i The mean of represents the historical mean of this dimension; Represents x i The standard deviation of , indicating the fluctuation range of the data; ∈ indicates a small positive number to prevent division by zero, used to avoid the problem of the denominator being zero; x i ' indicates the corrected data for subsequent processing.

[0086] Furthermore, the missing values ​​are filled using the historical sales of similar products. The filling formula is:

[0087]

[0088] Among them, x missing Indicates the missing value to be filled; α j Represents the similarity weight, calculated based on product history and sales data; x j,similar represents the historical sales volume of similar products; n represents the number of similar products that can be referenced. Output the cleaned data D clean , as the subsequent partition storage input.

[0089] Furthermore, the dimension partition storage process includes the following steps:

[0090] The cleaned data is partitioned and stored according to the time, product and region dimensions, defined as D partition , the specific operations are as follows:

[0091] The time dimension is divided into days, weeks, and months;

[0092] The product dimension is segmented by brand and category;

[0093] The regional dimension is divided by city and region;

[0094] Construct the synergy matrix S to represent the synergy relationship between products, where the synergy s ij Calculated by the following formula:

[0095]

[0096] Among them, s ij represents the synergy effect between products i and j; represents the sales data of product i at time t and region g; represents the sales data of product j at time t and region g; ∈ represents a small positive number used to prevent division by zero; the synergy matrix S is used for compression and index optimization, and t represents time.

[0097] Finally, the output data is D compressed , represents the data after partitioning, compression and indexing. This data is used as the input for subsequent hypergraph construction

[0098] S2. Construct a hypergraph model based on the processed data. Each node of the hypergraph model represents a product, and the hyperedge represents the collaborative relationship between multiple products. Each hyperedge has a corresponding weight. Then, a regularization term is designed to optimize the hypergraph. Finally, the optimized hypergraph model H is input, H = (V, ε), where V is a node set, representing the set of all products; ε is a hyperedge set, representing the collaborative relationship between products, and each hyperedge e k With weight w(e k ).

[0099] Specifically, the input data D from step 1 compressed It is the sales data after partitioning, compression and index optimization. The data contains multi-dimensional (time, product, region) information and uses the synergy matrix S, whose elements s ij It represents the synergy effect between product i and product j in the same time period and the same region. The present invention needs to construct a hypergraph based on these data for subsequent synergy effect learning.

[0100] Furthermore, the structure of the hypergraph model H is expressed as follows:

[0101] Each product i It is a node in the hypergraph. The attributes of the node are the sales data of the corresponding product in different periods and regions. To define;

[0102] A super edge k Connecting multiple products indicates that these products have significant synergy in the same space and time. The construction of hyperedges is based on the synergy matrix S, whose elements s ij Representative product i and p j synergy effect; if the synergy effect between multiple products reaches the set threshold, they are connected on the same hyperedge; for example, if s ij 、s ik and jk are both larger, then p i 、p j and p k Connected on the hyperedge 3 The weight of the hyperedge is the weighted average of these synergistic effects:

[0103]

[0104] Among them, w(ek ) represents the hyperedge e k The weight of the hyperedge, k is the number of products connected by the hyperedge, and s ij For product p i and p j The greater the weight, the stronger the synergy between products.

[0105] To avoid meaningless connections, the present invention defines a synergy threshold θ s , only when s ij >θ s Only when , products i and j are connected to the same hyperedge. This threshold can be optimized by cross-validation method based on historical data to ensure the rationality and efficiency of hyperedge connection.

[0106] Furthermore, the design regularization term optimizes the hypergraph, specifically:

[0107] In order to control the size of the hyperedge and avoid excessive connection of irrelevant products, a regularization term R(e k ), adjust the complexity of the hyperedge, expressed as follows

[0108]

[0109] Among them, R(e k ) represents the hyperedge e k The regularization term is used to balance the synergy strength and hyperedge complexity; α represents the weight for controlling the synergy strength; β represents the penalty coefficient for adjusting the number of nodes in the hyperedge to limit the excessive complexity of the hyperedge; k represents the hyperedge e k The number of nodes included; s ij represents the synergy between nodes i and j;

[0110] At the same time, in order to deal with the possible noise and abnormal synergy effects in large-scale sales data, a special regularization term L is designed special , which is used to further constrain the synergy matrix S, is expressed as follows:

[0111]

[0112] Among them, L special represents a special regularization term, which is used to smooth out excessive synergy effect values ​​and prevent abnormal growth of synergy effects between certain products; γ represents the intensity of controlling the regularization term; exp(-s ij ) is used to suppress extreme values, making abnormally large s ij is smoothed.

