Dynamic business decision deduction method based on deep learning

Abstract

The invention relates to the technical field of artificial intelligence, and discloses a business decision dynamic deduction method based on deep learning, comprising the following steps: step S1, collecting and preprocessing multi-source dynamic data; s2, dynamic feature construction and multi-modal fusion are carried out; s3, training a decision deduction model; s4, dynamic deduction and real-time optimization are carried out; s5, generating decision suggestions; by setting the step S1, the step S3 and the step S4, the limitation that a traditional method depends on single structured data is broken through by fusing multi-source dynamic data, the corresponding multi-source dynamic data are processed through the LSTM layer, the Transform layer and the CNN layer respectively, the modal weight is adaptively adjusted, the problem of heterogeneity of the multi-modal data is solved, and the multi-modal dynamic data fusion method has the advantages that the multi-modal dynamic data fusion efficiency is improved. And the accuracy of subsequent deduction is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, and more specifically to a method for dynamic extrapolation of business decisions based on deep learning. Background Technology

[0002] In the context of digital transformation, the business environment faced by enterprises is highly dynamic and complex: consumer preferences are rapidly evolving, supply chain nodes are experiencing increased fluctuations, and policies and regulations are frequently adjusted, making it difficult for traditional business decision-making methods to meet the needs.

[0003] Current technologies for dynamic extrapolation of business decisions mostly employ two approaches: static statistical models and rule-driven expert systems. Static statistical models, represented by regression analysis and time series forecasting, rely on linear or simple nonlinear assumptions and require pre-setting the stability of data distribution, but they cannot capture complex nonlinear relationships between multiple variables. Rule-driven expert systems pre-set rules based on the experience of business personnel, which lacks flexibility, ignores the potential patterns hidden in the data, and the rules rely more on human intervention, resulting in a delayed response. Therefore, current technologies only utilize structured historical data, do not fully integrate unstructured real-time data, have a single data dimension, and fixed model parameters, making it impossible to quickly calibrate with market changes, thus exhibiting weak dynamic adaptability.

[0004] In view of this, this application proposes a dynamic extrapolation method for business decision-making based on deep learning. By fusing multi-source dynamic data, training deep learning models, and optimizing in real time, it can achieve dynamic simulation, prediction, and adjustment of the effects of business decisions, thereby improving the scientific nature and adaptability of decision-making. Summary of the Invention

[0005] In order to overcome the above-mentioned deficiencies of the prior art, the present invention provides a deep learning-based dynamic extrapolation method for business decision-making to solve the problems existing in the background art.

[0006] This invention provides the following technical solution: a method for dynamic extrapolation of business decisions based on deep learning, comprising the following steps: Step S1: Multi-source dynamic data acquisition and preprocessing; Multi-source dynamic data is acquired and preprocessed, including structured data, unstructured text data and time-series signal data; Step S2: Dynamic Feature Construction and Multimodal Fusion; The preprocessed multi-source dynamic data is input into the multimodal fusion network to generate dynamic feature vectors. ,in, Indicates the current time step; Step S3: Train the decision-making inference model; input historical decisions and their corresponding historical dynamic feature vectors, and use the historical decision inputs and their corresponding actual decision effects as supervision signals to train the deep neural network model. As a decision-making inference model; Step S4, Dynamic Deduction and Real-time Optimization: Input the candidate decision schemes into the trained decision deduction model, output the deduction results and perform dynamic optimization; the deduction results are the predicted values ​​of the decision effects output by the decision deduction model, that is, the predicted effects of the candidate decision schemes; Step S5: Generate decision recommendations; Based on the deduction results output in step S4, obtain the optimal decision scheme through the natural selection optimization algorithm, and use the optimal decision scheme as the final decision recommendation.

[0007] Preferably, the structured data includes the company's historical sales, inventory levels, and costs; the unstructured text data includes user comments on social media and news, and industry news; the time-series signal data includes the company's store traffic and equipment operating status; and the structured data is represented as a matrix. ,in, Indicates the number of samples. Representing structured dimensions, The unstructured text data is represented as a matrix; the unstructured text data is denoted as a sequence. ,in, Indicates the total number of texts. Indicates the first A text, The time-series signal data is denoted as a matrix. ,in, Indicates the time step. This indicates the number of IoT sensors.

