Self-adaptive digital commercial decision reasoning optimization system and method

Abstract

The invention relates to the technical field of business intelligent decision, and provides a self-adaptive digital business decision reasoning optimization system and method, and the system comprises an acquisition module which collects the historical inflow / outflow data of a business entity; the classification module classifies entities through an iterative clustering algorithm based on commercial attributes and historical data; the calculation module quantifies an inflow / outflow influence value and generates a decision parameter matrix by analyzing historical data interaction between similar entities, and extracts first commercial data reflecting a change rule in combination with historical data; the reasoning module integrates the spatial-temporal characteristics through a decision reasoning engine, the time sequence analysis module processes the time sequence data to generate second commercial data, and the spatial relation processing module calculates entity association weights through a position encoder, processes the spatial-temporal characteristics in combination with a residual network and Transform, and finally outputs a commercial decision result of the next period.

Description

Technical Field

[0001] The present invention relates to the technical field of business intelligence decision-making, and more specifically, the present invention relates to an adaptive digital business decision-making reasoning optimization system and method. Background Art

[0002] In today's rapidly developing digital business environment, the formulation of business decisions has become increasingly complex and challenging. Traditional business decision-making methods mainly rely on manual experience analysis and simple data statistics methods. In the face of massive and complex business data, it is often difficult to accurately grasp the dynamic relationships between business entities and their changing patterns over time, resulting in insufficient scientific and forward-looking decision-making. With the rise of technologies such as big data and artificial intelligence, although some decision support systems based on data analysis have emerged, most of these systems only focus on single-dimensional data analysis, such as simple time series analysis or only considering spatial relationships, lacking comprehensive consideration of the multi-dimensional characteristics of business entities and being unable to comprehensively reflect the operating status and potential development trends of business entities. In addition, when dealing with the classification of business entities, existing systems often adopt fixed classification criteria, making it difficult to adapt to the dynamic changes in different business environments and business scenarios, lacking adaptability and flexibility.

[0003] In the process of implementing the embodiments of the present invention, the inventors found that there are at least the following problems or defects in the prior art: existing business decision-making systems cannot effectively integrate the historical inflow and outflow data, spatial relationship characteristics, and time series change laws of business entities, making it difficult to accurately predict the business decision results in the next decision-making cycle; in terms of the classification of business entities, there is a lack of an adaptive classification mechanism and it is unable to dynamically adjust the classification results according to the historical data of business entities and the business environment in which they are located, thus affecting the accuracy of decision-making; the calculation methods for the association weights between business entities in the prior art are relatively simple and cannot accurately reflect the complex interaction relationships between business entities; in the process of model training, there is a lack of an effective adaptive loss function, making it difficult to balance the prediction accuracy and generalization ability of the model, resulting in less than ideal performance of the model in actual applications. Summary of the Invention

[0004] The present invention provides an adaptive digital business decision-making reasoning optimization system and method.

[0005] In the first aspect of the present invention, an adaptive digital business decision-making reasoning optimization system is provided, including:

[0006] An acquisition module that acquires historical business data of each business entity, where the historical business data includes historical inflow data and historical outflow data;

[0007] Classification module, which classifies commercial entities based on the commercial attributes of the locations where the commercial entities are located and the historical commercial data corresponding to the commercial entities, and determines the categories to which the commercial entities belong;

[0008] Calculation module, which calculates the decision parameter matrix of each commercial entity based on the category to which each commercial entity belongs and the historical commercial data, including: determining other commercial entities in the category to which the target commercial entity belongs as the first commercial entities; determining first data based on the historical commercial data between the target commercial entity and each first commercial entity and the historical outflow data of the target commercial entity, where the first data represents the influence value of the outflow data of each first commercial entity on the target commercial entity; determining second data based on the historical commercial data between the target commercial entity and each first commercial entity and the historical inflow data of the target commercial entity, where the second data represents the influence value of the inflow data of each first commercial entity on the target commercial entity; obtaining the decision parameter matrix of the target commercial entity based on the first data and the second data; obtaining first commercial data based on the decision parameter matrix and the historical commercial data of the target commercial entity, and the first commercial data is used to reflect the historical change law of the target commercial entity;

[0009] Inference module, which processes the first commercial data and the spatial relationship features of each commercial entity obtained based on the historical outflow data by using the constructed decision inference engine, and infers the commercial decision result of the target commercial entity in the next decision cycle. The decision inference engine includes a time series analysis module and a spatial relationship processing module. The time series analysis module is used to process the input first commercial data to obtain second commercial data in the next decision cycle; the spatial relationship processing module includes processing the spatial relationship features of each input commercial entity through a position encoder to obtain the association weights between commercial entities; using a first residual network to process the association weights and the spatial relationship features of each commercial entity to obtain first feature information; processing the second commercial data and the first feature information through a Transformer network to obtain second feature information; obtaining the commercial decision result in the next decision cycle based on the second feature information.

[0010] Further, classifying commercial entities based on the commercial attributes of the locations where the commercial entities are located and the historical commercial data corresponding to the commercial entities, and determining the categories to which the commercial entities belong, including: the commercial attributes include retail, service, manufacturing, transportation hub, commercial center, and comprehensive;

[0011] Step 1: Randomly select the number of commercial entities consistent with the number of commercial attributes, and use their corresponding historical commercial data as the cluster centers;

[0012] Step 2: Calculate the correlation coefficients between the historical business data of the remaining business entities and the cluster centers, determine the maximum correlation coefficient of the remaining business entities and the corresponding cluster centers, and multiple clusters are formed by all business entities;

[0013] Step 3: Calculate the mean of the historical business data of the business entities in the same cluster, and use the mean as the new cluster center;

[0014] Step 4: Repeat Step 2 and Step 3 until the cluster centers no longer change.

[0015] Further, based on the first data and the second data, a decision parameter matrix of the target business entity is obtained, including: normalizing the first data based on the maximum value and the minimum value of the first data to obtain a normalized inflow feature;

[0016] Normalize the second data based on the maximum value and the minimum value of the second data to obtain a normalized outflow feature;

[0017] Perform element-wise non-linear transformation on the normalized inflow feature and the corresponding first parameter and the normalized outflow feature and the corresponding second parameter respectively, and add the transformation results to obtain a decision parameter matrix, where the first parameter and the second parameter are obtained based on historical business data.

[0018] Further, the spatial relationship processing module includes a plurality of first modules, a linear regression network, and a Softmax network connected in sequence; the first module includes a position encoder, a first residual network, two Transformer networks, a fully connected layer, a feedforward neural network, and a second residual network connected in sequence.

[0019] Further, the time series analysis module includes a plurality of LSTM networks connected in sequence. The LSTM network includes a time encoder, a multi-head attention mechanism, a third residual network, a feedforward neural network, a fully connected layer, and a fourth residual network connected in sequence; the input data is segmented into third business data of multiple consecutive time series by the time encoder, the multi-head attention mechanism learns the multiple third business data respectively, the third residual network and the feedforward neural network perform non-linear processing on the learning results, the fully connected layer processes the output results of the feedforward neural network to obtain global features; the fourth residual network learns the output results of the feedforward neural network and the output results of the fully connected layer to obtain the input results of the connected network, and the connected network includes the next LSTM network or the spatial relationship processing module.