[0113] It is understandable that the output of the constructed and regularized optimized hypergraph model is in is a node set, representing the set of all products; ε is a hyperedge set, representing the collaborative relationship between products. Each hyperedge e k With weight w(e k ) for subsequent synergy learning and prediction.

[0114] The nodes of the hypergraph model H contain multi-dimensional information about the products (sales, promotions, etc.), and the hyperedges describe the collaborative sales effects of these products in a specific time period and region. The regularized hyperedge set ε has a concise and meaningful structure.

[0115] Output: Hypergraph model H

[0116] Output: Hypergraph model It is the final result of this step and serves as the input of the subsequent Hypergraph Convolutional Network (HGCN). This model fully captures the synergy between products and ensures the comprehensiveness and effectiveness of information in the collaborative learning process.

[0117] Through this step, the hypergraph model not only effectively solves the problem of cross-product synergy modeling, but also optimizes complexity and outlier processing through regularization terms, providing a solid foundation for subsequent collaborative learning. The entire process meets the patent's requirements for synergy analysis of large-scale sales data, and innovative designs are made for the multidimensionality and dynamic change characteristics of the data to ensure that the model can efficiently and accurately handle complex cross-product relationships.

[0118] S3. A hypergraph convolutional network is designed to learn the hypergraph model and optimize the synergy between product nodes. The final feature representation of the prediction obtained after training the hypergraph convolutional network is mapped to a synergy score.

[0119] Furthermore, the update rule of the hypergraph convolutional network is as follows:

[0120] For each node p i ,The goal of the hypergraph convolution operation is to aggregate features from its neighboring nodes and propagate features through the convolution formula, which is expressed as follows:

[0121]

[0122] in, Represents node p i Feature representation at convolutional layer l+1; Represents the hyperedge e k Other nodes p j The features in the convolutional layer l; w(e k ) represents the hyperedge e kThe weight is used to measure the synergy between nodes; Represents the hyperedge e k The number of connected nodes, used as a normalization factor; W (l) represents the learnable weight matrix of the convolutional layer l, which is used for linear transformation of features; σ represents the activation function, which is used to introduce nonlinear operations;

[0123] Through this convolution operation, the node p i It not only combines its own characteristics, but also integrates the characteristics of the products that it sells in collaboration with, which can effectively capture the sales potential between products. k ) in each convolution layer plays a role in regulating the influence of different products on the synergy of the node. The more nodes in the hyperedge, the higher the normalization factor. This ensures that there is no excessive shifting of features when they are aggregated.

[0124] Then perform multi-layer convolution:

[0125] Initial Input For node p i The eigenvector x i , after multiple layers of hypergraph convolution, it is updated to That is, the final feature representation in the convolution layer L, where each layer of convolution updates the feature representation of the node. Each layer of convolution updates the feature representation of the node so that it gradually contains more information from the associated nodes.

[0126] Furthermore, the loss function of the hypergraph model includes node classification loss and synergy effect regularization loss;

[0127] The node classification loss is that some nodes have known synergy effect labels in historical data, and the cross entropy loss is used to train the hypergraph model to classify the known synergy effect labels, which is expressed as follows:

[0128]

[0129] Among them, y i is node p i The real label, is the label predicted by the model; the cross entropy loss will guide the model to correctly classify the synergy of nodes on the hypergraph.

[0130] In order to ensure that the synergy weights w(e k ) can be consistent with the actual synergy effect of the product. The present invention designs an innovative regularization term. This regularization term not only limits the hyperedge weight from being too large, but also guides the training of the model by combining the known synergy effect matrix S, and designs the regularization L regThe hyperedge weight is limited to be too large, and the model training is guided by combining the known synergy matrix S, which is expressed as follows:

[0131]

[0132] Among them, w(e k ) is the synergy weight; the regularization term is obtained by minimizing the hyperedge weight w(e k ) and the synergy effect between the nodes it connects ij The difference between the two ensures that the weights learned by the convolutional network can accurately reflect the synergy in the actual sales data. This not only improves the accuracy of the model, but also effectively prevents overfitting.