[0008] Preferably, the preprocessing in step S1 specifically includes: The structured data is filled with missing values ​​using time series interpolation, and then... Standardization processing; the unstructured text data uses The model is encoded as a vector, generating context-sensitive semantic vectors. ,in, express The dimension of the model encoding vector. Indicates the first The semantic vectors corresponding to each text are then used to obtain the text feature matrix through average pooling. ,in, The time series signal data is represented by the number of time steps; local features are extracted from the time series signal data using the sliding window method to obtain... The window size of the sliding window is selected. .

[0009] Preferably, the multimodal fusion network mainly includes layer, Layers and Layer; the The layer is used to process structured data, the The layer is used to process unstructured text data, the The layer is used to process the timing signal data from the sensor; The dynamic feature vector is expressed by the formula: ; in, , , These represent the corresponding modal weight coefficients, which are adaptively adjusted through an attention mechanism; Represents the Long Short-Term Memory network; This represents an encoder based on a self-attention mechanism; This represents a one-dimensional convolutional neural network.

[0010] Preferably, the historical decisions are historical business decisions, including promotional efforts and pricing strategies, and the decision effects are the sales growth rate and market share changes brought about by the business decisions; The decision inference model mainly includes an input layer, a hidden layer, a loss function, and an output layer; the input layer is used to input data, and the hidden layer contains two fully connected layers, employing... Activation function; the output layer is used to output the predicted value of the decision effect; The decision-making model uses mean squared error and The weighted sum of divergences, used as the loss function, is expressed by the following formula: ; in, Represents the loss function. Indicates the first A historical decision, This represents the total number of all historical decisions; Indicates the first The predictive effect of historical decisions; Indicates the first The actual effects of historical decisions; Represents the regularization coefficient, selected ; express Divergence; The first output of the model represents the... The confidence distribution of historical decisions, through Normalize the function; Indicates the first The prior distribution of historical decisions.

[0011] Preferably, the trained deep neural network model The training process for the decision inference model is as follows: Transform historical decisions into learnable feature vectors; input historical decisions into the MLP model and output the feature increments corresponding to the historical decisions; By concatenating the vector with the corresponding historical dynamic feature vector, we obtain a joint vector, which can be expressed by the formula: ;in, Represents a joint vector; Represents the feature vector corresponding to historical decisions; The joint vectors are divided into training, testing, and validation sets in a fixed ratio of 7:2:1. Actual decision performance values ​​corresponding to historical decisions are collected, and a one-to-one correspondence is established between the joint vectors and these actual decision performance values. The joint vectors from the training set are then input into the deep neural network model. The process involves obtaining the predicted decision performance values ​​corresponding to the joint vectors, using a loss function to calculate the error between the actual and predicted decision performance values, and continuing until the error no longer converges. At this point, the trained deep neural network model is obtained. .

[0012] Preferably, the candidate decision-making schemes are the business decision-making content selected by the enterprise that needs to be dynamically extrapolated; the dynamic feature vectors The data is input into a trained decision-making model to obtain predicted decision outcomes for candidate decision schemes, thereby enabling forward propagation prediction. The dynamic optimization specifically involves: Set deviation threshold If the deviation between the actual result and the predicted result exceeds a threshold, the online learning mechanism is triggered to update the model parameters through gradient descent; this can be expressed by the formula: ;in, This represents the updated model parameters. This represents the model parameters before the update. Represents the learning rate, selected ; This represents the gradient of the loss function with respect to the model parameters; the updated model parameters are then used as the new model parameters.

[0013] Preferably, obtaining the optimal decision scheme through the natural selection optimization algorithm specifically involves: Step A1: Initialize the population; The candidate decision schemes that meet the screening criteria are used as the initial population. Each individual in the initial population is regarded as a chromosome, that is, each chromosome corresponds to a candidate decision scheme. Step A2: Determine the fitness function; obtain the fitness function for each chromosome; Step A3: Perform natural selection on chromosomes in the population; Step A4: Perform crossover recombination on chromosomes in the population; Step A5: Perform mutations on the chromosomes in the population; Step A6: Obtain a new population, with a preset population generation number of Q and a fitness threshold of IU, where Q is an integer greater than 0 and IU is a real number greater than 0; repeat steps A3 to A5 until the new population reaches generation number Q or a chromosome in the new population has a fitness value greater than or equal to the fitness threshold IU, then the loop ends, and the candidate decision scheme corresponding to the chromosome with the highest fitness in the new population is obtained, which is the optimal decision scheme.

