Client demand analysis method based on AI artificial intelligence CRM system data

Abstract

The invention discloses a customer demand analysis method based on AI artificial intelligence CRM system data, and the method comprises the following steps: S1, collecting multi-source heterogeneous customer data, executing the preprocessing, and carrying out the fusion; s2, customer behavior features are extracted and fused to form feature vector representation, feature importance evaluation is carried out, and high-correlation feature subsets are screened; s3, inputting the high-correlation feature subset into a swarm intelligence optimization module, and outputting initial network configuration parameters; s4, constructing and initializing a customer demand prediction model according to the initial network configuration parameters; s5, training the customer demand prediction model, and dynamically adjusting the structure of the customer demand prediction model; s6, continuing to iteratively train the customer demand prediction model; and S7, generating a customer demand prediction result. According to the method, intelligent swarm optimization and deep learning methods are fused, accurate prediction of customer demands is realized, and the method has the advantages of high adaptability, accurate prediction and fast response.

Description

Technical Field

[0001] The present invention relates to the technical field of customer relationship management, and in particular to a method for analyzing customer needs based on AI (artificial intelligence) CRM system data. Background Art

[0002] With the continuous improvement of the level of enterprise informatization, customer relationship management (CRM) systems have become an important tool for enterprises to acquire, manage and maintain customer resources. Traditional CRM systems mainly focus on structured data, such as basic customer information, transaction records, service history, etc., and classify and manage customers and maintain relationships through rule engines or simple data analysis methods. Such systems perform well in processing static and regular customer data. However, with the popularization of the Internet, mobile terminals and social platforms, customer behavior data has become increasingly complex and heterogeneous, covering a large amount of unstructured information such as communication texts, social media interactions, browsing behavior trajectories, etc. Traditional CRM systems have obvious limitations when dealing with such massive, multi-source, heterogeneous and dynamically changing data. They are unable to fully tap into customers' potential needs, resulting in poor personalized recommendations and precision marketing results.

[0003] In recent years, the development of artificial intelligence (AI), particularly deep learning, has provided new insights for the intelligent processing of customer data. Some research has attempted to apply deep neural networks to customer data analysis, using trained predictive models to identify potential customers, recommend products, or predict customer churn. However, existing AI-based CRM applications often suffer from several issues. First, current applications often focus on modeling from a single data source, such as transactional or social data, lacking the unified integration of heterogeneous data from multiple sources. This results in a one-sided customer portrait and an inability to fully capture customer behavior. Second, in data preprocessing, most methods rely on manual rules or simple cleaning processes, lacking systematic design for differentiated preprocessing of structured and unstructured data, impacting the accuracy of subsequent feature extraction and modeling.

[0004] In the feature extraction stage, existing technologies usually use traditional statistical features or shallow models to generate feature vectors, failing to fully utilize the ability of deep learning to extract complex nonlinear relationships in high-dimensional feature space. At the same time, feature selection methods are mostly based on a single indicator, such as information gain or variance filtering, and fail to fully integrate multiple feature evaluation criteria for comprehensive evaluation and dynamic screening, resulting in insufficient relevance and representativeness of the input feature set, affecting the model training effect.

[0005] In terms of model structure design, existing methods mostly use fixed-structure neural networks, such as multi-layer perceptrons, convolutional neural networks, or long short-term memory networks. Since fixed structures cannot adapt to different data characteristics and dynamic business needs, the model lacks flexibility and is difficult to achieve personalized adaptive adjustment.

[0006] During the model training process, traditional training processes mostly use fixed hyperparameter settings such as learning rate, batch size and regularization coefficient, and lack dynamic optimization and adaptive adjustment mechanisms, resulting in a lack of flexibility and intelligence in the training process. Although some studies have introduced swarm intelligence algorithms for hyperparameter optimization, they are usually only performed on a coarse-grained basis in the initial stage, lacking refined adjustment strategies that are dynamically linked to training. In addition, existing methods have weak control over the complexity of the model structure during training, which can easily lead to model scale expansion and increased training overhead, while also causing the risk of overfitting and affecting the promotion and application of the model.

[0007] Traditional CRM systems rely heavily on periodic batch updates to process real-time new data, failing to achieve true real-time reasoning and online adaptive model updates. This limits the timeliness and accuracy of customer demand forecasts. While existing research has proposed some incremental learning frameworks, most CRM applications lack systematic designs for dynamic streaming of multi-source data, feature evolution, and continuous model optimization.