[0133] A special regularization term L is also introduced for some extreme cases special,extreme To suppress the occurrence of abnormal situations, the regular term is used to punish node p i Connected by too many hyperedges, represented as follows:

[0134]

[0135] Among them, |ε(p i )| indicates connection to node p i The hyperedge number is N, where N is the total number of hyperedges in the hypergraph. This item is used to prevent some products from excessively participating in irrelevant collaborative relationships, making the model more balanced and generalized.

[0136] The total loss function is expressed as follows:

[0137] L total =L cls +λ 1 L reg +λ 2 L special,extreme

[0138] Among them, λ 1 and λ 2 It is a hyperparameter that weighs various losses and is used to control the strength of regularization. By adjusting these parameters, the present invention can ensure that the model can accurately learn the synergy of products and effectively avoid overfitting or abnormal situations.

[0139] After the hypergraph convolutional network training, each product node p i The final feature representation Mapped to synergy score The score reflects the synergistic sales potential of the product in the overall sales network. These synergy scores will be used in subsequent steps to generate more targeted sales strategies and product recommendations.

[0140] S4. Construct a dynamic strategy generation model based on the synergy effect score. The dynamic strategy generation model generates real-time product recommendations and promotion strategies based on real-time product sales data, and outputs the optimal strategy.

[0141] Specifically, the real-time data flow D in the sales system real-time Contains multi-dimensional information, including new order data, user behavior data, inventory data, etc. This data is continuously updated while the system is running, reflecting the real-time sales of the product.

[0142] Order data D order : Includes new orders generated within each time window (such as every 5 seconds or every minute), and the specific content includes product ID, sales quantity, order amount, discount and other information.

[0143] User behavior data user : Reflects users’ real-time interactive behaviors, such as clicks, browsing history, user search data, the frequency of adding products to shopping carts, etc.

[0144] Inventory data D inventory :Contains the current inventory status of the product i This data is crucial for strategy generation. Products with too low inventory may need to have their recommendation frequency reduced, while products with too high inventory may need their recommendation priority increased.

[0145] The present invention divides these real-time data streams into time windows, each of which has a length of T (e.g., every 5 seconds, every minute). Within each window, the system will analyze the data in real time and combine it with the synergy effect score of the product. To dynamically adjust the strategy.

[0146] Furthermore, the steps of constructing the dynamic strategy generation model include:

[0147] Based on real-time order data order Calculate each node p i , that is, the real-time sales growth rate r of each product i (t), represents the sales growth rate of the product in time t, expressed as follows:

[0148]

[0149] Among them, S i (t) represents product p i Total sales at time t, S i (tT) represents the sales volume of the previous time window, ∈ is a smoothing term to avoid division by zero error; r i (t) used to reflect real-time sales trends of products;

[0150] Among them, inventory status Ii For each product p i The inventory status reflects the current inventory quantity, denoted as I i When the inventory level approaches the upper or lower limit, the recommended strategy will automatically adjust:

[0151] When I i When it is too low, you need to reduce the recommendation frequency of the product to avoid insufficient inventory;

[0152] When I i When it is high, the recommendation frequency of the product needs to be increased to reduce inventory pressure.

[0153] Score based on product synergy Real-time sales growth rate i (t) and inventory status I i Design objective function G i (t), expressed as follows:

[0154]

[0155] Among them, G i (t) represents the product p at time t i The strategy score is used to evaluate the product recommendation priority; κ, μ, ν represent the weight coefficients of controlling the influence of each factor; Indicates product p i The synergy score indicates the potential of the product to be sold together with other products; i Indicates product p i Inventory status. When the inventory is low, the recommendation weight will be weakened;

[0156] In the process of strategy optimization, a special regularization term L is designed special-hot , balances the diversity of strategy generation and avoids over-recommending certain popular products, which results in the neglect of sales of other products. This regularization term penalizes products with high concentration in strategy scores. The formula is as follows:

[0157]

[0158] Among them, δ represents the weight of the regularization term. The regularization term is scored by the suppression strategy S i (t) There are too many products that are too concentrated (i.e., some products have too high scores and occupy the recommendation list for a long time), encouraging more products to be recommended and maintaining the diversity of strategies.