[0014] Preferably, the screening condition is whether the predicted effect of the candidate decision scheme meets the set conditions. If the set conditions are met, the candidate decision scheme meets the screening condition. If the specific content of the decision effect is the sales growth rate, the set condition is that the predicted value of the sales growth rate is greater than the set sales growth rate. If the predicted effect of the candidate decision scheme, i.e., the predicted value of the sales growth rate, is greater than the set sales growth rate, the predicted effect of the candidate decision scheme meets the set conditions. The fitness function is the sales growth rate.

[0015] Preferably, the adaptive adjustment of the modal weight coefficients through the attention mechanism specifically involves: Perform feature projection by using a learnable matrix to project the matrix of each modality data onto the same dimension, and obtain the projected query matrix, key matrix and value matrix; The attention score is calculated using the following formula: ;in, Represents the attention score matrix, To represent the transpose of a matrix, Indicates the scaling factor; Obtain the modal weight coefficients; for use The weighting coefficients for each mode are obtained.

[0016] The technical effects and advantages of this invention are as follows: This invention, through steps S1, S3, and S4, facilitates the fusion of multi-source dynamic data, overcoming the limitations of traditional methods that rely on single structured data. It utilizes LSTM, Transformer, and CNN layers to process the corresponding multi-source dynamic data and adaptively adjusts modality weights, solving the heterogeneity problem of multi-modal data and effectively improving the accuracy of subsequent inferences. Simultaneously, it dynamically and adaptively optimizes the inference results by updating model parameters online through gradient descent, enabling rapid response to market changes and significantly improving response speed compared to static models. Furthermore, it combines KL divergence with mean squared error to obtain the loss function, using KL divergence regularization to constrain the model's confidence distribution, enhancing the model's interpretability. Therefore, by fusing multi-source dynamic data and training a deep neural network model, it achieves dynamic simulation, prediction, and adjustment of business decision-making effects, effectively improving the scientific rigor and adaptability of decision-making. Attached Figure Description

[0017] Figure 1 This is a flowchart of the deep learning-based dynamic deduction method for business decision-making according to the present invention. Detailed Implementation

[0018] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings. In addition, the forms of the various structures described in the following embodiments are merely illustrative. The deep learning-based dynamic deduction method for business decision-making involved in the present invention is not limited to the structures described in the following embodiments. All other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this invention provides a deep learning-based dynamic extrapolation method for business decision-making, comprising the following steps: Step S1: Multi-source dynamic data acquisition and preprocessing; Multi-source dynamic data is acquired and preprocessed, including structured data, unstructured text data and time-series signal data; Step S2: Dynamic Feature Construction and Multimodal Fusion; The preprocessed multi-source dynamic data is input into the multimodal fusion network to generate dynamic feature vectors. ,in, Indicates the current time step; Step S3: Train the decision-making inference model; input historical decisions and their corresponding historical dynamic feature vectors, and use the historical decision inputs and their corresponding actual decision effects as supervision signals to train the deep neural network model. As a decision-making inference model; Step S4, Dynamic Deduction and Real-time Optimization: Input the candidate decision schemes into the trained decision deduction model, output the deduction results and perform dynamic optimization; the deduction results are the predicted values ​​of the decision effects output by the decision deduction model, that is, the predicted effects of the candidate decision schemes; Step S5: Generate decision recommendations; Based on the deduction results output in step S4 and combined with the enterprise's business constraints, obtain the optimal decision scheme through the natural selection optimization algorithm, and use the optimal decision scheme as the final decision recommendation.