[0008] In summary, existing technologies have obvious deficiencies in standardized preprocessing of multi-source heterogeneous customer data, deep feature extraction, comprehensive evaluation of feature importance, efficient network structure optimization, dynamic hyperparameter adjustment during training, structural complexity control, and real-time reasoning capabilities, making it difficult to meet the current complex and changing customer demand prediction and practical application needs of personalized marketing. Summary of the Invention

[0009] One purpose of the present invention is to propose a method for analyzing customer demand based on AI artificial intelligence CRM system data. The present invention integrates the fractional-order cuckoo search algorithm and the improved black hole optimization algorithm, and combines deep learning methods to perform standardized modeling and feature extraction on multi-source heterogeneous customer data in the CRM system, thereby achieving accurate prediction of customer potential needs. It has the advantages of high optimization efficiency, strong model adaptability and fast prediction response, effectively improving the level of customer management intelligence and the quality of corporate marketing decision-making.

[0010] According to an embodiment of the present invention, a method for analyzing customer needs based on AI artificial intelligence CRM system data includes the following steps:

[0011] S1. Collect multi-source heterogeneous customer data in the CRM system, perform preprocessing on the collected customer data, fuse the preprocessed data, and output a standardized customer data set;

[0012] S2. Based on the standardized customer dataset, we use Embedding and Multi-layer Perceptron to extract customer behavior features, fuse them into feature vector representations, and evaluate the feature importance of the feature vector representations to select highly relevant feature subsets.

[0013] S3, input the high correlation feature subset into the swarm intelligence optimization module and output the initial network configuration parameters;

[0014] S4. Build and initialize a customer demand prediction model based on the initial network configuration parameters;

[0015] S5. Apply the backpropagation algorithm to perform local gradient updates on the customer demand forecasting model, evaluate the model's global performance indicators, and call the swarm intelligence optimization module based on the evaluation results to dynamically adjust the structure of the customer demand forecasting model.

[0016] S6. Continue iteratively training the customer demand prediction model, with the goal of minimizing the potential customer demand prediction error, until a preset convergence criterion or a maximum number of training rounds is reached;

[0017] S7. Input the new customer data collected in real time into the trained customer demand prediction neural network model to generate customer demand prediction results.

[0018] Optionally, the customer data includes structured data and unstructured data, the structured data includes basic customer information and transaction records, and the unstructured data includes customer communication texts, social media interaction content and browsing behavior records. The preprocessing of the structured data includes data format unification, missing value filling, outlier removal, numerical normalization and One-Hot encoding, and the preprocessing of the unstructured data includes data cleaning, missing value filling, outlier removal and text encoding.

[0019] Optionally, the customer behavior characteristics include historical purchase behavior characteristics, interactive behavior characteristics, interest preference characteristics and social behavior characteristics.

[0020] Optionally, the feature importance evaluation calculates the information gain and mutual information between feature vector representations, and takes the weighted sum of the information gain and mutual information as the importance score, sorts the importance scores from high to low, and filters out feature vector representations that satisfy an importance score greater than a threshold to form a high-correlation feature subset.

[0021] Optionally, the swarm intelligence optimization module applies a fractional-order cuckoo search algorithm and an improved black hole optimization algorithm, wherein the fractional-order cuckoo search algorithm performs a global search of network structure parameters, and the improved black hole optimization algorithm performs fine optimization of hyperparameters.

[0022] Optionally, the S3 specifically includes:

[0023] S31, generating a feature subset matrix based on the high-correlation feature subset;

[0024] S32. Perform normalization coding on the neural network object to be optimized to generate an initial parameter population. The neural network object to be optimized includes the number of hidden layers, the number of nodes in each layer, the connection mode coding, the learning rate, the regularization coefficient, the batch size, and the optimizer type coding:

[0025]

[0026] Among them, P (0) represents the initial parameter population, represents the parameter vector of the i-th group, m0 represents the initial population size;

[0027] S33, randomly dividing the initial parameter population into a fast exploration group and a fine mining group according to a set ratio;

[0028] S34. Define the performance evaluation index based on the feature subset matrix using the cross entropy loss function:

[0029]

[0030] in, represents the cross entropy loss of the i-th group of parameters in the t-th generation, N represents the number of validation samples, and y k represents the true label of the k-th sample, represents the predicted probability of the kth sample;

[0031] S35. Apply the fractional-order cuckoo search algorithm within the rapid exploration group and the fine mining group, respectively, and update the individual step size according to the asynchronous adaptive mechanism. The fractional-order cuckoo search algorithm introduces fractional-order calculus theory on the basis of the cuckoo search algorithm, and is a swarm intelligence optimization algorithm based on the motion trajectory of the fractional-order levy flight modeling population:

[0032]

[0033] in, represents the step length of the i-th individual in the t-th generation, γ0 represents the initial step length factor, β represents the step length attenuation coefficient, r i represents a random number in the interval (0,1), and exp represents the natural exponential function;

[0034] To perform an individual location update:

[0035]

[0036] in, represents the position of the i-th individual in the t+1 generation, represents the position of the i-th individual in the t-th generation, represents the step length of the i-th individual in the t-th generation, Levy(λ) represents the Levy flight step length vector, λ represents the stability index parameter of the Levy flight distribution, μ represents the fractional order influence weight, H represents the historical step length window, ω h represents the fractional weight coefficient, represents the position of the i-th individual in the th generation, represents the position of the i-th individual in the th-1 generation;

