AI commercial large model training method based on small, medium and micro commercial global operation

Abstract

The invention relates to the field of business data processing, and discloses an AI business large model training method based on small, medium and micro business global operation, comprising the following steps: collecting operation data of small, medium and micro enterprises; preprocessing the operation data; adopting a deep neural network model to obtain feature representation of the operation data; fusing the feature representations through a weighted fusion mechanism to obtain a comprehensive feature vector; optimizing the comprehensive feature vector by using a target loss function; further optimizing a fusion feature and a training process based on the optimized comprehensive feature vector and a loss function result; and performing incremental learning and online training on the optimization model. By analyzing and processing operation data of small, medium and micro enterprises and combining a deep neural network model and a weighted fusion mechanism, accurate decision support is provided for the enterprises, identification of operation bottlenecks is facilitated, so that the operation cost is reduced, the efficiency is improved, and real-time adjustment and online training enable the enterprises to obtain an efficient and flexible operation management scheme.

Description

Technical Field

[0001] The present invention relates to the field of commercial data processing, and specifically to an AI commercial large model training method based on the overall operation of small, medium and micro businesses. Background Art

[0002] With the rapid economic development and the acceleration of the digitalization process, small, medium and micro enterprises play an increasingly important role in the national economy. However, these enterprises face many challenges in the operation process, such as fierce market competition, changing consumer demands, high operation costs, and difficulties in digital transformation. At the same time, the rapid development of artificial intelligence technology provides new ideas and methods to solve these problems.

[0003] Traditional commercial operation models in commercial data processing are difficult to meet the needs of small, medium and micro enterprises for precision marketing, efficient operation and personalized services. When facing the complex business scenarios and diverse demands in the small, medium and micro commercial fields, general large models often cannot provide targeted solutions. Small, medium and micro commerce has unique business characteristics, such as a wide variety of categories, scattered customer groups, diverse marketing channels, etc., and an AI large model that can deeply understand these characteristics and provide customized services is needed.

[0004] At present, although there are some AI large models for specific industries, the large models specifically designed for the overall operation of small, medium and micro businesses are relatively few. Existing models still have deficiencies in data processing, algorithm optimization and application scenario adaptation, and cannot fully exert the potential of AI technology in the small, medium and micro commercial fields. Therefore, there is an urgent need for a more efficient and flexible solution that can provide accurate decision-making support for enterprises, identify operation bottlenecks, reduce waste of redundant resources, and improve operation efficiency. Summary of the Invention

[0005] In view of the deficiencies of the prior art, the present invention provides an AI commercial large model training method based on the overall operation of small, medium and micro businesses, which solves the problems that existing models cannot fully adapt to the complex business scenarios of small, medium and micro enterprises, have insufficient data processing capabilities, insufficient algorithm optimization, and cannot provide customized solutions.

[0006] To achieve the above objectives, the present invention is realized through the following technical solutions: An AI commercial large model training method based on the overall operation of small, medium and micro businesses, comprising the following steps: Collect the operation data of small, medium and micro enterprises, including sales data, customer data, inventory data and marketing data; Preprocess the operation data; Use a deep neural network model to extract features from the operation data to obtain the feature representation of the operation data; Fuse the feature representations through a weighted fusion mechanism to obtain a comprehensive feature vector; Optimize the comprehensive feature vector using the target loss function; Based on the optimized comprehensive feature vector and the loss function results, use reinforcement learning to dynamically adjust the hyperparameters of the deep neural network model to further optimize the fusion features and the training process; Perform incremental learning and online training on the optimized model to adapt to the changes and updates of data during the operation of small, medium and micro enterprises.

[0007] Preferably, the preprocessing includes duplicate removal, missing value filling, outlier handling, standardization and normalization.

[0008] Preferably, the deep neural network model includes a Transformer model, a convolutional neural network model and an LSTM model.

[0009] Preferably, the weighted fusion mechanism generates a comprehensive feature vector according to the weighting coefficients and by means of weighted summation, and the weighting coefficients are dynamically adjusted by an optimization algorithm.

[0010] Preferably, the optimization algorithm includes the gradient descent method, and updates the weighting coefficients through backpropagation to minimize the loss between the fusion features and the target business metrics.