[0020] In this embodiment, it should be specifically noted that the structured data includes, but is not limited to, the company's historical sales, inventory quantity, and costs; the unstructured text data includes, but is not limited to, user comments on social media and news, and industry news; the time-series signal data includes, but is not limited to, the company's store customer traffic and equipment operating status; the structured data can be obtained through the company's internal system, and the time-series signal data can be obtained by deploying IoT sensors; the structured data is represented as a matrix. ,in, Indicates the number of samples. Representing structured dimensions, The unstructured text data is represented as a matrix; the unstructured text data is denoted as a sequence. ,in, Indicates the total number of texts. Indicates the first A text, The time-series signal data is denoted as a matrix. ,in, Indicates the time step. Indicates the number of IoT sensors; The preprocessing in step S1 specifically includes: The structured data is filled with missing values ​​using time series interpolation, and then... Standardization processing is used to eliminate the influence of units; the unstructured text data uses... The model is encoded as a vector, generating context-sensitive semantic vectors. ,in, express The dimension of the model encoding vector. Indicates the first The semantic vectors corresponding to each text are then used to obtain the text feature matrix through average pooling. ,in, The time series signal data is represented by the number of time steps; local features are extracted from the time series signal data using the sliding window method to obtain... This is used to capture periodic fluctuations; the window size of the sliding window is selected. .

[0021] In this embodiment, it should be specifically noted that the multimodal fusion network mainly includes layer, Layers and Layer; the The layer is used to process structured data and capture time dependencies. The layer is used to process unstructured text data and identify semantic relationships. The layer is used to process the timing signal data from the sensor; The dynamic feature vector is expressed by the formula: ; in, , , These represent the corresponding modality weight coefficients, which can be adaptively adjusted through an attention mechanism to reflect the importance of different data modalities to the current decision. This represents a Long Short-Term Memory (LSTM) network, used to capture the temporal dependencies of structured data; This represents a self-attention-based encoder used to extract semantic associations from unstructured text data. This represents a one-dimensional convolutional neural network used to mine local patterns in time-series signals; The , , initial value selection After adaptive adjustment through attention mechanism , , The value is: , , The , , satisfy ; Its purpose is to map structured data, unstructured text data, and time-series signal data to a unified feature space through a multimodal fusion network, and to transform heterogeneous data into a unified dynamic feature representation through the multimodal fusion network, thereby enabling the simultaneous processing of structured data and unstructured text data.

[0022] In this embodiment, it should be specifically noted that the historical decisions are historical business decisions, including but not limited to promotional efforts and pricing strategies, and the decision effects are the sales growth rate and market share changes brought about by the business decisions. The decision inference model mainly includes an input layer, a hidden layer, a loss function, and an output layer; the input layer is used to input data, and the hidden layer contains two fully connected layers, employing... Activation function; the output layer is used to output the predicted value of the decision effect; The decision-making model uses mean squared error and The weighted sum of divergences, used as the loss function, is expressed by the following formula: ; in, Represents the loss function. Indicates the first A historical decision, This represents the total number of all historical decisions; Indicates the first The predictive effect of historical decisions; Indicates the first The actual effects of historical decisions; Represents the regularization coefficient, selected This is used to balance the accuracy and interpretability of predictions; express Divergence is used to constrain the confidence distribution of the model output, avoiding overconfidence. The first output of the model represents the... The confidence distribution of historical decisions, through The function is normalized to reflect the model's confidence level in the predicted values; Indicates the first The prior distribution of historical decisions; the mean square error is used to constrain the error between the predicted value and the true value, thereby improving accuracy.

[0023] In this embodiment, it should be specifically noted that the training deep neural network model The training process for the decision inference model is as follows: Transform historical decisions into learnable feature vectors; input historical decisions into the MLP model and output the feature increments corresponding to the historical decisions; for example, if the input historical decision is the rate of change of promotion intensity, then output the feature increment of promotion intensity as the feature vector. By concatenating the vector with the corresponding historical dynamic feature vector, we obtain a joint vector, which can be expressed by the formula: ;in, Represents a joint vector; Represents the feature vector corresponding to historical decisions; The joint vectors are divided into training, testing, and validation sets in a fixed ratio; in this embodiment, a fixed ratio of 7:2:1 is selected. Actual decision performance values ​​corresponding to historical decisions are collected, and the joint vectors are mapped one-to-one with these actual decision performance values. The joint vectors from the training set are then input into the deep neural network model. The process involves obtaining the predicted decision performance values ​​corresponding to the joint vectors, using a loss function to calculate the error between the actual and predicted decision performance values, and continuing until the error no longer converges. At this point, the trained deep neural network model is obtained. The validation set is used to evaluate the model's performance during training, thereby adjusting the model's hyperparameters, and the test set is used to finally evaluate the model's generalization ability.