[0159] Understandable, according to the strategy score S i(t), the system can generate a real-time recommendation list. The top K products with the highest scores will be recommended to users first. The system will dynamically adjust the recommendation list based on the synergy score, sales trend and inventory status to ensure that products with high synergy are recommended first.

[0160] Real-time promotion strategy is based on product combinations with high synergy scores. Based on sales performance within the current time window i (t) and inventory status I i , the system can dynamically adjust product discounts or promotion methods to ensure that the best promotion strategy is implemented. By monitoring inventory status in real time i The system can actively recommend products with high inventory pressure to increase their sales frequency. For products with tight inventory, the system will automatically lower their recommendation priority to maintain inventory safety.

[0161] The final output includes the following parts:

[0162] Product recommendation list: The generated recommendation list will be pushed to the e-commerce platform or user front end;

[0163] Promotional mix: The system generates dynamic promotional strategies and pushes them to users in real time;

[0164] Inventory adjustment plan: For products with high inventory, the system will generate inventory adjustment suggestions to assist operations in inventory management.

[0165] Feedback mechanism: The system will continuously update the real-time data stream based on user behavior feedback (such as click-through rate, purchase conversion rate) real-time , and then adjust the recommendation and promotion strategies to achieve closed-loop optimization.

[0166] S5. Apply the optimal strategy to the user front end to collect feedback data, and design feedback factors based on the feedback data. The feedback factors are the differences between the actual click-through rate and purchase conversion rate and the synergy score predicted by the hypergraph model. Finally, the feedback factors are integrated into the new process of the hypergraph convolutional network to optimize the prediction ability of the hypergraph model. At the same time, a time decay regularization mechanism is designed to control the hypergraph model’s dependence on historical data, so as to achieve a complete feedback closed-loop optimization.

[0167] Specifically, step 4 generates real-time product recommendations, promotion strategies, and inventory management strategies. After these strategies are applied to the user front end, feedback data will be generated. This invention records these data as F real-time , including the following three types of information:

[0168] Click Feedback F click : Records the frequency of users clicking on recommended products.

[0169] Purchase Feedback F purchase: Records the purchase conversion rate of users for recommended products.

[0170] Inventory Feedback inventory : Reflects the real-time inventory changes of recommended products.

[0171] These feedback data are a reflection of the actual market performance of the previous round of recommendations or promotion strategies, and will serve as input to help the present invention adjust the parameters in the model and update the product synergy score. and strategy generation mechanism to form a closed-loop optimization process.

[0172] Further, product synergy effects are scored through feedback data Updates are made so that the model can better adapt to real-time market performance in future strategy generation.

[0173] In order to quantify the difference between actual market performance and predicted synergy effect, the present invention is based on feedback data F real-time A feedback factor Δy is designed i The feedback factor is calculated by comparing the difference between the actual click-through rate and purchase conversion rate and the synergy score predicted by the model, and is designed as follows:

[0174]

[0175] Among them, Δy i represents the feedback factor, represents the product p i Synergy score of actual performance and model prediction The difference between them; ρ represents the weight of click feedback, which is used to control the weight of click feedback in adjusting the synergy score; ω represents the weight of purchase feedback, which is used to control the weight of purchase feedback in adjusting the synergy score; F click (p i ) indicates product p i The click feedback indicates the frequency of users clicking on the recommended products; purchase (p i ) indicates product p i The purchase feedback indicates the actual purchase conversion rate of users for recommended products.

[0176] After obtaining the feedback factor Δy i Afterwards, the present invention needs to integrate it into the update process of the hypergraph convolutional network to optimize the prediction ability of the model. i It is introduced into the loss function of the hypergraph convolutional network to serve as a guide for updating model parameters in future training processes.

[0177] Update mechanism of hypergraph convolutional network: parameter W of hypergraph convolutional network (l)The update is performed by minimizing the feedback adjustment loss. The present invention designs a loss function L that combines feedback data feedback , which quantifies the synergy score predicted by the model Compared with the actual performance (score y after feedback correction) i ):

[0178]

[0179] Among them, L feedback represents the feedback adjustment loss, which is used to minimize the predicted synergy score of the product and the corrected score y after actual feedback i The difference between i represents the product p after feedback adjustment i The real synergy score represents the synergy performance corrected by user click and purchase feedback.