[0020] Further, the spatial relationship features of the input business entities are processed by the position encoder to obtain the association weights between business entities, and its expression is:

[0021]

[0022] Among them, p i and p j respectively represent the positions of commercial entity i and commercial entity j in the business network, d is the dimension of the position encoding, α and β are parameters obtained by training based on historical business data, and W ij represents the association weight between commercial entity i and commercial entity j.

[0023] Furthermore, a decision inference engine is learned using a gradient-based adaptive loss function, and the expression of the loss function is:

[0024]

[0025] Among them, N is the number of samples, is the predicted value, y i is the true value, and λ is the balance parameter.

[0026] In the second aspect of the present invention, an adaptive digital business decision inference optimization method is provided, including:

[0027] Obtain the historical business data of each commercial entity, where the historical business data includes historical inflow data and historical outflow data;

[0028] Classify commercial entities based on the business attributes of the locations where each commercial entity is located and the historical business data corresponding to each commercial entity, and determine the category to which each commercial entity belongs;

[0029] Based on the category to which each commercial entity belongs and the historical business data, calculate the decision parameter matrix of each commercial entity, including: determining other commercial entities in the category to which the target commercial entity belongs as the first commercial entity; determining the first data based on the historical business data between the target commercial entity and each first commercial entity and the historical outflow data of the target commercial entity, where the first data represents the influence value of the outflow data of each first commercial entity on the target commercial entity; determining the second data based on the historical business data between the target commercial entity and each first commercial entity and the historical inflow data of the target commercial entity, where the second data represents the influence value of the inflow data of each first commercial entity on the target commercial entity; obtaining the decision parameter matrix of the target commercial entity based on the first data and the second data;

[0030] Obtain the first business data based on the decision parameter matrix and the historical business data of the target commercial entity, where the first business data is used to reflect the historical change law of the target commercial entity;

[0031] The constructed decision - making and reasoning engine processes the first commercial data and the spatial relationship features of each commercial entity obtained based on historical outflow data, and infers the commercial decision - making result of the target commercial entity in the next decision - making cycle. The decision - making and reasoning engine includes a time - series analysis module and a spatial relationship processing module. The time - series analysis module is used to process the input first commercial data to obtain the second commercial data in the next decision - making cycle. The spatial relationship processing module includes using a position encoder to process the spatial relationship features of each input commercial entity to obtain the association weights between commercial entities; using a first residual network to process the association weights and the spatial relationship features of each commercial entity to obtain first - feature information; using a Transformer network to process the second commercial data and the first - feature information to obtain second - feature information; and obtaining the commercial decision - making result in the next decision - making cycle based on the second - feature information.

[0032] Furthermore, classify commercial entities based on the commercial attributes of the locations where each commercial entity is located and the historical commercial data corresponding to each commercial entity, and determine the category to which each commercial entity belongs, including: the commercial attributes include retail, service, manufacturing, transportation hub, commercial center, and comprehensive categories;

[0033] Step 1: Randomly select the number of commercial entities equal to the number of commercial attributes, and use their corresponding historical commercial data as the cluster centers;

[0034] Step 2: Calculate the correlation coefficients between the historical commercial data of the remaining commercial entities and each cluster center, determine the maximum correlation coefficient of the remaining commercial entities and the corresponding cluster center, and all commercial entities form multiple clusters;

[0035] Step 3: Calculate the mean of the historical commercial data of the commercial entities in the same cluster, and use the mean as the new cluster center;

[0036] Step 4: Repeat Step 2 and Step 3 until the cluster centers no longer change.

[0037] Furthermore, based on the first data and the second data, obtain the decision - parameter matrix of the target commercial entity, including: normalizing the first data based on the maximum and minimum values of the first data to obtain the normalized inflow feature;

[0038] Normalizing the second data based on the maximum and minimum values of the second data to obtain the normalized outflow feature;

[0039] Perform element - level non - linear transformation on the normalized inflow feature and the corresponding first parameter and the normalized outflow feature and the corresponding second parameter respectively, and add the transformation results to obtain the decision - parameter matrix, where the first parameter and the second parameter are obtained based on historical commercial data.

[0040] The above embodiments of the present invention have at least the following beneficial effects: The present invention can improve the accuracy of business prediction and decision-making efficiency. Through the intelligent classification module, accurate grouping of business entities can be carried out to ensure the comparability of data among entities of the same type; the calculation module can deeply explore the potential correlation rules among business entities by quantifying the influence values of inflows and outflows and constructing a dynamic parameter matrix; the inference engine integrates time series analysis and spatial relationship processing technologies, and can capture the temporal evolution characteristics and spatial interaction relationships of business data simultaneously, so as to generate more reliable business trend predictions. Based on its full-process automated processing ability, the system can realize an intelligent closed-loop from data collection to decision output, continuously improve the classification accuracy through an iterative optimization clustering algorithm, and the dynamic calculation of the parameter matrix can reflect the changes in the business environment in real time.

[0041] The system can enhance the adaptability and intelligence level of business decisions. Time series analysis using the LSTM network combined with the attention mechanism can accurately identify long-term dependence relationships, and spatial relationship processing based on position encoding and Transformer can accurately model the spatial influence among business entities. By optimizing the model parameters through an adaptive loss function, the system performance can be continuously improved. Its unique normalization processing technology can eliminate the difference in data dimensions, and the non-linear transformation can enhance the feature expression ability. The finally output business decision suggestions can help enterprises optimize resource allocation, reduce operation risks, and provide data support for strategic planning, so as to maximize business value. The overall architecture of the system can flexibly adapt to the needs of different business scenarios such as retail, manufacturing, and services, and has broad application prospects and business value. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] By reading the following detailed description with reference to the accompanying drawings, the above and other objects, features, and advantages of the exemplary embodiments of the present invention will become readily understood. In the drawings, several embodiments of the present invention are shown in an exemplary rather than restrictive manner, wherein:

[0043] Figure 1 It is a schematic structural diagram of an adaptive digital business decision inference optimization system provided by an embodiment of the present invention;

[0044] Figure 2 It is a schematic flowchart of an adaptive digital business decision inference optimization method provided by an embodiment of the present invention;

[0045] Figure 3 It schematically shows a schematic structural diagram of an electronic device according to an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0046] The principles and spirit of the present invention will be described below with reference to several exemplary embodiments. It should be understood that these embodiments are provided only to enable those skilled in the art to better understand and implement the present invention, rather than to limit the scope of the present invention in any way. On the contrary, these embodiments are provided to make the present invention more thorough and complete, and to convey the scope of the present invention fully to those skilled in the art.

[0047] Those skilled in the art know that the embodiments of the present invention can be implemented as a system, a device, a device, a method, or a computer program product. Therefore, the present invention can be specifically implemented in the following forms, namely: complete hardware, complete software (including firmware, resident software, microcode, etc.), or a combination of hardware and software.

[0048] It should be noted that the number of any element in the drawings is for illustration rather than limitation, and any naming is only for distinction and does not have any limiting meaning.