[0011] Preferably, the reinforcement learning includes a learning rate, a regularization coefficient and a batch size, and the reinforcement learning evaluates the performance of the model on the validation set through a reward function.

[0012] Preferably, the reward function gives a reward value by comprehensively considering the decrease of the loss function and the improvement of the business metrics, and updates the hyperparameters of the model according to the current state and the reward value.

[0013] Preferably, the loss function includes an error metric between the fusion features and the actual business objectives, and the error metric includes mean square error and cross-entropy error.

[0014] Preferably, the incremental learning and online training update the existing model every time new data arrives.

[0015] Preferably, the weighted fusion mechanism performs weighted summation on the feature representations, multiplies the feature values of the feature representations by their corresponding weighting coefficients, and the weighting coefficients are dynamically adjusted by an optimization algorithm during the training process.

[0016] The present invention provides a training method for an AI business large model based on the overall operation of small, medium and micro businesses. It has the following beneficial effects: 1. By analyzing and optimizing the operation data of small, medium and micro enterprises, and combining with a deep neural network model and a weighted fusion mechanism, the present invention can provide accurate decision-making support for enterprises, thereby helping enterprises identify operation bottlenecks, reduce waste of redundant resources, significantly reduce operation costs and improve operation efficiency. Moreover, through real-time adjustment and online training, an efficient and flexible operation management solution can be provided for enterprises.

[0017] 2. By using a deep neural network model and reinforcement learning to analyze operation data, the present invention can deeply explore the personalized needs and potential preferences of consumers, thereby providing accurate marketing solutions for small, medium and micro enterprises. Through feature extraction and weighted fusion of operation data, a high-quality consumer portrait can be generated, providing personalized recommendation content and customized marketing strategies for enterprises, not only improving the utilization efficiency of marketing resources, but also effectively enhancing customer stickiness and loyalty, and enhancing the market competitiveness of enterprises.

[0018] 3. By using this model to provide operation strategies for small, medium and micro enterprises, analyzing operation data in real time, optimizing processes and making accurate decisions for enterprises, the intelligent level of business operations can be improved, and the purpose of helping enterprises with digital transformation can be achieved. This not only enhances the competitiveness of enterprises, but also improves their innovation ability, enabling enterprises to better adapt to market changes and industry development trends, and promoting enterprises to enter the digital economy era. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a schematic flow chart of the method of the present invention; DETAILED DESCRIPTION OF THE EMBODIMENTS

[0020] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described in conjunction with the drawings in the specification of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0021] For a better understanding of the present invention, the above content will be described in detail below in conjunction with specific embodiments.

[0022] Please refer to the attached Figure 1 , the embodiment of the present invention provides an AI business large model training method based on the overall operation of small, medium and micro businesses, including the following steps: Collect the operation data of small, medium and micro enterprises, including sales data, customer data, inventory data and marketing data; In this embodiment, the operation data of small and medium-sized enterprises is collected first. These data cover multiple key areas, including sales data, customer data, inventory data, and marketing data. Through the comprehensive collection of these data, a solid foundation can be provided for subsequent model training and business decision-making.

[0023] Specifically, the operation data of small and medium-sized enterprises is obtained through multiple data sources. Specifically, the sales data includes information such as commodity sales records, sales volume, sales amount, and price fluctuations, which can reflect the market performance and profitability of the enterprise; the customer data covers the basic information, purchase behavior, loyalty, etc. of customers, and through these data, the behavior patterns and demand characteristics of customers can be accurately depicted; the inventory data includes the quantity of commodity inventory, inventory turnover rate, supply chain status, etc., which helps the enterprise optimize inventory management and reduce inventory costs; the marketing data covers information such as the investment, return, and effect evaluation of the enterprise in various marketing activities, and is used to evaluate the effectiveness of marketing activities and make corresponding optimization adjustments.