[0180] It can be understood that by integrating the feedback factor into the loss function, the model can adaptively update the hyperedge weights and node features in real time, capture the ever-changing user behavior patterns, and optimize future strategy generation.

[0181] Furthermore, the synergy effect and user interest of products are not static and may decay or increase over time. Therefore, the present invention introduces a time decay regularization mechanism L time , which is used to control the model’s reliance on historical data and ensure that the model can quickly adapt to new market changes.

[0182] Furthermore, the time decay regularization formula is: The hyperedge weight w(e k ) The update at time t should take into account its changing trend in the past time tT. The time decay regularization term designed in the present invention limits the drastic changes in weights, so that the model maintains stability in a dynamic environment:

[0183]

[0184] Among them, L time represents the time decay regularization term, which limits the drastic changes of hyperedge weights in different time periods; w(e k ,t) represents the hyperedge e k The weight at the current time t; w(e k ,tT) represents the hyperedge e kThe weight in the previous time window tT; λ represents the time decay coefficient, which is used to control the model's reliance on historical synergies. By introducing the time decay regularization term, the model can dynamically adjust the impact of historical data on the current strategy, avoid over-reliance on historical synergy weights, and better adapt to market changes.

[0185] Furthermore, in combination with feedback data and feedback factor Δy i and the time-decay regularization term L time Finally, the present invention realizes a complete feedback closed-loop optimization process. This process ensures that the system can dynamically adjust product synergy scoring and strategy generation, and adaptively optimize when new real-time data is input.

[0186] Complete closed-loop optimization process:

[0187] The system executes real-time strategies and collects user feedback on clicks, purchases, and inventory;

[0188] Calculate the feedback factor Δy based on the feedback data i , used to modify the synergy score of the product;

[0189] The modified synergy score is incorporated into the training of the hypergraph convolutional network, and the network weights W are updated. (l) To optimize the model;

[0190] Apply the time decay regularization mechanism L time , ensuring that the model adapts to real-time data changes and avoids over-reliance on historical data;

[0191] The updated model will be used for the next round of strategy generation, forming an adaptive closed-loop optimization.

[0192] Furthermore, in order to ensure that the model can quickly adapt to market changes, the system will update the model at fixed time intervals (such as every hour or every day) based on real-time feedback data. In each time interval, new feedback data will be used to retrain the model, and the hyperedge weights and node features will be dynamically adjusted through the above feedback mechanism and regularization terms. During the continuous updating process, the model can dynamically adjust parameters such as α, β, λ to adapt to the needs of different market cycles or promotional activities. For example, during large-scale promotions, it may be necessary to increase the purchase weight β in the feedback data, while during regular sales, it is more dependent on the weight α of click feedback.

[0193] This step is based on the feedback data of the real-time strategy generated in step 4. Through the closed-loop optimization mechanism, the weights of the hypergraph convolutional network and the product synergy score are continuously adjusted to ensure that the model can adapt to the changing market demand. Through the time decay regularization mechanism, the model can avoid over-reliance on historical data and enhance the adaptability of the model in a dynamic market. Importantly,

[0194] This step designs a feedback factor Δy i The model adaptive update mechanism is introduced, and the time decay regularization term L is introduced. time , ensuring a smooth transition of synergy weights in different periods. The solution combines real-time feedback and historical data, and through a closed-loop optimization process, ensures that the model can continuously adjust recommendation and promotion strategies to improve the accuracy and stability of actual sales results.

[0195] The above is a preferred embodiment of the present invention. It should be pointed out that a person skilled in the art can make several improvements and modifications without departing from the principle of the present invention. These improvements and modifications are also considered to be within the scope of protection of the present invention.