[0037] When the maximum number of iterations is reached or the change in the optimal validation set loss of the group within w consecutive generations is less than the convergence judgment threshold, the fractional cuckoo search algorithm terminates the iteration;

[0038] S36. Dynamically adjust the population size based on the cross entropy loss decrease rate of the current generation group:

[0039] m (t+1) =m (t) ×(1-δR t );

[0040] Among them, m (t+1) represents the population size of the t+1 generation, m (t) represents the size of the t-generation population, δ represents the population adjustment coefficient, R t Indicates the cross entropy loss decrease rate of the current generation group;

[0041] S37. After the cuckoo search termination condition is met, parameters that satisfy the cross entropy loss less than the loss screening threshold are selected from the optimized population to construct a high-quality parameter set;

[0042] S38. Execute an improved black hole optimization algorithm based on a high-quality parameter set.

[0043] Optionally, the S38 specifically includes:

[0044] S381. Extract the hyperparameter part from the high-quality parameter set and select the first k groups of individuals with the smallest cross entropy loss as the multi-black hole centers;

[0045] S382. The remaining individuals in the high-quality parameter set except the black hole center are regarded as ordinary individuals. In the black hole optimization phase, each ordinary individual selects the nearest black hole based on the Euclidean distance and updates its position according to the adaptive attraction coefficient:

[0046]

[0047] Among them, p ′ i (t+1) represents the position of the ith common individual in the t+1th generation, p ′ i (t)represents the position of the i-th common individual in the t-th generation, represents the attraction coefficient of the i-th individual in the t-th generation, represents the jth black hole center individual of the tth generation;

[0048] The adaptive attraction coefficient is dynamically calculated and updated based on the distance between the current individual and the black hole:

[0049]

[0050] in, represents the attraction coefficient of the i-th individual in the t-th generation, ρ0 represents the initial attraction coefficient, κ represents the attraction contraction control coefficient, represents the distance from the ith individual of the tth generation to the nearest black hole, represents the maximum distance between individuals in the t-th generation population;

[0051] S383. During the black hole optimization iteration process, if the performance of the validation set after the normal individual position is updated is better than that of the corresponding black hole individual, the corresponding black hole position is replaced by the normal individual;

[0052] S384. When the black hole optimization reaches the maximum iteration number or meets the convergence condition, a black hole with the best verification performance is selected from the black hole set, and the initial network configuration parameters are output.

[0053] Optionally, the S4 specifically includes:

[0054] S41, extracting initial network configuration parameters, including the number of hidden layers, the number of nodes in each layer, the connection mode encoding, the learning rate, the regularization coefficient, the batch size, and the optimizer type encoding;

[0055] S42. Initializing a customer demand prediction model based on the extracted parameters. The customer demand prediction model adopts a hierarchical structure of an encoding module, a hidden representation module, and a decoding module. The connection relationship between the modules is determined by the connection mode code.

[0056] S43, the encoding module receives the input feature vector, performs linear transformation and activation operations, and obtains a preliminary hidden representation. The hidden representation module calculates the state of each hidden layer layer by layer, and the decoding module generates a potential customer demand prediction output based on the final hidden layer;

[0057] S44. After the structure of the customer demand prediction model is completed, the highly correlated feature subsets are divided according to the batch size, the initial learning rate is set, the optimizer and regularization coefficient are applied, and the compilation initialization of the customer demand prediction model is completed;

[0058] S45. To support dynamic topological changes during the training phase, a node growth mechanism is defined. When the local gradient norm of a layer is greater than the node growth trigger threshold during training, the number of newly added nodes is:

[0059]

[0060] Where, ΔN l Indicates the number of new nodes in the first layer, α g Indicates the node growth ratio coefficient, represents the second norm of the local gradient of the first layer, τ g Indicates the node growth trigger threshold;

[0061] S46. Synchronously define a node pruning mechanism. When the sparsity of the node output of a certain layer is higher than the pruning threshold, remove the low-activity nodes and update the node connection matrix. The sparsity is measured using the L1 norm.

[0062] Optionally, the S5 specifically includes:

[0063] S51, initialize the training cycle counter, set it to zero, configure the initial learning rate, regularization coefficient and batch size parameters, and divide the standardized customer data set into multiple small batch data sets;

[0064] S52. For each training batch, extract an input feature vector from the small batch data set, input the feature vector into the customer demand prediction model, and generate a corresponding prediction output;

[0065] S53. Based on the predicted output and the actual label data, the batch loss value is calculated, the batch loss value is used as training feedback, and the back propagation process is performed to calculate the local gradient of the model parameter set;

[0066] S54. Apply the optimizer rule to locally update the model parameters based on the calculated local gradient to generate an updated parameter set after the current training batch;

[0067] S55. After completing all batch training, perform inference based on the validation set data using the currently updated model parameters to calculate the corresponding validation set overall loss value.