[0024] In this embodiment, it should be specifically noted that the candidate decision scheme is the business decision content selected by the enterprise that needs to be dynamically deduced; the dynamic feature vector The data is input into a trained decision-making model to obtain predicted decision outcomes for candidate decision schemes, thereby enabling forward propagation prediction. The dynamic optimization specifically involves: Set deviation threshold If the deviation between the actual result and the predicted result exceeds a threshold, the online learning mechanism is triggered to update the model parameters through gradient descent; this can be expressed by the formula: ;in, This represents the updated model parameters. This represents the model parameters before the update. Represents the learning rate, selected ; This represents the gradient of the loss function with respect to the model parameters; the updated model parameters are then used as the new model parameters. The deviation threshold can be set by those skilled in the art, and this embodiment does not specifically limit this value. The online learning mechanism also includes collecting recent... Real-time multi-source dynamic data at each time step is used to ensure sufficient data volume for updating model parameters. In this embodiment, the following is selected: After updating the model parameters, verify the loss of the updated model on the validation set. If the loss increases, roll back the parameters.

[0025] In this embodiment, it should be specifically noted that the enterprise business constraints include cost constraints, inventory constraints, brand constraints, and legal constraints; the brand constraint means that the decision-making scheme cannot contain prohibited words, and the legal constraint means that the decision-making scheme cannot violate the content of laws and regulations; those skilled in the art can select any one or more of the enterprise business constraints according to the actual situation, or they can choose not to set enterprise business constraints. The specific steps for obtaining the optimal decision scheme through the natural selection optimization algorithm are as follows: Step A1: Initialize the population; The candidate decision schemes that meet the screening criteria are used as the initial population. Each individual in the initial population is regarded as a chromosome, that is, each chromosome corresponds to a candidate decision scheme. The screening criteria are whether the prediction effect of the candidate decision scheme meets the set conditions. If the set conditions are met, the candidate decision scheme meets the screening criteria. The set conditions are specifically set by those skilled in the art based on the specific content of the decision effect. For example, if the specific content of the decision effect is the sales growth rate, the set condition can be that the predicted value of the sales growth rate is greater than the set sales growth rate. Therefore, if the prediction effect of the candidate decision scheme, that is, the predicted value of the sales growth rate, is greater than the set sales growth rate, the prediction effect of the candidate decision scheme meets the set conditions. Step A2: Determine the fitness function; obtain the fitness function for each chromosome; the fitness function is set according to the specific content of the decision effect. For example, if the specific content of the decision effect is the sales growth rate, then the fitness function is selected as the sales growth rate; in this embodiment, the fitness function is selected as the sales growth rate. Step A3: Perform natural selection on chromosomes in the population; Step A4: Perform crossover recombination on chromosomes in the population; Step A5: Perform mutations on the chromosomes in the population; Step A6: Obtain a new population, with a preset generation number of Q and a fitness threshold of IU, where Q is an integer greater than 0 and IU is a real number greater than 0; repeat steps A3 to A5 until the generation number of the new population is Q or a chromosome in the new population has a fitness greater than or equal to the fitness threshold IU, then the loop ends, and the candidate decision scheme corresponding to the chromosome with the highest fitness in the new population is obtained, which is the optimal decision scheme; for example, if the preset generation number of the population is 1, then the chromosomes in the initial population are subjected to natural selection, crossover recombination, and mutation to obtain a new population, at which point the generation number of the new population is 1, and therefore the loop ends; The natural selection can be carried out using a combination of the elite method and the rotation method, and the crossover and recombination can be carried out using the PMX method; the elite method, the rotation method and the PMX method are all existing technical means, and will not be described in detail in this embodiment.