Claims

1. A sales data analysis method based on big data and cloud computing, characterized in that: The method comprises: S1. Collect product sales data from multiple channels, clean and process them by dimension partitions, wherein the dimension partition storage process is partition storage and constructing a synergy matrix S in the partition storage process to represent the synergy relationship between products, and output the processed data; S2. Construct a hypergraph model based on the processed data. Each node of the hypergraph model represents a product, and the hyperedge represents the collaborative relationship between multiple products. Each hyperedge has a corresponding weight. Then, a regularization term is designed to optimize the hypergraph. Finally, the optimized hypergraph model H is input, H = (V, ε), where V is a node set, representing the set of all products; ε is a hyperedge set, representing the collaborative relationship between products, and each hyperedge e k With weight w(e k ); S3. Design a hypergraph convolutional network to learn the hypergraph model and optimize the synergy between product nodes. The final feature representation of the prediction obtained after training the hypergraph convolutional network is mapped to a synergy score. S4. constructing a dynamic strategy generation model according to the synergy effect score, wherein the dynamic strategy generation model generates real-time product recommendations and promotion strategies according to the real-time product sales data flowing in, and outputs the optimal strategy; S5. Apply the optimal strategy to the user front end to collect feedback data, and design feedback factors based on the feedback data. The feedback factors are the differences between the actual click-through rate and purchase conversion rate and the synergy score predicted by the hypergraph model. Finally, the feedback factors are integrated into the new process of the hypergraph convolutional network to optimize the prediction ability of the hypergraph model. At the same time, a time decay regularization mechanism is designed to control the hypergraph model’s dependence on historical data, and achieve complete feedback closed-loop optimization.

2. The sales data analysis method based on big data and cloud computing according to claim 1 is characterized in that: The dimension partition storage process includes the following steps: The cleaned data is partitioned and stored according to the time, product and region dimensions, defined as D partition , the specific operations are as follows: The time dimension is divided into days, weeks, and months; The product dimension is segmented by brand and category; The regional dimension is divided by city and region; Construct the synergy matrix S to represent the synergy relationship between products, where the synergy s ij Calculated by the following formula: Among them, s ij represents the synergy effect between products i and j; represents the sales data of product i at time t and region g; represents the sales data of product j at time t and region g; ∈ represents a small positive number used to prevent division by zero; the synergy matrix S is used for compression and index optimization, and t represents time.

3. The sales data analysis method based on big data and cloud computing according to claim 1 is characterized in that: The hypergraph model structure is expressed as follows: Each product i It is a node in the hypergraph. The attributes of the node are the sales data of the corresponding product in different periods and regions. t,pi,g To define; A super edge k Connecting multiple products indicates that these products have significant synergy in the same space and time. The construction of hyperedges is based on the synergy matrix S, whose elements s ij Representative product i and p j If the synergy between multiple products reaches the set threshold, they are connected on the same hyperedge; At the same time, design the synergistic effect threshold θ s , only when s ij >θ s Only when , products i and j are connected to the same hyperedge.

4. The sales data analysis method based on big data and cloud computing according to claim 3 is characterized in that: The design regularization term optimizes the hypergraph, specifically: In order to control the size of the hyperedge and avoid excessive connection of irrelevant products, a regularization term R(e k ), adjust the complexity of the hyperedge, expressed as follows Among them, R(e k ) represents the hyperedge e k The regularization term is used to balance the synergy strength and hyperedge complexity; α represents the weight for controlling the synergy strength; β represents the penalty coefficient for adjusting the number of nodes in the hyperedge to limit the excessive complexity of the hyperedge; k represents the hyperedge e k The number of nodes included; s ij represents the synergy between nodes i and j; At the same time, in order to deal with the possible noise and abnormal synergy effects in large-scale sales data, a special regularization term L is designed special , which is used to further constrain the synergy matrix S, is expressed as follows: Among them, L special represents a special regularization term, which is used to smooth out excessive synergy effect values ​​and prevent abnormal growth of synergy effects between certain products; γ represents the intensity of controlling the regularization term; exp(-s ij ) is used to suppress extreme values, making abnormally large s ij is smoothed.

5. The sales data analysis method based on big data and cloud computing according to claim 1 is characterized in that: The update rules of the hypergraph convolutional network are as follows: For each node p i ,The goal of the hypergraph convolution operation is to aggregate features from its neighboring nodes and propagate features through the convolution formula, which is expressed as follows: in, Represents node p i Feature representation at convolutional layer l+1; Represents the hyperedge e k Other nodes p j The features in the convolutional layer l; w(e k ) represents the hyperedge e k The weight is used to measure the synergy between nodes; Represents the hyperedge e k The number of connected nodes, used as a normalization factor; W (l) represents the learnable weight matrix of the convolutional layer l, which is used for linear transformation of features; σ represents the activation function, which is used to introduce nonlinear operations; Then perform multi-layer convolution: Initial Input For node p i The eigenvector x i , after multiple layers of hypergraph convolution, it is updated to That is, the final feature representation in the convolutional layer L, where each layer of convolution updates the feature representation of the node.