[0024] As an option, in practical applications, the data collection method can be divided into two types. One is regular collection, that is, exporting data from each business system according to a fixed time period; the other is real-time collection, using sensors, Internet platforms, mobile devices, etc. to dynamically collect data to ensure the timeliness and comprehensiveness of the data. To ensure the quality and accuracy of the data, preprocessing can also be performed on the collection process to ensure that the data has been de-duplicated, missing values have been filled, and anomalies have been eliminated before entering the deep learning model.

[0025] In a possible implementation, data is automatically collected by connecting various enterprise management systems, third-party platforms, and API interfaces and combining the existing technical architecture. For example, sales data can be obtained in real time by connecting with e-commerce platforms or POS systems, customer data can be collected through multiple channels such as CRM systems, social media platforms, and customer feedback, inventory data can be docked with warehouse management systems, and marketing data can be directly obtained through the data interfaces of advertising platforms and analysis tools. This method can effectively reduce the data latency of manual collection and improve the efficiency of data collection.

[0026] To ensure the comprehensiveness and high quality of the data, all data should comply with the enterprise's privacy policy and data security standards during the collection process. For example, when collecting customer data, relevant privacy protection regulations such as GDPR should be followed to ensure the effective protection of users' personal information. To improve the accuracy and practicality of the data, external data sources can also be combined for supplementation, such as industry reports, market research, etc.

[0027] Therefore, the collected operation data will provide rich feature information for the subsequent training of the deep neural network model, further improving the training effect and decision-making support ability of the AI commercial large model. In the subsequent steps, these collected data will undergo feature extraction and weighted fusion, ultimately providing comprehensive and accurate operation support for the enterprise.

[0028] Preprocess the operation data; In this embodiment, data preprocessing is a crucial step to ensure the efficient and accurate training of the subsequent deep neural network model. By preprocessing the operation data, redundant information can be removed, missing data can be filled, outliers can be corrected, and data quality can be guaranteed, thereby improving the accuracy and robustness of model training. This step provides a clean and standardized basis for the input data of the deep learning model.

[0029] Specifically, the preprocessing of operation data includes multiple sub-steps, mainly involving operations such as deduplication, filling missing values, outlier handling, standardization, and normalization. First, the deduplication operation ensures that each piece of data in the dataset is unique, avoiding interference caused by duplicate data. Especially when processing sales data and customer data, deduplication can prevent the same transaction record or customer information from being double-counted. For filling missing values, different strategies can be adopted according to the importance and missing situation of the data, such as mean filling, median filling, or interpolation method, to ensure the rationality and continuity of the filled data.

[0030] Generally, during the outlier handling process, outliers in the data are first detected and identified through statistical analysis (such as the standard deviation method or the IQR method). If some data points deviate far from the normal range, they will be corrected or removed to avoid these abnormal data affecting the accuracy of model training. For example, if a data value in the inventory data is extremely abnormal, it may be an input error or other unforeseen problems, and at this time, appropriate correction or deletion is required.

[0031] As an option, in the data standardization and normalization steps, standardization is usually used to ensure that different features are compared under the same dimension, while normalization adjusts the data to a unified range (such as [0, Business Data Processing 1]). These two steps help prevent certain features from having an adverse impact on model training due to overly large or small numerical values during the training process. The standardization process usually adopts the Z-score standardization method, that is, for each feature in the dataset, its mean and standard deviation are calculated and transformed through the formula:

[0032] where, is the original data, is the mean of this feature, is the standard deviation of this feature, It is the data after standardization.

[0033] Specifically, for the normalization operation, the common method is to map the data into the interval [0,1], which is achieved through the following formula:

[0034] where is the original data, and are the minimum and maximum values in this feature respectively, is the data after normalization.

[0035] In a possible implementation, the preprocessing process may combine multiple data sources. For example, sales data and inventory data may come from different systems or databases and are uniformly processed through ETL (Extract, Transform, Load) technology to ensure the consistency and compatibility of data formats. In addition, the data processing process can be carried out in the cloud or on a local server to efficiently process large-scale data and improve the processing speed and response ability.

[0036] The preprocessed data will go through further feature extraction steps to provide rich and high-quality input data for the deep neural network model. Each step in this process can effectively ensure the accuracy, integrity, and consistency of the data, providing a solid data foundation for subsequent AI model training.