[0049] The following reference Figure 1 , Figure 1 is a schematic structural diagram of an adaptive digital business decision-making inference optimization system provided for an embodiment of the present invention. As Figure 1 shown, an adaptive digital business decision-making inference optimization system includes:

[0050] An acquisition module 101 that acquires historical business data of each business entity, where the historical business data includes historical inflow data and historical outflow data;

[0051] A classification module 102 that classifies business entities based on the business attributes of the locations where the business entities are located and the historical business data corresponding to the business entities, and determines the categories to which the business entities belong;

[0052] A calculation module 103 that calculates a decision parameter matrix of each business entity based on the category to which the business entity belongs and the historical business data, including: determining other business entities in the category to which the target business entity belongs as the first business entities; determining first data based on the historical business data between the target business entity and each first business entity and the historical outflow data of the target business entity, where the first data represents the influence value of the outflow data of each first business entity on the target business entity; determining second data based on the historical business data between the target business entity and each first business entity and the historical inflow data of the target business entity, where the second data represents the influence value of the inflow data of each first business entity on the target business entity; obtaining a decision parameter matrix of the target business entity based on the first data and the second data; and obtaining first business data based on the decision parameter matrix and the historical business data of the target business entity, where the first business data is used to reflect the historical change law of the target business entity;

[0053] The reasoning module 104 uses the constructed decision reasoning engine to process the first business data and the spatial relationship features of each business entity obtained based on the historical outflow data, and infers the business decision results of the target business entity in the next decision cycle. The decision reasoning engine includes a time series analysis module and a spatial relationship processing module. The time series analysis module is used to process the input first business data to obtain the second business data of the next decision cycle; the spatial relationship processing module includes processing the spatial relationship features of each input business entity through a position encoder to obtain the association weights between the business entities; using the first residual network to process the association weights and the spatial relationship features of each business entity to obtain the first feature information; processing the second business data and the first feature information through a Transformer network to obtain the second feature information; and obtaining the business decision results of the next decision cycle based on the second feature information.

[0054] It should be noted that the core of this system is to collect historical business data of each business entity through the acquisition module. These data include historical inflow data and historical outflow data. The historical inflow data refers to the data records of various resources (such as funds, goods, customer traffic, etc.) received by the business entity in the past time period, while the historical outflow data refers to the data records of various resources output by the business entity in the past time period. The classification module is used to classify the business entity according to its business attributes and the corresponding historical business data to determine its category. The business attribute refers to the industry type or functional positioning of the business entity, such as retail, service, etc. The calculation module calculates the decision parameter matrix based on the category of the business entity and the historical business data. The matrix is a mathematical structure used to quantify the mutual influence relationship between business entities, including the influence value of inflow and outflow data. The reasoning module uses the constructed decision reasoning engine, combined with time series analysis and spatial relationship processing, to comprehensively process the historical data and spatial relationship characteristics of the business entity, so as to infer the business decision results of the next decision cycle. The decision reasoning engine is the core processing unit of the system. It realizes the optimization reasoning of business decisions through the coordinated work of the time series analysis module and the spatial relationship processing module.

[0055] Specifically, a business entity refers to a unit with independent economic behavior and financial accounting in business activities, such as a store, a factory, or a logistics center, etc. Historical business data are various data records generated by business entities during a past time period, including but not limited to data such as capital flow, goods in and out, and customer flow. When classifying business entities, the classification module makes a comprehensive judgment based on the business attributes of the locations where the business entities are located, such as retail, service, manufacturing, etc., in combination with their historical business data. The first data refers to the influence value of the outflow data of each first business entity (i.e., other business entities of the same category as the target business entity and having an impact on it) on the target business entity, that is, the degree of influence of the outflow data of the first business entity on the target business entity; the second data refers to the influence value of the inflow data of each first business entity on the target business entity, that is, the degree of influence of the inflow data of the first business entity on the target business entity. The decision parameter matrix is obtained by adding the normalized inflow features and outflow features after non-linear transformation with corresponding parameters, and these parameters are trained based on historical business data and are used to adjust and optimize the calculation results of the decision parameter matrix. The time series analysis module processes the first business data to predict the second business data in the next decision cycle, while the spatial relationship processing module analyzes and processes the spatial relationship features between business entities through components such as a position encoder, a residual network, and a Transformer network to obtain the association weights and feature information between business entities.

[0056] Preferably, the construction process of the decision inference engine in the inference module is as follows: First, the time series analysis module uses multiple LSTM networks, and each LSTM network includes components such as a time encoder, a multi-head attention mechanism, a residual network, a feed-forward neural network, a fully connected layer, and a residual network. The time encoder divides the input data into multiple data segments of consecutive time series, the multi-head attention mechanism learns these data segments to extract key features, the residual network and the feed-forward neural network perform non-linear processing on the learning results, and the fully connected layer integrates the processed results to obtain global features. The spatial relationship processing module processes the spatial relationship features of business entities through a position encoder to calculate the association weights between business entities, and the output result of the position encoder will be further processed through a residual network and a Transformer network to obtain the first feature information. Finally, the Transformer network fuses the second business data obtained by the time series analysis module and the first feature information obtained by the spatial relationship processing module to obtain the second feature information, and based on the second feature information, the business decision result in the next decision cycle is obtained. During the data processing process, for example, when normalizing the first data and the second data, they are scaled to the interval [0, 1] according to their maximum and minimum values to eliminate the dimensional difference and numerical range difference between different data, making the data more suitable for model processing.

[0057] In some embodiments, the business entities are classified based on the business attributes of the locations where the business entities are located and the historical business data corresponding to the business entities to determine the categories to which the business entities belong, including: the business attributes include retail, service, manufacturing, transportation hub, commercial center, and comprehensive categories;

[0058] Step 1: Randomly select the number of business entities equal to the number of business attributes, and use the corresponding historical business data as the cluster centers;

[0059] Step 2: Calculate the correlation coefficients between the historical business data of the remaining business entities and each cluster center, determine the maximum correlation coefficient of the remaining business entities and the corresponding cluster center, and all business entities form multiple clusters;

[0060] Step 3: Calculate the mean of the historical business data of the business entities in the same cluster, and use the mean as the new cluster center;

[0061] Step 4: Repeat Step 2 and Step 3 until the cluster centers no longer change.

[0062] It should be noted that the classification of business entities is based on the business attributes of their locations and the corresponding historical business data. Business attributes refer to the industry type or functional positioning of business entities, such as retail, service, manufacturing, etc. These attributes reflect the main functions and roles of business entities in economic activities. Through classification, entities with similar business characteristics can be grouped into one category to more accurately analyze and predict their business behaviors. The classification process includes steps such as randomly selecting cluster centers, calculating correlation coefficients, and updating cluster centers. These steps work together to enable business entities to be reasonably divided into different categories according to their historical data and business attributes, providing a basis for subsequent calculation of decision parameter matrices and business decision reasoning.

[0063] Specifically, the commercial attribute is an important basis for classifying commercial entities, including but not limited to retail, service, manufacturing, transportation hub, commercial center, and comprehensive categories, etc. Retail commercial entities are mainly involved in the sale of goods, service entities provide various services, manufacturing entities are responsible for product production, transportation hub entities are distribution centers for logistics and people flow, commercial center entities are usually concentrated areas for various commercial activities, and comprehensive entities include multiple functions. During the classification process, first, randomly select the same number of commercial entities as the number of commercial attributes, and use their corresponding historical commercial data as the cluster centers. The cluster center is the initial reference point in the classification process and is used to calculate the correlation between other commercial entities and it. The correlation coefficient is used to measure the similarity between the historical commercial data of a commercial entity and the cluster center. By calculating the correlation coefficient, it can be determined which cluster center has the highest similarity with each commercial entity, and then it is assigned to the corresponding cluster. The update of the cluster center is completed by calculating the mean value of the historical commercial data of all commercial entities in the same cluster. This process will be repeated continuously until the cluster center no longer changes, that is, it reaches the convergence state. At this time, the classification result is considered stable.