6. The sales data analysis method based on big data and cloud computing according to claim 1 is characterized in that: The loss function of the hypergraph model includes node classification loss and synergy effect regularization loss; The node classification loss is that some nodes have known synergy effect labels in historical data, and the cross entropy loss is used to train the hypergraph model to classify the known synergy effect labels, which is expressed as follows: Among them, y i is node p i The real label, is the label predicted by the model; At the same time, design the regularization L reg The hyperedge weight is limited to be too large, and the model training is guided by combining the known synergy matrix S, which is expressed as follows: Among them, w(e k ) is the synergy effect weight; A special regularization term L is also introduced for some extreme cases special,extreme To suppress the occurrence of abnormal situations, the regular term is used to punish node p i Connected by too many hyperedges, represented as follows: Among them, |ε(p i )| indicates connection to node p i The number of hyperedges in the hypergraph, N is the total number of hyperedges in the hypergraph; The total loss function is expressed as follows: L total =L cls +λ1L reg +λ2L special,extreme Among them, λ i and λ2 are hyperparameters that weigh various losses and are used to control the strength of regularization.

7. The sales data analysis method based on big data and cloud computing according to claim 1 is characterized in that: The real-time streaming product sales data includes order data, user behavior data and inventory data.

8. The sales data analysis method based on big data and cloud computing according to claim 7 is characterized in that: The steps of constructing the dynamic strategy generation model include: Based on real-time order data order Calculate each node p i , that is, the real-time sales growth rate r of each product i (t), represents the sales growth rate of the product in time t, expressed as follows: Among them, S i (t) represents product p i Total sales at time t, S i (tT) represents the sales volume of the previous time window, ∈ is a smoothing term to avoid division by zero error; r i (t) used to reflect real-time sales trends of products; Score based on product synergy Real-time sales growth rate i (t) and inventory status I i Design objective function G i (t), expressed as follows: Among them, G i (t) represents the product p at time t i The strategy score is used to evaluate the product recommendation priority; κ, μ, ν represent the weight coefficients of controlling the influence of each factor; Indicates product p i The synergy score indicates the potential of the product to be sold together with other products; i Indicates product p i Inventory status. When the inventory is low, the recommendation weight will be weakened; In the process of strategy optimization, a special regularization term L is designed special-hot , balances the diversity of strategy generation and avoids over-recommending certain popular products, which results in the neglect of sales of other products. This regularization term penalizes products with high concentration in strategy scores. The formula is as follows: Among them, δ represents the regularization term weight.

9. The sales data analysis method based on big data and cloud computing according to claim 1, characterized in that: The feedback factor is designed as follows: Among them, Δy i represents the feedback factor, represents the product p i Synergy score of actual performance and model prediction The difference between them; ρ represents the weight of click feedback, which is used to control the weight of click feedback in adjusting the synergy score; ω represents the weight of purchase feedback, which is used to control the weight of purchase feedback in adjusting the synergy score; F click (p i ) indicates product p i The click feedback indicates the frequency of users clicking on the recommended products; purchase (p i ) indicates product p i The purchase feedback indicates the actual purchase conversion rate of users for recommended products.

10. The sales data analysis method based on big data and cloud computing according to claim 9, characterized in that: The feedback factor is introduced into the loss function of the hypergraph convolutional network to serve as a guide for updating model parameters in future training processes, as shown below: Among them, L feedback represents the feedback adjustment loss, which is used to minimize the predicted synergy score of the product and the corrected score y after actual feedback i The difference between The time decay regularization mechanism is expressed as follows: Among them, L time represents the time decay regularization term, which limits the drastic changes of hyperedge weights in different time periods; w(e k ,t) represents the hyperedge e k The weight at the current time t; w(e k ,tT) represents the hyperedge e k The weight in the previous time window tT; λ represents the time decay coefficient, which is used to control the degree of dependence of the hypergraph model on historical synergy effects.

Images