[0037] Therefore, by strictly preprocessing the data, noise can be eliminated and the data quality can be improved, providing efficient and accurate input for the deep learning model. This process plays a crucial role in the present invention, ensuring the accuracy and reliability of subsequent model training and decision support.

[0038] Use a deep neural network model to extract features from the operation data and obtain the feature representation of the operation data; In this embodiment, a deep neural network model is used to extract features from the operation data, aiming to extract effective feature representations. Through feature extraction, complex original data can be transformed into structured information suitable for machine learning and model training. This process provides the necessary input data for subsequent weighted fusion, optimization, and decision support.

[0039] Specifically, the deep neural network model is used to automatically learn and extract features from operation data. By training the model, the rules, patterns, and potential correlations hidden behind the data can be identified. Generally, the deep neural network model gradually abstracts and extracts high-order features in the data through a multi-layer structure. First, the input layer receives the preprocessed data, including sales data, customer data, inventory data, and marketing data, etc. Next, the data is passed through multiple layers of neural networks, and more abstract features are gradually extracted at each layer.

[0040] As an option, different types of deep neural network models can be used during the feature extraction process. For example, a Transformer-based model can be used for text data, a convolutional neural network (CNN) for image data, and a long short-term memory network (LSTM) for time series data. These deep neural network models extract features of the data from different perspectives through different structures and training methods. For example, the Transformer model performs well in processing long texts and data with complex relationships, while the CNN is good at extracting spatial features from image data, and the LSTM is suitable for capturing the temporal dependencies in time series data.

[0041] For each type of data, the model is processed through different network layers. For example, for sales data, convolutional layers may be needed to capture local patterns in the data, while for customer behavior data, the long short-term memory network can capture the historical behaviors and preference changes of customers. The output of each network layer is the feature representation abstracted by that layer.

[0042] In a possible implementation, by using activation functions (such as ReLU, Sigmoid, or Tanh, etc.) in the neural network, the features extracted at each layer can be further non-linearly transformed to enhance the expressive power of the model. Through the backpropagation algorithm, the model continuously adjusts the weights according to the loss function of the target task, thereby optimizing the effect of feature extraction and ensuring that the network can more accurately capture the potential rules in the data.

[0043] In some embodiments, after the feature vectors are extracted, subsequent weighted fusion and optimization steps will be performed on these features to further enhance the prediction ability of the model. By weighted fusing multiple feature sources and combining different data sources and modality information, the operational status of the enterprise can be more comprehensively understood, thereby providing higher-quality support for decision-making.

[0044] During the feature extraction process, the deep neural network model will be adjusted and optimized according to different data types and task requirements. For example, it may be necessary to adjust the hyperparameters of the model (such as learning rate, number of network layers, number of nodes, etc.) to further improve the accuracy and adaptability of feature extraction.

[0045] Therefore, through these feature extraction steps of the deep neural network model, the original operation data can be transformed into feature vectors with high expressive ability, which are used as the input for subsequent optimization, decision-making, and prediction models. Therefore, it can effectively capture the key features in the operation data and provide guiding analysis results for enterprises to help them make accurate decisions in the complex business environment.

[0046] The feature representations are fused through a weighted fusion mechanism to obtain a comprehensive feature vector; In this embodiment, the weighted fusion mechanism of the feature representations is an important step in fusing multi-modal features extracted from the deep neural network model. Through the weighted fusion mechanism, the weights of each modal feature can be dynamically adjusted according to the importance and contribution degree of different features, and finally a comprehensive feature vector is generated. This feature vector can more comprehensively represent the operation status of small, medium, and micro enterprises. This comprehensive feature vector will be used as the core input for subsequent optimization and decision support.

[0047] Specifically, after feature extraction, multiple feature representations from different operation data (such as sales data, customer data, inventory data, and marketing data, etc.) are obtained. To effectively fuse these features, it is necessary to perform weighted summation on them through a weighted fusion mechanism to obtain a comprehensive feature vector that integrates information from each modality. Generally, the weighted fusion process of the feature representations is achieved by calculating the weight coefficients of each modal feature. These weight coefficients reflect the contribution degree of each modal data to the final prediction result. The calculation of the weights usually depends on the optimization algorithm during the training process.