[0064] Preferably, during the classification process, the selection and update of the cluster center are key steps. The initial cluster center can be selected by random sampling, randomly selecting the historical commercial data of a commercial entity from each commercial attribute category as the starting point. When calculating the correlation coefficient, the Pearson correlation coefficient or other suitable statistical methods can be used, which can quantify the linear correlation between two data sets. When updating the cluster center, it is necessary to summarize and calculate the historical commercial data of all commercial entities in the same cluster to obtain the new mean value as the new cluster center. This process can be implemented through an iterative algorithm. Each iteration will recalculate the correlation coefficient and update the cluster center until the change in the cluster center is less than a preset threshold, such as less than 0.01, or reaches a preset maximum number of iterations, such as 100 iterations. In this way, the accuracy and stability of the classification result can be ensured, providing a reliable basis for subsequent business decisions.

[0065] In some embodiments, based on the first data and the second data, a decision parameter matrix of the target commercial entity is obtained, including: normalizing the first data based on the maximum and minimum values of the first data to obtain a normalized inflow feature;

[0066] Normalizing the second data based on the maximum and minimum values of the second data to obtain a normalized outflow feature;

[0067] Performing element-wise non-linear transformation on the normalized inflow feature and the corresponding first parameter and on the normalized outflow feature and the corresponding second parameter respectively, and adding the transformation results to obtain the decision parameter matrix, where the first parameter and the second parameter are obtained based on historical commercial data.

[0068] It should be noted that the decision parameter matrix is obtained based on the first data and the second data, and this process is an important step for the system to quantify the mutual influence relationship between business entities. The first data represents the influence value of each first business entity on the outflow data of the target business entity, while the second data represents the influence value of each first business entity on the inflow data of the target business entity. Through normalization processing and non-linear transformation, the system can convert these data into a decision parameter matrix, thereby providing a key quantitative basis for subsequent business decision-making reasoning. The construction of the decision parameter matrix involves the processing of inflow and outflow characteristics, as well as the application of the first parameter and the second parameter obtained based on historical business data. These parameters are obtained in the system learning and optimization process and are used to adjust and optimize the calculation results of the decision parameter matrix.

[0069] Specifically, the first data and the second data are the key inputs when the system calculates the decision parameter matrix. The first data reflects the influence degree of other business entities on the outflow data of the target business entity. For example, if a supplier (the first business entity) supplies goods to the target business entity (the retailer), then the shipment volume of the supplier (outflow data) will have a direct impact on the inventory of the retailer (inflow data). The second data reflects the influence degree of other business entities on the inflow data of the target business entity. For example, the quantity of goods received by the retailer from the supplier. The normalized inflow characteristics and the normalized outflow characteristics are obtained by normalizing the first data and the second data. The purpose of the normalization process is to scale the data to a unified range (such as [0, 1]) to eliminate the dimensional difference and numerical range difference between different data. The first parameter and the second parameter are obtained by training based on historical business data, and they are used to adjust the weights of the normalized characteristics to reflect the importance of different characteristics in the decision-making process.

[0070] Preferably, the specific steps of the normalization process are to divide the first data and the second data by the difference between their respective maximum and minimum values, and then scale the results to the range of [0, 1]. For example, if the maximum value of the first data is 100 and the minimum value is 10, then a data point with a value of 50 will become (50 - 10) / (100 - 10) = 0.44 after normalization. The non-linear transformation can be achieved through activation functions such as ReLU or Sigmoid, which can map the input data to a non-linear output space, thereby increasing the expressive power of the model. When calculating the decision parameter matrix, the normalized incoming features are multiplied by the first parameter and then added to the result of multiplying the normalized outgoing features by the second parameter to obtain the final decision parameter matrix. The specific values of the first parameter and the second parameter can be automatically adjusted during the training process through machine learning algorithms such as the gradient descent method to minimize the prediction error. In this way, the system can dynamically adjust the parameters according to historical data, thus more accurately reflecting the mutual influence relationship between business entities.

[0071] In some embodiments, the spatial relationship processing module includes a plurality of first modules, a linear regression network, and a Softmax network connected in sequence; the first module includes a position encoder, a first residual network, two Transformer networks, a fully connected layer, a feedforward neural network, and a second residual network connected in sequence.

[0072] It should be noted that the spatial relationship processing module is an important part of the system for processing the spatial relationship features of business entities. Its structure includes a plurality of sub-modules connected in sequence, such as a position encoder, a first residual network, two Transformer networks, a fully connected layer, a feedforward neural network, a second residual network, etc. These sub-modules work together to encode, transform, and fuse the spatial relationship features of business entities, and extract feature information that can reflect the association weights and spatial interaction patterns between business entities. The position encoder is responsible for converting the spatial position information of business entities into a processable feature vector, while the Transformer network uses its powerful parallel processing ability and self-attention mechanism to deeply analyze and extract the spatial relationships between business entities, thereby providing richer spatial dimension information for business decision-making reasoning.

[0073] Specifically, each sub-module in the spatial relationship processing module has its specific functions and roles. The position encoder converts the spatial position information of commercial entities (such as longitude and latitude, geographical regions, etc.) into a feature vector of a fixed dimension, so that the subsequent network can process it. The first residual network is used to alleviate the problem of gradient disappearance in deep networks. By introducing residual connections, the network can learn complex feature mapping relationships more effectively. The two Transformer networks use their self-attention mechanisms to model the spatial relationships between commercial entities, and can dynamically focus on the interaction strengths between different commercial entities. The role of the fully connected layer is to integrate and transform the extracted features, providing a more compact feature representation for subsequent processing. The feed-forward neural network further performs non-linear transformations on the features to enhance the expressive power of the model. The second residual network then uses residual connections again in the final stage to ensure the integrity and stability of the feature information. These sub-modules form a complete processing flow through sequential connection, enabling the spatial relationship features to be gradually extracted and optimized.

[0074] Preferably, the construction and optimization of the spatial relationship processing module can be achieved through the following steps. First, the input of the position encoder is the spatial position information of commercial entities, such as position coordinates represented by longitude and latitude, and the output is a feature vector of a fixed dimension. When designing the position encoder, a combination of sine and cosine functions can be used to map the position information into the feature space while retaining the periodicity and relativity of the position. For the Transformer network, its input is the feature vector processed by the position encoder, as well as the association weights between commercial entities (such as weights calculated through commercial transaction frequency or geographical distance). The self-attention mechanism in the Transformer network will calculate the attention scores between different commercial entities based on these inputs, thereby dynamically adjusting the weights of the features. In the fully connected layer and the feed-forward neural network, the performance of the model can be optimized by adjusting the number of layers and neurons in the network. For example, the output dimension of the fully connected layer can be set to 128 or 256, and the hidden layer dimension of the feed-forward neural network can be set to 512. Finally, the extracted features are further optimized through the second residual network to ensure the integrity and accuracy of the feature information, providing high-quality spatial relationship features for subsequent commercial decision-making inferences.

[0075] In some embodiments, the time series analysis module includes a plurality of LSTM networks connected in sequence. The LSTM network includes a time encoder, a multi-head attention mechanism, a third residual network, a feed-forward neural network, a fully connected layer, and a fourth residual network connected in sequence. The time encoder is used to split the input data into third commercial data of multiple continuous time series. The multi-head attention mechanism is used to learn the multiple third commercial data respectively. The third residual network and the feed-forward neural network perform non-linear processing on the learning results. The fully connected layer processes the output results of the feed-forward neural network to obtain global features. The fourth residual network is used to learn the output results of the feed-forward neural network and the output results of the fully connected layer to obtain the input results of the network connected thereto. The network connected thereto includes the next LSTM network or the spatial relationship processing module.