[0026] In this embodiment, it should be specifically explained that the adaptive adjustment of the modal weight coefficients through the attention mechanism specifically refers to: Feature projection is performed, and the learnable matrix is ​​used to project the matrices of each modality data to the same dimension to obtain the projected query matrix, key matrix, and value matrix; the features of different modalities are transformed to the same dimensional space through linear projection to achieve dimensional unification of cross-modal features, laying the foundation for obtaining modality weight coefficients in the future. Attention scores are obtained to measure the correlation between modalities; expressed by the formula: ;in, This represents the attention score matrix, used to measure the correlation strength between features of different modalities. To represent the transpose of a matrix, This represents the scaling factor; its purpose is to quantize the correlation between modes through the dot product and scaling, while avoiding excessively large dot product results due to excessively large unified dimensions, which could lead to problems in subsequent operations. The problem of vanishing gradients; Obtain the modal weight coefficients; for use Thus, the weighting coefficients for each mode are obtained, expressed by the formula: ; ; ; in, express Functions used to convert The original scores are converted into a probability distribution form to ensure that the sum of the weights is 1; The first value of the attention score matrix represents the... Row vector elements, representing the first row element. The correlation score between a mode and other modes; .

[0027] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

[0028] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A deep learning-based dynamic extrapolation method for business decision-making, characterized in that: The method comprises the following steps: Step S1, multi-source dynamic data acquisition and preprocessing; Step S1, multi-source dynamic data acquisition and preprocessing; Step S2, dynamic feature construction and multi-modal fusion; input the preprocessed multi-source dynamic data into the multi-modal fusion network to generate a dynamic feature vector wherein, denotes the current time step; Step S3, training the decision deduction model; inputting the historical decisions and the corresponding historical dynamic feature vectors, taking the historical decision input and the corresponding decision effect actual value as the supervision signal, and training the deep neural network model as a decision deduction model; The multi-source dynamic data comprises structured data, unstructured text data and time series signal data; Step S4, dynamic deduction and real-time optimization; the candidate decision scheme is input into the trained decision deduction model, and a deduction result is output and dynamically optimized; the deduction result is a decision effect prediction value output by the decision deduction model, that is, a predicted effect of the candidate decision scheme; 2. The deep learning-based dynamic business decision-making reasoning method according to claim 1, characterized in that: The structured data includes historical sales, inventory quantity and cost of the enterprise; the unstructured text data includes user reviews of social media and news information, industry dynamic messages; the time series signal data includes enterprise store passenger flow and equipment running state; the structured data is recorded as a matrix , wherein represents the number of samples, represents the structured dimension, is a matrix representation; the unstructured text data is recorded as a sequence , wherein represents the total number of texts, represents the th text, ; the time series signal data is recorded as a matrix , wherein represents the time step, represents the number of Internet of Things sensors.

3. The deep learning-based dynamic business decision-making reasoning method of claim 2, wherein: Step S5, generating a decision suggestion; based on the deduction result output in step S4, an optimal decision scheme is obtained through a natural selection optimization algorithm, and the optimal decision scheme is taken as a final decision suggestion. The structured data is filled with missing values by using time series interpolation method, and then is processed by standardization; the unstructured text data is encoded into vectors by using model to generate context-related semantic vectors , wherein, represents the dimension of the model encoding vector, represents the semantic vector corresponding to the first text, and the text feature matrix is obtained by average pooling , wherein, represents the time step number; the time series signal data is processed by sliding window method to extract local features and obtain ; the window size of the sliding window is selected as .

4. The deep learning-based business decision dynamic deduction method according to claim 3, characterized in that: The multimodal fusion network mainly includes layer, Layers and Layer; the The layer is used to process structured data, the The layer is used to process unstructured text data, the The layer is used to process the timing signal data from the sensor; The preprocessing in step S1 is specifically: ； wherein, , , respectively represent corresponding modal weight coefficients, which are adaptively adjusted through an attention mechanism; represents a long short-term memory network; represents an encoder based on a self-attention mechanism; represents a one-dimensional convolutional neural network.

5. The deep learning-based dynamic business decision-making reasoning method according to claim 4, characterized in that: The dynamic feature vector is represented by a formula: The decision deduction model mainly comprises an input layer, a hidden layer, a loss function and an output layer; the input layer is used for inputting data, the hidden layer comprises two full connection layers, adopts an activation function; and the output layer is used for outputting a decision effect prediction value. The decision inference model adopts a weighted sum of mean square error and divergence as a loss function, which is expressed by a formula as follows: ； in, Represents the loss function. Indicates the first A historical decision, This represents the total number of all historical decisions; Indicates the first The predictive effect of historical decisions; Indicates the first The actual effects of historical decisions; Represents the regularization coefficient, selected ; express Divergence; The first output of the model represents the... The confidence distribution of historical decisions, through Normalize the function; Indicates the first The prior distribution of historical decisions.