[0048] As an option, a neural network-based weighted mechanism can be adopted to learn the weighted coefficients of each modal feature. Specifically, assume there are modal feature representations, and the feature vector of each modality is , then the result of the weighted fusion can be expressed as:

[0049] where, is the feature vector of the th modality, is the weighted coefficient of the th modal feature, and , that is, the sum of all weights is 1. Through the backpropagation algorithm, the model will dynamically update these weight coefficients, enabling it to continuously optimize during the training process. Eventually, the feature representation after weighted fusion can better reflect the impact of each modality data on the final decision.

[0050] Specifically, the core of the weighted fusion mechanism lies in how to design the learning process of the weight coefficients. In some embodiments, these weight coefficients can be optimized by minimizing a loss function. The loss function can include the prediction error of the model and the difference between the weighted feature representation and the actual business objective. To ensure the rationality and effectiveness of the fusion result, regularization methods are usually applied during the training process to avoid overfitting and ensure the stability of the fusion effect of each feature.

[0051] In a possible implementation, to further improve the fusion effect, a weighted fusion method based on the attention mechanism can also be adopted. In this method, the network will automatically calculate the "importance" scores of each modality feature and adjust the weights accordingly. Specifically, the model based on the attention mechanism will dynamically adjust the weights according to the context information of the input features, enabling important features to obtain higher weights and thus making a greater contribution to the final comprehensive feature vector.

[0052] In some embodiments, to further improve the accuracy of weighted fusion, the weighting coefficients can also be jointly optimized with the hyperparameters of the model (such as the learning rate, regularization coefficient, etc.), and these parameters can be dynamically adjusted through reinforcement learning or other optimization algorithms to further optimize the fusion effect and training performance of the model.

[0053] Therefore, the comprehensive feature vector obtained through the weighted fusion mechanism will provide more comprehensive and accurate input data for subsequent model optimization, decision analysis, and business prediction. In the present invention, this mechanism provides strong support for the efficient utilization of multimodal data, enhancing the generalization ability and practical application effect of the deep learning model.

[0054] Optimizing the comprehensive feature vector using the target loss function; In this embodiment, optimizing the comprehensive feature vector using the target loss function is an important step to ensure that the model can be effectively adjusted according to the actual business requirements. The goal of this optimization process is to optimize the comprehensive feature vector by minimizing the loss function, enabling the final model to exhibit higher accuracy and robustness on the given task. Through this process, the deep neural network model can better fit the training data and make more accurate predictions for new unseen data.

[0055] Specifically, the comprehensive feature vector is generated through a weighted fusion mechanism, which combines information from different data modalities (such as sales data, customer data, inventory data, and marketing data). In order to maximize the effectiveness of the comprehensive feature vector in subsequent models, it needs to be optimized through the target loss function. In general, the loss function is used to measure the difference between the output of the model and the target value, and its purpose is to minimize the difference, thereby improving the prediction accuracy of the model.

[0056] As an option, in the optimization process, the loss function usually includes multiple parts. For example, for regression problems, a commonly used loss function is the mean square error (MSE), which is calculated as:

[0057] in, is the actual value, is the predicted value, is the total number of samples. In the present invention, the comprehensive feature vector is used as the input of the model, and the role of the target loss function is to continuously adjust the weights and biases of the network so that the prediction results of the model are as close to the actual business goals as possible. When learning multimodal data, the design of a weighted loss function may be involved to adjust the contribution of different modal data in the optimization process according to their quality and importance.

[0058] Specifically, the optimization process updates the model parameters by using optimization algorithms such as gradient descent. The goal is to minimize the loss function, thereby continuously optimizing the representation of the comprehensive feature vector so that the vector can better reflect the operating data of small and medium-sized enterprises, thereby improving the prediction accuracy of the model. During the training process, the loss function is minimized through the back-propagation algorithm, which involves calculating the gradient of the loss function to each model parameter and updating these parameters to reduce the prediction error.