[0076] It should be noted that the time series analysis module is a key part of the system for processing the time characteristics of the historical data of commercial entities. It processes the input first commercial data through a plurality of LSTM networks to predict the second commercial data in the next decision cycle. The LSTM network is a special recurrent neural network structure that can effectively process the long-term dependence relationships in time series data. The time encoder splits the input data into data segments of multiple continuous time series. The multi-head attention mechanism learns these data segments to extract key features. The third residual network and the feed-forward neural network perform non-linear processing on the learning results. The fully connected layer integrates the output results of the feed-forward neural network to obtain global features. The fourth residual network is used to learn the output results of the feed-forward neural network and the output results of the fully connected layer to provide inputs for subsequent modules. In this way, the time series analysis module can capture the variation rules of commercial data in the time dimension and provide prediction results in the time series dimension for commercial decisions.

[0077] Specifically, the LSTM network in the time series analysis module is the core component, which consists of multiple sub-modules, including a time encoder, a multi-head attention mechanism, a third residual network, a feed-forward neural network, a fully connected layer, and a fourth residual network. The function of the time encoder is to split the input first commercial data into multiple consecutive time series segments in chronological order. For example, monthly data is split into multiple weekly data segments, so that the model can better process time series features. The multi-head attention mechanism can simultaneously focus on important features in multiple time series segments, and learn different subsets of features through multiple heads respectively, thereby improving the model's expression ability for time series data. The third residual network and the feed-forward neural network are used to perform non-linear transformations on the output of the multi-head attention mechanism to enhance the model's learning ability. The role of the fully connected layer is to further integrate the features after non-linear transformation and extract global features. The fourth residual network ensures the integrity and stability of feature information during the transmission process through residual connections, providing high-quality feature inputs for subsequent modules.

[0078] Preferably, the construction and optimization of the time series analysis module can be achieved through the following steps. First, the input of the time encoder is the first commercial data, such as time series data of the historical sales volume or inventory of a commercial entity. In the design of the time encoder, the segmentation granularity can be set according to the time granularity of the data (such as daily, weekly, monthly). For example, monthly data is split into four weekly data segments. The number of heads of the multi-head attention mechanism can be set according to the complexity of the data, such as set to 8 or 16. Each head will learn different parts of the input data, thereby extracting richer features. The specific parameter settings of the third residual network and the feed-forward neural network can be adjusted according to the scale and feature complexity of the data. For example, the hidden layer dimension of the feed-forward neural network is set to 512 or 1024. The output dimension of the fully connected layer can be set according to the requirements of subsequent modules, such as set to 128 or 256. Finally, the fourth residual network ensures the integrity of feature information through residual connections, providing reliable inputs for subsequent modules. Through these optimization steps, the time series analysis module can more effectively capture the time features of commercial data and provide more accurate prediction results for commercial decisions.

[0079] In some embodiments, the spatial relationship features of each input commercial entity are processed by a position encoder to obtain the association weights between commercial entities, and its expression is:

[0080]

[0081] where p i and p j respectively represent the positions of commercial entity i and commercial entity j in the commercial network, d is the dimension of the position encoding, α and β are parameters trained based on historical commercial data, and W ijRepresents the association weight between business entity i and business entity j.

[0082] It should be noted that the position encoder is used in the spatial relationship processing module to calculate the association weight between business entities. The association weight is a quantitative indicator that measures the degree of mutual influence between business entities and reflects the spatial relationship of business entities in the business network. The position encoder calculates the association weight between them based on the position information of business entities in the business network through a specific calculation formula. This calculation method can effectively reflect the spatial distance, geographical proximity or closeness of business connections between business entities, thus providing important basic data for subsequent spatial relationship processing.

[0083] Specifically, the input of the position encoder is the position information of business entities, usually represented by the position coordinates of business entities in the business network, such as longitude and latitude or other geographical space coordinates. The output of the position encoder is the association weight between business entities, which is a value between 0 and 1, indicating the degree of mutual influence between two business entities. The calculation formula of the association weight is based on the distance or other spatial features between business entities and is adjusted by specific parameters (such as parameters obtained by training based on historical business data) to reflect the actual connection strength between business entities. For example, if the distance between two business entities is relatively close, their association weight may be higher, indicating that they have a closer connection in business activities.

[0084] Preferably, the construction and optimization of the position encoder can be achieved through the following steps. First, determine the source of the position information of business entities, such as obtaining the precise longitude and latitude coordinates of business entities through a Geographic Information System (GIS). Then, select an appropriate distance metric method to calculate the distance between business entities, such as Euclidean distance or Manhattan distance. When calculating the association weight, parameters obtained by training based on historical business data can be introduced, and these parameters can be optimized through machine learning algorithms (such as gradient descent method) to better reflect the actual connection strength between business entities. For example, if historical data shows that there are frequent business transactions between certain business entities, even if their geographical distance is relatively far, the association weight between them can be increased by adjusting the parameters. In this way, the position encoder can calculate the association weight between business entities more accurately and provide high-quality input data for the subsequent spatial relationship processing module.

[0085] In some embodiments, a decision inference engine is learned using a gradient-based adaptive loss function, and the expression of the loss function is:

[0086]

[0087] where N is the number of samples, is the predicted value, yi is the true value, and λ is the balance parameter.

[0088] It should be noted that the gradient-based adaptive loss function mentioned in the present invention is a key technical means for optimizing the decision inference engine. The loss function is a quantitative index for measuring the difference between the predicted value and the true value of the model. Its purpose is to optimize the parameters of the model by minimizing the loss value, thereby improving the prediction accuracy of the model. In the present invention, the adaptive loss function combines the forms of squared error and absolute error, and adjusts the weights of the two through the balance parameter. This design not only considers the sensitivity of the squared error to large errors, but also utilizes the robustness of the absolute error to outliers, enabling the model to more comprehensively adapt to different data distribution situations during the training process, thereby improving the generalization ability and prediction accuracy of the model.

[0089] Specifically, in the expression of the adaptive loss function, the number of samples refers to the total number of data points used to train the model, the predicted value is the output result calculated by the model based on the input data, and the true value is the actually observed data value. The balance parameter is a value between 0 and 1, which is used to adjust the weight ratio of the squared error term and the absolute error term in the total loss. When the balance parameter is close to 0, the loss function is closer to the squared error form and is more sensitive to large errors; when the balance parameter is close to 1, the loss function is closer to the absolute error form and is more robust to outliers. By adjusting the balance parameter, the training process of the model can be flexibly optimized according to the specific application scenario and data characteristics to achieve better prediction effects.

[0090] Preferably, the construction and optimization of the adaptive loss function can be achieved through the following steps. First, according to the specific business decision application scenario, determine the initial value of the appropriate balance parameter. For example, in the case of relatively stable data with fewer outliers, the balance parameter can be set to a smaller value (such as 0.1) to pay more attention to the optimization of the squared error; while in the case of large data fluctuations and more outliers, the balance parameter can be set to a larger value (such as 0.5 or 0.7) to enhance the robustness of the model to outliers. Second, during the model training process, the balance parameter can be dynamically adjusted through methods such as cross-validation to find the optimal parameter combination and further improve the performance of the model. Finally, by minimizing the adaptive loss function, the parameters of the model are iteratively updated using optimization algorithms such as gradient descent until the prediction error of the model reaches a satisfactory range or meets the preset convergence conditions. In this way, the adaptive loss function can effectively guide the training process of the model, enabling it to better adapt to the complex business decision environment, thereby improving the accuracy and reliability of the decision.