[0068] S56. When the decrease in the overall loss value of the validation set is less than the set threshold within a set number of consecutive training cycles, the swarm intelligence optimization module is started to execute the structure adaptive adjustment process;

[0069] S57, in the structural adaptation stage, based on the current latest parameter set, the optimization parameter population is initialized as the input of the swarm intelligence optimization process;

[0070] S58. In the swarm intelligence optimization process, the overall loss minimization of the verification set is used as the search goal, and the parameter population is iteratively searched and evolved to dynamically optimize the structure and parameter configuration of the customer demand prediction model.

[0071] Optionally, the judgment condition of the preset convergence criterion is: if the decrease in the prediction error is less than the error convergence threshold, the training process is terminated.

[0072] The beneficial effects of the present invention are:

[0073] First of all, the present invention provides a method for analyzing customer demand based on AI artificial intelligence CRM system data, which can effectively solve the multiple bottlenecks of existing technologies in heterogeneous processing of customer data, deep feature extraction, model structure flexibility, intelligent control of training and real-time prediction, and significantly improve the accuracy of customer demand prediction, the real-time response of the system and the overall intelligence level of model operation.

[0074] Secondly, by introducing a classification preprocessing strategy for structured and unstructured data, combined with processing processes such as data format unification, missing value filling, outlier elimination, numerical normalization and text encoding, the CRM system's compatibility with multi-source heterogeneous data has been significantly enhanced, and the standardization and high quality of data input have been achieved, ensuring the basic reliability of subsequent model training from the source. The feature extraction method that combines Embedding and multi-layer perceptron not only enhances the ability to express complex customer behavior characteristics, but also can fully explore the high-order correlations between historical purchases, interest preferences, interactive behaviors and social behaviors, thereby improving the precision and discriminability of customer portraits.

[0075] Thirdly, in response to the problem of the effectiveness of input features, the present invention proposes a feature importance evaluation mechanism based on the fusion of information gain and mutual information, and generates a high-correlation feature subset through weighted scoring and dynamic threshold screening, thereby realizing adaptive optimization of feature input dimensions, reducing model redundancy, and improving training efficiency and generalization performance. In terms of model parameter and structure optimization, the fractional-order cuckoo search algorithm and the improved black hole optimization algorithm are innovatively introduced into the training process in a coordinated manner, which not only realizes the global search of network structure parameters and fine-tuning of hyperparameters, but also enhances the global exploration ability and local convergence accuracy of the optimization algorithm through mechanisms such as asynchronous adaptive step size and dynamic shrinkage of search radius, greatly reducing the dependence on manual parameter adjustment and improving the system's automatic modeling capability.

[0076] In addition, the present invention further proposes a dynamic adjustment and compression mechanism for the neural network structure. During training, by real-time monitoring of the local gradient norm and node output sparsity, adaptive node growth, node pruning, connection sparsification, and subnetwork pruning operations are performed, enabling the network structure to achieve a dynamic balance between model complexity and predictive performance, significantly alleviating the overfitting problem and reducing computing resource overhead. Combining a dynamic linkage strategy with swarm intelligence optimization and structural reorganization mechanisms, the network topology can be continuously optimized during the training cycle, improving the model's structural adaptability and training convergence speed.

[0077] Finally, the present invention introduces a termination mechanism based on the reduction of prediction error in model training convergence control, ensuring that the model is fully trained while avoiding over-learning; during the deployment phase, the system can receive new customer data in real time, quickly generate customer demand forecast results through the trained model, and support personalized recommendation decisions, providing enterprises with more intelligent, accurate and sustainable customer management and marketing support. BRIEF DESCRIPTION OF THE DRAWINGS

[0078] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0079] Figure 1 This is a flowchart of a method for analyzing customer needs based on AI artificial intelligence CRM system data proposed by the present invention;

[0080] Figure 2 This is a structural diagram of the initial network configuration parameter optimization using a fractional-order cuckoo search algorithm and an improved black hole optimization algorithm based on AI artificial intelligence CRM system data customer demand analysis proposed in the present invention;

[0081] Figure 3 This is a schematic diagram of the dynamic optimization mechanism of local gradient update and swarm intelligence linkage in the training process based on customer demand analysis of AI artificial intelligence CRM system data proposed by the present invention. DETAILED DESCRIPTION

[0082] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0083] refer to Figure 1-3 A method for analyzing customer needs based on AI artificial intelligence CRM system data includes the following steps:

[0084] S1. Collect multi-source heterogeneous customer data in the CRM system, perform preprocessing on the collected customer data, fuse the preprocessed data, and output a standardized customer data set;

[0085] S2. Based on the standardized customer dataset, we use Embedding and Multi-layer Perceptron to extract customer behavior features, fuse them into feature vector representations, and evaluate the feature importance of the feature vector representations to select highly relevant feature subsets.