[0059] In one possible implementation, in order to optimize the model more accurately, it may be necessary to combine regularization methods to avoid overfitting problems. Common regularization methods include L1 regularization and L2 regularization, which prevent the model from over-relying on the noise in the training data by penalizing the complexity of the model, thereby improving its generalization ability. For example, the loss function of L2 regularization can be expressed as:

[0060] in, is the regularization parameter, is the weight of the model, is the number of weights in the model, that is, the number of model parameters. This regularization term penalizes the weights to prevent the model from overfitting the training data and ensure the adaptability and stability of the model in the real environment.

[0061] In some embodiments, when optimizing the objective loss function, an optimization algorithm using an adaptive learning rate may also be involved, such as the Adam optimizer. This optimizer can adaptively adjust the learning rate according to the gradient of each parameter, thereby accelerating convergence and avoiding problems such as gradient vanishing or gradient explosion. The update rule of the Adam optimizer is as follows:

[0062]

[0063]

[0064] where is the bias correction of the first moment estimate in the -th iteration, representing the estimate of the first moment (i.e., the mean) of the gradient, and the value obtained after bias correction, is the first moment estimate in the -th iteration, representing the weighted average of the current gradient, reflecting the trend of the gradient, is the decay rate of the first moment estimate, set to be close to 1, controlling the weight of the first moment estimate in the historical gradients. is the bias correction of the second moment estimate in the -th iteration, representing the estimate of the second moment (i.e., the variance) of the gradient, and the value obtained after bias correction. is the second moment estimate in the -th iteration, representing the weighted average of the square of the current gradient, reflecting the degree of fluctuation of the gradient. is the decay rate of the second moment estimate, set to be close to 1, controlling the weight of the second moment estimate in the squares of the historical gradients. is the parameter of the model in the -th iteration. This parameter is updated according to the gradient in each iteration, and the goal is to minimize the loss function. is the model parameter in the -th iteration. is the learning rate, controlling the step size of the adjustment of the model parameter in each update. A smaller learning rate can ensure the stable update of the model parameter, while a larger learning rate may lead to an overly large update and non-convergence. is a constant to prevent division by zero, usually set to a very small value to ensure that there is no zero value in the division operation and avoid numerical stability problems.

[0065] Among them, the Adam optimizer combines momentum (through the first moment estimate) and adaptive learning rate (through the second moment estimate), which can effectively improve the convergence speed of the model and avoid the problems of gradient explosion or gradient vanishing that may occur in the traditional gradient descent method.

[0066] Based on the optimized comprehensive feature vectors and the results of the loss function, reinforcement learning is used to dynamically adjust the hyperparameters of the deep neural network model to further optimize the fused features and the training process; In this embodiment, based on the optimized comprehensive feature vectors and the results of the loss function, reinforcement learning is adopted to dynamically adjust the hyperparameters of the deep neural network model, aiming to further optimize the fused features and the training process. As a learning method based on a reward mechanism, reinforcement learning can dynamically adjust the hyperparameters of the model (such as the learning rate, regularization coefficient, etc.) during the training process to achieve more efficient feature representation and training optimization. Through this optimization process, the performance of the deep neural network can be improved, enabling the model to better adapt to and predict in complex business data.

[0067] Specifically, through the optimization of the aforementioned weighted fusion mechanism and the objective loss function, an optimized comprehensive feature vector is obtained. These feature vectors are used as the input of the deep neural network. In the further optimization process, it is necessary to dynamically adjust the hyperparameters of the model to better adapt to the changes in business data and the training requirements of the model. Reinforcement learning can learn how to adjust these hyperparameters according to the current training situation, thereby accelerating the training process and improving the generalization ability of the final model.

[0068] Generally, the core idea of reinforcement learning is to make decisions based on the reward signal through interaction with the environment. In the present invention, the environment refers to the state during the model training process, including the current comprehensive feature vector, the results of the loss function, and the current hyperparameters of the model. By continuously exploring and adjusting the hyperparameters, the reinforcement learning agent can learn a set of optimal hyperparameter settings, enabling the training process to proceed more efficiently and significantly improving the performance of the model.