6. The deep learning-based dynamic business decision-making reasoning method according to claim 5, characterized in that: The training deep neural network model The training process as a decision inference model is specifically: The historical decision is a historical business decision content, the business decision comprises a promotion intensity and a pricing strategy, and the decision effect is a sales growth rate and a market share change brought by the business decision; The historical decision is converted into a learnable feature vector; the historical decision is input into an MLP model, and a feature increment corresponding to the historical decision is output; ; wherein, represents a joint vector; represents a feature vector corresponding to a historical decision; The joint vector is divided into a training set, a test set and a validation set in a fixed ratio, and a fixed ratio of 7:2:1 is selected; the actual value of the decision effect corresponding to the historical decision is collected, the joint vector is one-to-one corresponding to the actual value of the decision effect, the joint vector in the training set is input into the deep neural network model , the decision effect prediction value corresponding to the joint vector is obtained, the error between the actual value of the decision effect and the decision effect prediction value is obtained using a loss function, and the process is stopped until the error no longer converges, and the deep neural network model at the end of training is obtained .

7. The deep learning-based dynamic business decision-making reasoning method according to claim 6, characterized in that: The candidate decision scheme is a business decision content selected by the enterprise and requiring dynamic reasoning; the dynamic feature vector The candidate decision scheme is a business decision content selected by the enterprise and requiring dynamic reasoning; the dynamic feature vector The dynamic optimization is specifically: Setting a deviation threshold If the deviation between the actual result and the predicted result exceeds the threshold, an online learning mechanism is triggered to update the model parameters through gradient descent. ; wherein, denotes the updated model parameters, denotes the model parameters before the update, denotes the learning rate, which is chosen ; denotes the gradient of the loss function with respect to the model parameters; the updated model parameters are used as the new model parameters.

8. The deep learning-based dynamic business decision-making reasoning method according to claim 7, characterized in that: The joint vector is obtained by splicing the corresponding historical dynamic feature vector, and is represented by a formula: Step A1, initialize population; take a candidate decision scheme satisfying the screening condition as the initial population, and take each individual in the initial population as a chromosome, i.e. each chromosome corresponds to a candidate decision scheme; The optimal decision scheme is obtained through a natural selection optimization algorithm, and is specifically: Step A2, determining a fitness function; a fitness function of each chromosome is obtained; Step A3: natural selection is performed on the chromosomes in the population; Step A4: the chromosomes in the population are subjected to crossover and recombination; Step A5: the chromosomes in the population are subjected to mutation; 9. The deep learning-based dynamic business decision-making reasoning method according to claim 8, characterized in that: Step A6: a new population is obtained, a preset population number is Q, a fitness threshold is IU, Q is an integer greater than 0, and IU is a real number greater than 0; steps A3 to A5 are cycled until the number of the new population is Q or there is a chromosome in the new population corresponding to a fitness greater than or equal to the fitness threshold IU, the cycle is ended, and a candidate decision scheme corresponding to the chromosome corresponding to the maximum fitness in the new population is obtained, that is, the optimal decision scheme.

10. The deep learning-based dynamic business decision-making reasoning method of claim 9, wherein: The screening condition is whether the predicted effect of the candidate decision scheme meets a set condition, if the set condition is met, the candidate decision scheme meets the screening condition, if the specific content of the decision effect is a sales growth rate, the set condition is that the prediction value of the sales growth rate is greater than a set sales growth rate, if the predicted effect of the candidate decision scheme, that is, the prediction value of the sales growth rate, is greater than the set sales growth rate, the predicted effect of the candidate decision scheme meets the set condition; the fitness function is the sales growth rate. The adaptive adjustment of the modality weight coefficient through the attention mechanism is specifically: Feature projection is performed, the matrices of the various modal data are projected to the same dimension by using a learnable matrix, and a projected query matrix, a key matrix and a value matrix are obtained; An attention score is obtained, and is represented by a formula: ; wherein, denotes the attention score matrix, denotes the transpose of a matrix, denotes a scaling factor; Obtaining the modal weight coefficient; to Using Obtaining the weight coefficient of each modal.

Images