[0086] S3, input the high correlation feature subset into the swarm intelligence optimization module and output the initial network configuration parameters;

[0087] S4. Build and initialize a customer demand prediction model based on the initial network configuration parameters;

[0088] S5. Apply the backpropagation algorithm to perform local gradient updates on the customer demand forecasting model, evaluate the model's global performance indicators, and call the swarm intelligence optimization module based on the evaluation results to dynamically adjust the structure of the customer demand forecasting model.

[0089] S6. Continue iteratively training the customer demand prediction model, with the goal of minimizing the potential customer demand prediction error, until a preset convergence criterion or a maximum number of training rounds is reached;

[0090] S7. Input the new customer data collected in real time into the trained customer demand prediction neural network model to generate customer demand prediction results.

[0091] The present invention proposes a method for analyzing customer demand based on AI artificial intelligence CRM system data. By standardizing multi-source heterogeneous customer data, extracting deep behavioral features, and combining swarm intelligence optimization to dynamically adjust the customer demand prediction model, the method ultimately achieves accurate prediction of customer potential needs. This method effectively overcomes the problems of traditional CRM systems in processing single data, rigid models, and low prediction accuracy, improves the model's adaptability, real-time response, and prediction accuracy, and enhances the company's intelligence level and decision-making efficiency in customer relationship management and personalized marketing.

[0092] In this embodiment, the customer data includes structured data and unstructured data. The structured data includes basic customer information and transaction records, and the unstructured data includes customer communication texts, social media interaction content, and browsing behavior records. The preprocessing of the structured data includes data format unification, missing value filling, outlier removal, numerical normalization, and one-hot encoding. The preprocessing of the unstructured data includes data cleaning, missing value filling, outlier removal, and text encoding.

[0093] The present invention formulates differentiated preprocessing processes for structured data and unstructured data respectively, to achieve data format unification, defect repair, anomaly cleaning and effective encoding, thereby ensuring high quality and standardization of input data. Compared with traditional rough preprocessing methods, the present invention can better adapt to different types of customer information sources, lay a solid foundation for subsequent feature extraction and model training, greatly improve data utilization efficiency and modeling effect, and reduce the interference and impact of data noise on the customer demand prediction model training process.

[0094] In this embodiment, the customer behavior characteristics include historical purchase behavior characteristics, interactive behavior characteristics, interest preference characteristics and social behavior characteristics.

[0095] The present invention explicitly extracts multi-dimensional customer behavior characteristics such as historical purchasing behavior, interactive behavior, interest preferences and social behavior to form a comprehensive customer behavior portrait. By systematically integrating behavioral characteristics of different dimensions, compared with single-dimensional modeling, it can more deeply explore customers' potential needs and interests, improve feature expression capabilities and model recognition accuracy, provide rich and accurate behavioral basic support for customer demand prediction, and enhance the model's ability to model and understand complex customer behavior patterns.

[0096] In this embodiment, the feature importance evaluation calculates the information gain and mutual information between feature vector representations, and takes the weighted sum of the information gain and mutual information as the importance score. The importance scores are sorted from high to low, and feature vector representations that meet the importance score greater than the threshold are screened out to form a high-correlation feature subset.

[0097] The present invention proposes to evaluate the importance of features by means of a weighted fusion of information gain and mutual information, screen out highly relevant feature subsets for input into the model, and effectively remove redundant features. Compared with the traditional single-indicator screening method, the present invention can take into account both the correlation and complementarity of features, dynamically adjust feature inputs, improve training efficiency and model generalization ability, reduce the risk of overfitting, and provide a more compact, efficient, and expressive input feature set for the customer demand prediction model.

[0098] In this embodiment, the swarm intelligence optimization module applies a fractional-order cuckoo search algorithm and an improved black hole optimization algorithm. The fractional-order cuckoo search algorithm performs a global search of network structure parameters, and the improved black hole optimization algorithm performs fine optimization of hyperparameters.

[0099] This invention innovatively applies the fractional-order cuckoo search algorithm and the improved black hole optimization algorithm in the customer demand forecasting task, and collaboratively performs global and local optimization of the network structure and hyperparameters. By introducing fractional-order calculus and a dynamic search radius mechanism, it significantly improves the convergence speed and accuracy of parameter optimization, avoids falling into local optimality, reduces the cost of manual parameter adjustment, enhances the adaptability of the model structure and the intelligence of the training process, and provides a guarantee for the accurate and efficient construction of a customer demand forecasting model.