[0069] As an option, the optimization process of reinforcement learning can be guided by defining a reward function. Specifically, the reward function can be designed as an inverse relationship based on the loss function value, that is, the smaller the loss function value, the higher the reward. Suppose the current loss function is , then the reward function can be defined as:

[0070] where is the loss value under the current model parameters . Through this reward function, the reinforcement learning agent can continuously adjust the hyperparameters and update its policy according to the current reward value in each training process. Through multiple iterations, the agent can learn to select appropriate hyperparameters at different training stages, thereby optimizing the feature fusion and the model training process.

[0071] Specifically, reinforcement learning dynamically selects and adjusts hyperparameters, such as the learning rate, weight decay coefficient, etc., through interaction with the training process. The reinforcement learning agent can control the magnitude of each step update by adjusting the learning rate to ensure the convergence and stability of the training process. Suppose the current learning rate is , then the agent selects an appropriate value to ensure the convergence and stability of the training process. The reinforcement learning agent can also adjust the regularization parameter to control the model complexity and prevent overfitting.

[0072] In one possible implementation, reinforcement learning algorithms such as Deep Q-Network (DQN) can be used to adjust hyperparameters. In this method, the agent generates a value function by evaluating the training effects under different hyperparameter configurations. This value function is used to measure the "goodness" of selecting a certain hyperparameter in a specific state and its impact on the model performance. The update of the value function follows the following rule:

[0073] where is the value of taking action in state , is the learning rate of the value function, is the current reward, is the discount factor, is the next state and is the maximum value of all possible actions

[0074] In some embodiments, other reinforcement learning algorithms, such as policy gradient methods or evolutionary strategies, can be combined to further improve the accuracy and efficiency of hyperparameter optimization. The policy gradient method directly selects the optimal hyperparameter configuration by optimizing a probability distribution, while the evolutionary strategy simulates the natural selection process and searches for the optimal hyperparameter combination through genetic algorithms.

[0075] Therefore, through this optimization process of reinforcement learning, the hyperparameters of the deep neural network can be adjusted more precisely, enabling the comprehensive feature vector to better adapt to complex business data, further improving the training efficiency and prediction accuracy of the model. Ultimately, the present invention can achieve a more intelligent and efficient model optimization process, providing more accurate and effective decision-making support for small, medium, and micro enterprises.

[0076] Incremental learning and online training are performed on the optimized model to adapt to the changes and updates of data during the operation of small, medium and micro enterprises.

[0077] In this embodiment, in order to enable the model to adapt to the continuously changing and updated data during the operation of small, medium and micro enterprises, an incremental learning and online training method is adopted. This method enables the deep neural network model to dynamically adjust when receiving new data, thereby ensuring that the model continues to be optimized during the data update process without having to train from scratch. Through incremental learning and online training, the model can be quickly updated according to new business data and timely adapt to changes in the market environment.

[0078] Specifically, on the basis of the aforementioned weighted fusion mechanism and loss function optimization, incremental learning and online training strategies are further introduced. The core idea of these strategies is that in the case of continuously changing data, the model can adapt to new data and optimize its prediction ability at the lowest computational cost. Specifically, during incremental learning, the model only uses new data to update existing knowledge without having to retrain the entire model. Compared with traditional batch training methods, incremental learning can process dynamic data streams more efficiently.

[0079] Generally, the goal of incremental learning is to enable the model to only perform a small number of model updates when receiving new data samples. In this way, the model can effectively utilize new data and be optimized in a short time without repeating the entire training process. For example, assume that at the parameters of the model are , after receiving new data, the updated parameters are , so the incremental update can be calculated by the following rule:

[0080] where represents the incremental update of the model parameters, adjusting the model based on new data. The size of the incremental update can be set according to the importance of the data and the learning ability of the model to ensure the stability of the model after each update.

[0081] As an option, incremental learning can be combined with online training. During online training, the model can process newly input data in real time and update immediately. Different from traditional offline training, the latter usually requires collecting a large amount of data over a period of time and performing batch processing. Online training methods usually involve continuous processing of data streams, that is, as each new data point arrives, the model can be quickly updated. For example, in online training, assume that each data point is and the corresponding label is , the model adapts to new data by gradually adjusting its weights. The update formula can be expressed as:

[0082] where is the gradient of the loss function with respect to the model parameters, representing the partial derivative of the loss function with respect to each parameter at the current parameters. It reflects the change trend of the loss function and indicates how the model parameters should be adjusted to reduce the prediction error. Specifically, is the prediction error of the model on the data point business data processing for the label .