[0091] The above embodiments of the present invention have the following beneficial effects: The adaptive digital business decision-making reasoning optimization system can improve the accuracy and intelligence level of business decisions. By acquiring historical business data through the acquisition module and combining the intelligent clustering method based on business attributes and data characteristics in the classification module, the rationality of business entity classification can be ensured. The calculation module can accurately capture the dynamic association characteristics between business entities by quantifying the inflow and outflow influence values between entities of the same type and constructing a decision parameter matrix. The reasoning module integrates the dual mechanisms of time series analysis and spatial relationship processing, and can simultaneously learn the temporal evolution law and spatial interaction characteristics of business data, so as to generate more reliable prediction results.

[0092] This system can enhance the adaptability and robustness of the decision-making model. The spatial relationship processing module constructed by the position encoder and the residual network can accurately calculate the association weights between business entities. The time series analysis module adopts an LSTM network combined with a multi-head attention mechanism, which can effectively extract long-term dependence features. The adaptive loss function adopted by the system can balance the prediction accuracy and generalization ability, and the parameter training method based on gradient optimization can ensure the continuous iterative improvement of the model. The finally output business decision can provide data-driven intelligent support for enterprises, helping to optimize business strategies and improve operational efficiency.

[0093] As Figure 2 shown, an adaptive digital business decision-making reasoning optimization method for some embodiments, the method includes:

[0094] Acquire historical business data of each business entity, where the historical business data includes historical inflow data and historical outflow data;

[0095] Classify business entities based on the business attributes of the locations where each business entity is located and the historical business data corresponding to each business entity, and determine the category to which each business entity belongs;

[0096] Based on the category to which each business entity belongs and the historical business data, calculate the decision parameter matrix of each business entity, including: determining other business entities in the category to which the target business entity belongs as the first business entity; based on the historical business data between the target business entity and each first business entity and the historical outflow data of the target business entity, determining the first data, where the first data represents the influence value of the outflow data of each first business entity on the target business entity; based on the historical business data between the target business entity and each first business entity and the historical inflow data of the target business entity, determining the second data, where the second data represents the influence value of the inflow data of each first business entity on the target business entity; based on the first data and the second data, obtain the decision parameter matrix of the target business entity;

[0097] The first commercial data is obtained based on the decision parameter matrix and the historical commercial data of the target commercial entity, and the first commercial data is used to reflect the historical change law of the target commercial entity;

[0098] The constructed decision inference engine is used to process the first commercial data and the spatial relationship features of each commercial entity obtained based on the historical outflow data, and infer the commercial decision result of the target commercial entity in the next decision cycle. The decision inference engine includes a time series analysis module and a spatial relationship processing module. The time series analysis module is used to process the input first commercial data to obtain the second commercial data in the next decision cycle; the spatial relationship processing module includes processing the spatial relationship features of each input commercial entity through a position encoder to obtain the association weight between commercial entities; using a first residual network to process the association weight and the spatial relationship features of each commercial entity to obtain first feature information; processing the second commercial data and the first feature information through a Transformer network to obtain second feature information; obtaining the commercial decision result in the next decision cycle based on the second feature information.

[0099] It can be understood that the steps described in this adaptive digital commercial decision inference optimization method correspond to the various modules in the adaptive digital commercial decision inference optimization system described in the reference Figure 1 Therefore, the modules, features, and beneficial effects described above for the adaptive digital commercial decision inference optimization system also apply to the adaptive digital commercial decision inference optimization method and the operations included therein, and will not be elaborated here.

[0100] In some embodiments, the commercial entities are classified based on the commercial attributes of the locations where the commercial entities are located and the historical commercial data corresponding to the commercial entities to determine the categories to which the commercial entities belong, including: the commercial attributes include retail, service, manufacturing, transportation hub, commercial center, and comprehensive categories;

[0101] Step 1: Randomly select the number of commercial entities consistent with the number of commercial attributes, and use their corresponding historical commercial data as the cluster centers;

[0102] Step 2: Calculate the correlation coefficients between the historical commercial data of the remaining commercial entities and each cluster center, determine the maximum correlation coefficient of the remaining commercial entities and the corresponding cluster center, and all commercial entities form multiple clusters;

[0103] Step 3: Calculate the mean value of the historical commercial data of the commercial entities in the same cluster, and use the mean value as the new cluster center;

[0104] Step 4: Repeat Step 2 and Step 3 until the cluster centers no longer change.

[0105] It should be noted that the adaptive digital business decision-making reasoning optimization method of the present invention is a systematic business decision-making support process. This method understands the operating conditions of a business entity by obtaining the historical business data of the business entity, including inflow and outflow data. Based on these data and the business attributes of the location where the business entity is located, the business entity is classified to determine its category. Business attributes refer to the industry type or functional positioning of the business entity, such as retail, service, etc. Through classification, the behavior of the business entity can be analyzed and predicted more accurately. Further, by calculating the decision parameter matrix, the mutual influence relationship between business entities is quantified, providing data support for business decisions. Finally, using the constructed decision-making reasoning engine, combining time series analysis and spatial relationship processing, the historical data and spatial relationship characteristics of the business entity are comprehensively processed to infer the business decision result of the next decision-making cycle, providing a scientific decision-making basis for business activities.

[0106] Specifically, the historical business data of the business entity is the basic input of this method. The inflow data refers to the data records of various resources received by the business entity (such as funds, goods, customer traffic, etc.), and the outflow data refers to the data records of various resources output by the business entity. Business attributes are an important basis for classifying business entities, including but not limited to retail, service, manufacturing, etc. The classification process is achieved by calculating the correlation coefficient between the historical business data of the business entity and the cluster center point. The cluster center point is the randomly selected historical business data of the business entity, used as the initial reference point for classification. The correlation coefficient is used to measure the similarity between business entities, and the category to which the business entity belongs is determined by calculating the maximum correlation coefficient. The decision parameter matrix is obtained by normalizing the inflow characteristics and outflow characteristics and adding them after non-linear transformation with the corresponding parameters. These parameters are obtained through training based on historical business data and are used to adjust and optimize the calculation result of the decision parameter matrix. The time series analysis module and the spatial relationship processing module are the core parts of the decision-making reasoning engine. The time series analysis module is used to process the time characteristics of business data, and the spatial relationship processing module is used to process the spatial relationship characteristics between business entities.

[0107] Preferably, the implementation of this method can be refined into the following steps. First, when obtaining historical business data, it is necessary to ensure the integrity and accuracy of the data, and noise data and outliers can be removed through data cleaning and preprocessing steps. During the classification process, the selection of cluster centers can be optimized through multiple iterations to improve the accuracy of classification. For example, a convergence threshold can be set, and when the change in the cluster center is less than this threshold, it is considered that the classification process has converged. When calculating the decision parameter matrix, the specific steps of normalization are to divide the inflow data and the outflow data by the difference between their maximum and minimum values respectively, and then scale them into the range of [0,1]. Nonlinear transformation can be achieved through activation functions (such as ReLU or Sigmoid), and these functions can map the input data to a non-linear output space, thereby increasing the expressive power of the model. In the construction of the decision inference engine, the time series analysis module can adopt multiple LSTM networks, and each LSTM network includes components such as a time encoder, a multi-head attention mechanism, a residual network, a feed-forward neural network, a fully connected layer, and a residual network. The spatial relationship processing module can process the spatial relationship features of business entities through a position encoder, calculate the association weights between business entities, and then further process them through a residual network and a Transformer network to obtain the first feature information. Finally, the second business data obtained by the time series analysis module and the first feature information obtained by the spatial relationship processing module are fused through the Transformer network to obtain the second feature information, thereby inferring the business decision result of the next decision cycle.