[0100] In this embodiment, S3 specifically includes:

[0101] S31, generating a feature subset matrix based on the high-correlation feature subset;

[0102] S32. Perform normalization coding on the neural network object to be optimized to generate an initial parameter population. The neural network object to be optimized includes the number of hidden layers, the number of nodes in each layer, the connection mode coding, the learning rate, the regularization coefficient, the batch size, and the optimizer type coding:

[0103]

[0104] Among them, P (0) represents the initial parameter population, represents the parameter vector of the i-th group, m0 represents the initial population size;

[0105] S33, randomly dividing the initial parameter population into a fast exploration group and a fine mining group according to a set ratio;

[0106] S34. Define the performance evaluation index based on the feature subset matrix using the cross entropy loss function:

[0107]

[0108] in, represents the cross entropy loss of the i-th group of parameters in the t-th generation, N represents the number of validation samples, and y k represents the true label of the k-th sample, represents the predicted probability of the kth sample;

[0109] S35. Apply the fractional-order cuckoo search algorithm within the rapid exploration group and the fine mining group, respectively, and update the individual step size according to the asynchronous adaptive mechanism. The fractional-order cuckoo search algorithm introduces fractional-order calculus theory on the basis of the cuckoo search algorithm, and is a swarm intelligence optimization algorithm based on the motion trajectory of the fractional-order levy flight modeling population:

[0110]

[0111] in, represents the step length of the i-th individual in the t-th generation, γ0 represents the initial step length factor, β represents the step length attenuation coefficient, r i represents a random number in the interval (0,1), and exp represents the natural exponential function;

[0112] To perform an individual location update:

[0113]

[0114] in, represents the position of the i-th individual in the t+1 generation, represents the position of the i-th individual in the t-th generation, represents the step length of the i-th individual in the t-th generation, Levy(λ) represents the Levy flight step length vector, λ represents the stability index parameter of the Levy flight distribution, μ represents the fractional order influence weight, H represents the historical step length window, ω h represents the fractional weight coefficient, represents the position of the i-th individual in the th generation, represents the position of the i-th individual in the th-1 generation;

[0115] When the maximum number of iterations is reached or the change in the optimal validation set loss of the group within w consecutive generations is less than the convergence judgment threshold, the fractional cuckoo search algorithm terminates the iteration;

[0116] S36. Dynamically adjust the population size based on the cross entropy loss decrease rate of the current generation group:

[0117] m (t+1) =m (t) ×(1-δR t );

[0118] Among them, m (t+1) represents the population size of the t+1 generation, m (t) represents the size of the t-generation population, δ represents the population adjustment coefficient, R t Indicates the cross entropy loss decrease rate of the current generation group;

[0119] S37. After the cuckoo search termination condition is met, parameters that satisfy the cross entropy loss less than the loss screening threshold are selected from the optimized population to construct a high-quality parameter set;

[0120] S38. Execute an improved black hole optimization algorithm based on a high-quality parameter set.

[0121] The present invention generates an initial parameter population through normalized coding based on feature subsets, divides the population into a rapid exploration population and a fine mining population, and applies a fractional-order cuckoo search algorithm within each population for asynchronous adaptive step-size update. Combined with a dynamic population size adjustment strategy, efficient neural network initial parameter optimization is achieved. This mechanism improves the global and local balance of the search, greatly accelerates the optimization convergence speed, and enhances the rationality of the initial structural parameters of the neural network and the upper limit of the model performance.

[0122] In this embodiment, the S38 specifically includes:

[0123] S381. Extract the hyperparameter part from the high-quality parameter set and select the first k groups of individuals with the smallest cross entropy loss as the multi-black hole centers;

[0124] S382. The remaining individuals in the high-quality parameter set except the black hole center are regarded as ordinary individuals. In the black hole optimization phase, each ordinary individual selects the nearest black hole based on the Euclidean distance and updates its position according to the adaptive attraction coefficient:

[0125]

[0126] Among them, p ′ i (t+1) represents the position of the ith common individual in the t+1th generation, p ′ i(t) represents the position of the i-th common individual in the t-th generation, represents the attraction coefficient of the i-th individual in the t-th generation, represents the jth black hole center individual of the tth generation;

[0127] The adaptive attraction coefficient is dynamically calculated and updated based on the distance between the current individual and the black hole:

[0128]

[0129] in, represents the attraction coefficient of the i-th individual in the t-th generation, ρ0 represents the initial attraction coefficient, κ represents the attraction contraction control coefficient, represents the distance from the ith individual of the tth generation to the nearest black hole, represents the maximum distance between individuals in the t-th generation population;

[0130] S383. During the black hole optimization iteration process, if the performance of the validation set after the normal individual position is updated is better than that of the corresponding black hole individual, the corresponding black hole position is replaced by the normal individual;

[0131] S384. When the black hole optimization reaches the maximum iteration number or meets the convergence condition, a black hole with the best verification performance is selected from the black hole set, and the initial network configuration parameters are output.