[0083] Specifically, by combining the strategies of incremental learning and online training, efficient update of the model when receiving new data can be achieved. By appropriately adjusting the learning rate and the incremental update strategy, it can be ensured that the model converges stably during the update process and will not result in performance degradation due to too rapid or drastic data changes. Especially in the operating environment of small, medium and micro enterprises, the data may change very frequently, and through these strategies, the model can adapt to new market demands, customer changes and other dynamic factors.

[0084] In a possible implementation, in order to ensure the effectiveness of incremental learning and online training, a dynamic learning strategy can also be introduced. By automatically adjusting the learning rate according to the current performance of the model at each update, the instability caused by too large update steps can be effectively avoided. Specifically, the learning rate can be adaptively adjusted according to the error or performance feedback of the model during the training process, usually using exponential decay or dynamic adjustment based on the model performance.

[0085] In some embodiments, in order to improve the efficiency of incremental learning, transfer learning techniques can be combined to utilize the knowledge of existing models to accelerate the learning process of new data. For example, by pre-training an initial model and then fine-tuning it with new data in the incremental learning stage, the ability of the model to adapt to new data can be improved and learning from scratch can be avoided.

[0086] Therefore, by combining incremental learning and online training, it can be ensured that the deep neural network model can be optimized in real time and continuously when facing the data changes in the operation process of small, medium and micro enterprises, providing more accurate decision-making support. At the same time, incremental learning and online training can also greatly reduce the consumption of training time and computing resources, enabling small, medium and micro enterprises to more efficiently utilize limited resources for model update and optimization.

[0087] Although embodiments of the present invention have been shown and described, it will be understood by those of ordinary skill in the art that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention, and the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. The AI ​​business model training method based on the global operation of small and medium-sized businesses is characterized by: The following steps are involved: Collect operational data of small and medium-sized enterprises, including sales data, customer data, inventory data and marketing data; Pre-processing of operational data; Use a deep neural network model to extract features from operational data and obtain feature representation of operational data; The feature representations are fused through a weighted fusion mechanism to obtain a comprehensive feature vector; Optimize the comprehensive feature vector using the objective loss function; Based on the optimized comprehensive feature vector and loss function results, reinforcement learning is used to dynamically adjust the hyperparameters of the deep neural network model to further optimize the fusion features and training process; Perform incremental learning and online training on the optimization model to adapt to changes and updates in data during the operation of small and medium-sized enterprises.

2. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The preprocessing includes deduplication, filling missing values, outlier processing, standardization and normalization.

3. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The deep neural network model includes a Transformer model, a convolutional neural network model and an LSTM model.

4. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The weighted fusion mechanism generates a comprehensive feature vector according to a weighted summation method based on a weighted coefficient, and the weighted coefficient is dynamically adjusted through an optimization algorithm.

5. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 4 is characterized in that: The optimization algorithm includes a gradient descent method, which updates weight coefficients through back propagation to minimize the loss between the fused features and the target business indicators.

6. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The reinforcement learning includes a learning rate, a regularization coefficient and a batch size, and the reinforcement learning evaluates the performance of the model on a validation set through a reward function.

7. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 6 is characterized in that: The reward function gives a reward value by comprehensively considering the decrease of the loss function and the improvement of the business indicators, and updates the hyperparameters of the model according to the current state and the reward value.

8. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The loss function includes an error metric between the fused feature and the actual business target, and the error metric includes a mean square error and a cross entropy error.

9. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The incremental learning and online training update the existing model every time new data arrives.

10. The AI ​​business big model training method based on the global operation of small and medium-sized businesses according to claim 1 is characterized in that: The weighted fusion mechanism performs weighted summation on the feature representations, multiplying the feature values ​​of the feature representations by their corresponding weighting coefficients, and the weighting coefficients are dynamically adjusted during the training process through an optimization algorithm.

Images