[0108] In some embodiments, obtaining a decision parameter matrix of a target business entity based on the first data and the second data includes: normalizing the first data based on the maximum and minimum values of the first data to obtain a normalized inflow feature;

[0109] normalizing the second data based on the maximum and minimum values of the second data to obtain a normalized outflow feature;

[0110] performing element-wise non-linear transformation on the normalized inflow feature and the corresponding first parameter and the normalized outflow feature and the corresponding second parameter respectively, and adding the transformation results to obtain a decision parameter matrix, where the first parameter and the second parameter are obtained based on historical business data.

[0111] It should be noted that in the adaptive digital business decision-making reasoning optimization method of the present invention, the process of classifying business entities is based on the business attributes of their locations and the corresponding historical business data. Business attributes refer to the industry types or functional positions of business entities, such as retail, service, manufacturing, etc. These attributes reflect the main functions and roles of business entities in economic activities. Through classification, entities with similar business characteristics can be grouped into one category to more accurately analyze and predict their business behaviors. The classification process includes steps such as randomly selecting cluster center points, calculating correlation coefficients, and updating cluster center points. These steps work together to enable business entities to be reasonably divided into different categories according to their historical data and business attributes, thus providing a basis for subsequent calculation of the decision parameter matrix and business decision-making reasoning.

[0112] Specifically, business attributes are an important basis for classifying business entities, and they include but are not limited to retail, service, manufacturing, transportation hub, commercial center, and comprehensive types, etc. Retail business entities are mainly involved in the sale of goods, service types provide various services, manufacturing types are responsible for product production, transportation hub types are the gathering places for logistics and people flow, commercial center types are usually concentrated areas for various business activities, and comprehensive types include multiple functions. In the classification process, first, randomly select the number of business entities consistent with the number of business attributes, and use their corresponding historical business data as the cluster center points. The cluster center points are the initial reference points in the classification process and are used to calculate the correlation with other business entities. The correlation coefficient is used to measure the similarity between the historical business data of a business entity and the cluster center point. By calculating the correlation coefficient, it can be determined which cluster center point each business entity has the highest similarity with, and thus it is assigned to the corresponding cluster. The update of the cluster center point is completed by calculating the mean value of the historical business data of all business entities in the same cluster. This process will be repeated continuously until the cluster center point no longer changes, that is, it reaches a convergence state. At this time, the classification result is considered stable.

[0113] Preferably, the implementation of the classification process can be refined into the following steps. First, when randomly selecting cluster center points, historical business data of a business entity can be randomly selected from each business attribute category as the initial cluster center points. When calculating the correlation coefficient, the Pearson correlation coefficient or other suitable statistical methods can be used, which can quantify the linear correlation between two data sets. When updating the cluster center points, it is necessary to summarize and calculate the historical business data of all business entities in the same cluster to obtain the new mean as the new cluster center points. This process can be achieved through an iterative algorithm. Each iteration will recalculate the correlation coefficient and update the cluster center points until the change in the cluster center points is less than a preset threshold (e.g., less than 0.01), or the preset maximum number of iterations (e.g., 100 iterations) is reached. In this way, the accuracy and stability of the classification results can be ensured, providing a reliable basis for subsequent business decisions.

[0114] Reference is now made to Figure 3 , which shows a schematic structural diagram of an electronic device 300 suitable for use in implementing some embodiments of the present invention. The electronic devices in some embodiments of the present invention may include, but are not limited to, mobile terminals such as mobile phones, laptop computers, digital broadcast receivers, PDAs (Personal Digital Assistants), PADs (Tablet Computers), PMPs (Portable Multimedia Players), vehicle terminals (e.g., in-vehicle navigation terminals), etc., and fixed terminals such as digital TVs, desktop computers, etc. Figure 3 The terminal devices shown are merely examples and should not impose any limitations on the functions and usage scopes of the embodiments of the present invention.

[0115] As Figure 3 shown, the electronic device 300 may include a processing device (e.g., a central processing unit, a graphics processing unit, etc.) 301, which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 302 or a program loaded from a storage device 308 into a random access memory (RAM) 303. In the RAM 303, various programs and data required for the operation of the electronic device 300 are also stored. The processing device 301, the ROM 302, and the RAM 303 are connected to each other through a bus 304. An input / output (I / O) interface 305 is also connected to the bus 304.

[0116] Generally, the following devices may be connected to the I / O interface 305: an input device 306 including, for example, a touch screen, a touchpad, a keyboard, a mouse, a camera, a microphone, an accelerometer, a gyroscope, etc.; an output device 307 including, for example, a liquid crystal display (LCD), a speaker, a vibrator, etc.; a storage device 308 including, for example, a magnetic tape, a hard disk, etc.; and a communication device 309. The communication device 309 can allow the electronic device 300 to communicate with other devices wirelessly or wireline to exchange data. AlthoughFigure 3 An electronic device 300 with various devices is shown, but it should be understood that it is not required to implement or have all the shown devices. More or fewer devices may be implemented or had alternatively. Figure 3 Each block shown in may represent a device or multiple devices as needed.

[0117] Furthermore, the storage medium of the embodiments of the present application stores program instructions capable of implementing all the above methods. Among them, the program instructions may be stored in the above storage medium in the form of a software product, including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) or a processor to execute all or part of the steps of the methods described in the various embodiments of the present application. And the aforementioned storage medium includes: various media that can store program codes such as USB flash drives, mobile hard disks, read-only memories (ROM, Read-Only Memory), random access memories (RAM, Random Access Memory), magnetic disks, or optical discs, or terminal devices such as computers, servers, mobile phones, and tablets.

[0118] The above description is only some preferred embodiments of the present invention and an explanation of the technical principles applied. Those skilled in the art should understand that the scope of the invention involved in the embodiments of the present invention is not limited to the technical solutions formed by the specific combination of the above technical features, and should also cover other technical solutions formed by any combination of the above technical features or their equivalent features without departing from the above inventive concept. For example, technical solutions formed by mutually replacing the above features with (but not limited to) technical features having similar functions disclosed in the embodiments of the present invention.

Claims

1. An adaptive digital business decision-making reasoning optimization system, characterized in that, Including: An acquisition module that acquires the historical business data of each business entity, where the historical business data includes historical inflow data and historical outflow data; A classification module that classifies business entities based on the business attributes of the locations where the business entities are located and the historical business data corresponding to the business entities, and determines the categories to which the business entities belong; A calculation module that calculates the decision parameter matrix of each business entity based on the categories to which the business entities belong and the historical business data, including: determining other business entities in the category to which the target business entity belongs as the first business entities; determining first data based on the historical business data between the target business entity and each first business entity and the historical outflow data of the target business entity; determining second data based on the historical business data between the target business entity and each first business entity and the historical inflow data of the target business entity; obtaining the decision parameter matrix of the target business entity based on the first data and the second data; obtaining first business data based on the decision parameter matrix and the historical business data of the target business entity; An inference module that processes the first business data and the spatial relationship features of each business entity obtained based on the historical outflow data by using a constructed decision inference engine, and infers the business decision result of the target business entity in the next decision cycle. The decision inference engine includes a time series analysis module and a spatial relationship processing module. The time series analysis module is used to process the input first business data to obtain second business data in the next decision cycle; the spatial relationship processing module includes processing the spatial relationship features of each input business entity through a position encoder to obtain the association weights between business entities; using a first residual network to process the association weights and the spatial relationship features of each business entity to obtain first feature information; processing the second business data and the first feature information through a Transformer network to obtain second feature information; obtaining the business decision result in the next decision cycle based on the second feature information.