[0132] The present invention implements an improved black hole optimization algorithm based on high-quality parameters, adopts a multi-black hole mechanism and an adaptive attraction update strategy, improves the local fine-grained hyperparameter optimization capability, and enhances the local search accuracy and stability of the parameter space through mechanisms such as ordinary individuals dynamically approaching the nearest black hole and dynamically updating the black hole position. It further refines and supplements the initial coarse-grained optimization results, effectively improving the accuracy and overall performance of the final parameter configuration of the customer demand prediction neural network.

[0133] In this embodiment, the S4 specifically includes:

[0134] S41, extracting initial network configuration parameters, including the number of hidden layers, the number of nodes in each layer, the connection mode encoding, the learning rate, the regularization coefficient, the batch size, and the optimizer type encoding;

[0135] S42. Initializing a customer demand prediction model based on the extracted parameters. The customer demand prediction model adopts a hierarchical structure of an encoding module, a hidden representation module, and a decoding module. The connection relationship between the modules is determined by the connection mode code.

[0136] S43, the encoding module receives the input feature vector, performs linear transformation and activation operations, and obtains a preliminary hidden representation. The hidden representation module calculates the state of each hidden layer layer by layer, and the decoding module generates a potential customer demand prediction output based on the final hidden layer;

[0137] S44. After the structure of the customer demand prediction model is completed, the highly correlated feature subsets are divided according to the batch size, the initial learning rate is set, the optimizer and regularization coefficient are applied, and the compilation initialization of the customer demand prediction model is completed;

[0138] S45. To support dynamic topological changes during the training phase, a node growth mechanism is defined. When the local gradient norm of a layer is greater than the node growth trigger threshold during training, the number of newly added nodes is:

[0139]

[0140] Where, ΔN l Indicates the number of new nodes in the first layer, α g Indicates the node growth ratio coefficient, represents the second norm of the local gradient of the first layer, τ g Indicates the node growth trigger threshold;

[0141] S46. Synchronously define a node pruning mechanism. When the sparsity of the node output of a certain layer is higher than the pruning threshold, remove the low-activity nodes and update the node connection matrix. The sparsity is measured using the L1 norm.

[0142] Based on the initial network configuration parameters, the present invention constructs a customer demand prediction model composed of an encoding module, a hidden module and a decoding module. It supports dynamic changes in network topology and, through the node growth and node pruning mechanism, can adaptively adjust the network scale according to the local gradient norm and output sparsity during the training process, ensuring that the model controls complexity while maintaining prediction accuracy, reducing computing resource consumption, improving system operation efficiency, and enhancing the flexibility and scalability of the model to adapt to different data characteristics.

[0143] In this embodiment, the S5 specifically includes:

[0144] S51, initialize the training cycle counter, set it to zero, configure the initial learning rate, regularization coefficient and batch size parameters, and divide the standardized customer data set into multiple small batch data sets;

[0145] S52. For each training batch, extract an input feature vector from the small batch data set, input the feature vector into the customer demand prediction model, and generate a corresponding prediction output;

[0146] S53. Based on the predicted output and the actual label data, the batch loss value is calculated, the batch loss value is used as training feedback, and the back propagation process is performed to calculate the local gradient of the model parameter set;

[0147] S54. Apply the optimizer rule to locally update the model parameters based on the calculated local gradient to generate an updated parameter set after the current training batch;

[0148] S55. After completing all batch training, perform inference based on the validation set data using the currently updated model parameters to calculate the corresponding validation set overall loss value.

[0149] S56. When the decrease in the overall loss value of the validation set is less than the set threshold within a set number of consecutive training cycles, the swarm intelligence optimization module is started to execute the structure adaptive adjustment process;

[0150] S57, in the structural adaptation stage, based on the current latest parameter set, the optimization parameter population is initialized as the input of the swarm intelligence optimization process;

[0151] S58. In the swarm intelligence optimization process, the overall loss minimization of the verification set is used as the search goal, and the parameter population is iteratively searched and evolved to dynamically optimize the structure and parameter configuration of the customer demand prediction model.

[0152] The present invention sets up a dynamic training control mechanism during the training process, and through the linkage of periodic verification set evaluation and swarm intelligence optimization, intelligently judges and dynamically adjusts the customer demand prediction model structure. Compared with the traditional fixed structure and fixed training hyperparameters, the present invention can flexibly call the optimization module according to the progress of training, continuously evolve the model structure and parameter configuration, significantly improving the training convergence speed, model generalization ability and system stability, adapting to the ever-changing customer data characteristics, and improving prediction accuracy.

[0153] In this embodiment, the judgment condition of the preset convergence criterion is: if the decrease in the prediction error is less than the error convergence threshold, the training process is terminated.

[0154] In setting the training convergence criterion, this invention uses the reduction in potential customer demand forecasting error as the core indicator, sets a reasonable error convergence threshold, and intelligently determines when to terminate training. This mechanism avoids overfitting caused by excessive model training, ensures sufficient training, and improves the final performance stability of the model. This termination strategy improves training efficiency, reduces resource waste, and ensures that the final customer demand forecasting model has better generalization performance and application results in practical applications.