2. An adaptive digital business decision-making reasoning optimization system according to claim 1, wherein, The first data represents the influence value of the outflow data of each first business entity on the target business entity; the second data represents the influence value of the inflow data of each first business entity on the target business entity; the first business data is used to reflect the historical change law of the target business entity; Classifying business entities based on the business attributes of the locations where the business entities are located and the historical business data corresponding to the business entities, and determining the categories to which the business entities belong, including: the business attributes include retail, service, manufacturing, transportation hub, commercial center, and comprehensive; Step 1: Randomly select the number of business entities equal to the number of business attributes, and use their corresponding historical business data as the cluster centers; Step 2: Calculate the correlation coefficients between the historical business data of the remaining business entities and each cluster center, determine the maximum correlation coefficient of the remaining business entities and the corresponding cluster center, and all business entities form multiple clusters; Step 3: Calculate the mean value of the historical business data of the business entities in the same cluster, and use the mean value as the new cluster center; Step 4: Repeat Step 2 and Step 3 until the cluster centers no longer change.

3. An adaptive digital business decision-making reasoning optimization system according to claim 2, characterized in that Based on the first data and the second data, a decision parameter matrix of the target business entity is obtained, including: normalizing the first data based on the maximum and minimum values of the first data to obtain a normalized inflow feature; normalizing the second data based on the maximum and minimum values of the second data to obtain a normalized outflow feature; performing element-wise non-linear transformation on the normalized inflow feature and the corresponding first parameter and on the normalized outflow feature and the corresponding second parameter respectively, and adding the transformation results to obtain a decision parameter matrix, where the first parameter and the second parameter are obtained based on historical business data.

4. An adaptive digital business decision-making reasoning optimization system according to claim 1, characterized in that, The spatial relationship processing module includes a plurality of first modules, a linear regression network, and a Softmax network connected in sequence; the first module includes a position encoder, a first residual network, two Transformer networks, a fully connected layer, a feedforward neural network, and a second residual network connected in sequence.

5. An adaptive digital business decision-making reasoning optimization system according to claim 1, characterized in that The time series analysis module includes a plurality of LSTM networks connected in sequence. The LSTM network includes a time encoder, a multi-head attention mechanism, a third residual network, a feedforward neural network, a fully connected layer, and a fourth residual network connected in sequence; the input data is segmented into third business data of multiple continuous time series by the time encoder, the multiple third business data are respectively learned by the multi-head attention mechanism, the learning results are non-linearly processed by the third residual network and the feedforward neural network, and the output result of the feedforward neural network is processed by the fully connected layer to obtain a global feature; the output result of the feedforward neural network and the output result of the fully connected layer are learned by the fourth residual network to obtain the input result of the connected network, and the connected network includes the next LSTM network or the spatial relationship processing module.

6. An adaptive digital business decision-making inference optimization system according to claim 1, characterized in that The spatial relationship features of the input business entities are processed by the position encoder to obtain the association weights between business entities, and the expression is: Among them, p i and p j respectively represent the positions of commercial entity i and commercial entity j in the business network, d is the dimension of the position encoding, α and β are parameters obtained by training based on historical business data, and W ij represents the association weight between commercial entity i and commercial entity j.

7. An adaptive digital business decision-making inference optimization system according to claim 1, characterized in that The decision inference engine is learned using a gradient-based adaptive loss function, and the expression of the loss function is: where N is the number of samples, is the predicted value, and y i is the true value, and λ is the balance parameter.

8. An adaptive digital business decision-making reasoning optimization method, characterized in that, including: Obtain the historical business data of each business entity, where the historical business data includes historical inflow data and historical outflow data; Classify business entities based on the business attributes of the locations where the business entities are located and the historical business data corresponding to the business entities, and determine the category to which each business entity belongs; Based on the category to which each business entity belongs and the historical business data, calculate the decision parameter matrix of each business entity, including: determining other business entities in the category to which the target business entity belongs as the first business entities; based on the historical business data between the target business entity and each first business entity and the historical outflow data of the target business entity, determine the first data, where the first data represents the influence value of the outflow data of each first business entity on the target business entity; based on the historical business data between the target business entity and each first business entity and the historical inflow data of the target business entity, determine the second data, where the second data represents the influence value of the inflow data of each first business entity on the target business entity; based on the first data and the second data, obtain the decision parameter matrix of the target business entity; The first commercial data is obtained based on the decision parameter matrix and the historical commercial data of the target commercial entity, and the first commercial data is used to reflect the historical change law of the target commercial entity; The constructed decision inference engine is used to process the first commercial data and the spatial relationship features of each commercial entity obtained based on the historical outflow data, and infer the commercial decision result of the target commercial entity in the next decision cycle. The decision inference engine includes a time series analysis module and a spatial relationship processing module. The time series analysis module is used to process the input first commercial data to obtain the second commercial data in the next decision cycle; the spatial relationship processing module includes obtaining the association weights between commercial entities by processing the spatial relationship features of each input commercial entity through a position encoder; using a first residual network to process the association weights and the spatial relationship features of each commercial entity to obtain first feature information; obtaining second feature information by processing the second commercial data and the first feature information through a Transformer network; and obtaining the commercial decision result in the next decision cycle based on the second feature information.

9. An adaptive digital business decision-making reasoning optimization method according to claim 8, wherein, Classify commercial entities based on the commercial attributes of the locations where each commercial entity is located and the historical commercial data corresponding to each commercial entity, and determine the category to which each commercial entity belongs, including: the commercial attributes include retail, service, manufacturing, transportation hub, commercial center, and comprehensive categories; Step 1: Randomly select the number of commercial entities consistent with the number of commercial attributes, and use their corresponding historical commercial data as the cluster centers; Step 2: Calculate the correlation coefficients between the historical commercial data of the remaining commercial entities and each cluster center, determine the maximum correlation coefficient of the remaining commercial entities and the corresponding cluster center, and all commercial entities form multiple clusters; Step 3: Calculate the mean value of the historical commercial data of the commercial entities in the same cluster, and use the mean value as the new cluster center; Step 4: Repeat Step 2 and Step 3 until the cluster centers no longer change.

10. An adaptive digital business decision-making reasoning optimization method according to claim 8, characterized in that, Based on the first data and the second data, obtain the decision parameter matrix of the target commercial entity, including: normalizing the first data based on the maximum and minimum values of the first data to obtain the normalized inflow feature; Normalize the second data based on the maximum and minimum values of the second data to obtain the normalized outflow feature; Perform element-wise non-linear transformation on the normalized inflow feature and the corresponding first parameter and the normalized outflow feature and the corresponding second parameter respectively, and add the transformation results to obtain the decision parameter matrix, where the first parameter and the second parameter are obtained based on the historical commercial data.

Images