[0155] Example 1:

[0156] To verify the feasibility of the present invention in practice, the present invention was applied to the customer relationship management platform of an e-commerce enterprise. The platform is headquartered in Shanghai and mainly manages the consumption and interaction data of 2 million registered users nationwide. The platform originally used a traditional CRM system, which mainly relied on a rule engine and simple statistical features for customer classification and recommendation. This resulted in a high customer churn rate, a low personalized marketing conversion rate, and a high cost of manual intervention, which seriously restricted the enterprise's user stickiness and performance growth.

[0157] In actual applications, the historical customer data accumulated by the company is first collected and processed. The data covers three categories: structured data includes basic customer information, historical transaction records, and accumulated points; unstructured data includes communication texts between customers and customer service, social media comments, product browsing time and click records, and other behavioral trajectories. The data collection period is from January 2022 to June 2023, with a total of approximately 2.3TB of data, of which structured data accounts for approximately 40% and unstructured data accounts for approximately 60%. In order to achieve data standardization, unified field encoding, missing value filling, outlier removal, numerical normalization, One-Hot encoding and other operations are performed on structured data; for unstructured data, text cleaning, missing value supplementation, abnormal content removal and TF-IDF text vector encoding are performed, and finally a unified standard customer data set is generated for subsequent feature extraction.

[0158] In the feature extraction stage, the Embedding layer combined with the multi-layer perceptron was applied to extract deep behavioral features, including purchase cycle features, category preference features, interaction frequency features, and social topic interest features, through weighted evaluation of information gain and mutual information.

[0159] During the initial network parameter optimization process, a combined optimization strategy of the fractional-order cuckoo search algorithm and the improved black hole optimization algorithm was applied. First, parameters such as the number of hidden layers, the number of nodes in each layer, the connection sparsity, the learning rate, the regularization coefficient, the batch size, and the optimizer type were converted into a continuous optimization space through normalized encoding. The constructed customer demand prediction neural network model adopts an encoder-hidden-decoder structure, initially configured with four hidden layers, 512-256-128-64 nodes, and a sparse connection ratio of 30%. The initial learning rate was set to 0.001, and the regularization coefficient was 0.0005. During training, the model dynamically monitors the local gradient norm. When the gradient norm of a layer exceeds the set threshold of 0.05, the number of nodes in that layer is automatically increased. When the node output sparsity exceeds 0.7, node pruning is automatically performed to ensure dynamic evolution of the model structure and control complexity. Regarding training cycle control, the convergence termination condition is triggered when the validation set error decreases by less than 0.5% for five consecutive cycles, thus avoiding overfitting.

[0160] Table 1 Experimental data comparison table

[0161]

[0162]

[0163] According to the customer data collection and processing statistics in Table 1, the application scenarios of the present invention cover multi-source and multi-type customer data, with a total data volume of 2.3TB. Among them, structured data mainly comes from registration databases and order systems, accounting for about 40%, while unstructured data comes from customer service communication texts, social media content, and website browsing behavior logs, accounting for about 60%. Compared with the limitations of traditional CRM systems that mainly process structured information, the present invention uses a customized preprocessing process to standardize and unify processing for different data types, thereby improving the reliability of data fusion and subsequent modeling.

[0164] In terms of parameter optimization and model training indicator comparison, it can be clearly seen that the method of the present invention has achieved significant improvements in a number of key performance indicators. The accuracy of the optimized customer demand prediction neural network model on the validation set has increased from 75.2% of the original model to 92.4%, an increase of 17.2%, indicating that the fractional-order cuckoo search algorithm and improved black hole optimization algorithm adopted by the present invention have obvious advantages in the selection and optimization of initial network configuration parameters. At the same time, the customer retention period has been extended from 8.2 months to 10.6 months, an increase of 29.3%, which directly reflects the enhancing effect of accurate customer demand prediction on customer stickiness.

[0165] The optimized system of the present invention has improved its push response rate from 8.3% to 14.7%, and its conversion rate from 3.9% to 6.5%, an increase of 77.1% and 66.7% respectively. This fully demonstrates that through deep feature extraction and dynamic model structure adjustment, the present invention can effectively capture customers' potential interests and purchasing intentions, thereby improving the success rate of personalized marketing.

[0166] In terms of training efficiency, the average number of training rounds was reduced from 95 to 62, and the average training time was shortened from 14.5 hours to 9.2 hours, with the training speed increased by nearly 35%. This shows that the introduction of the swarm intelligence optimization mechanism not only improves model performance, but also has significant effects on training resource utilization and time cost control.

[0167] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A method for analyzing customer needs based on AI artificial intelligence CRM system data, characterized in that: The steps include: S1. Collect multi-source heterogeneous customer data in the CRM system, perform preprocessing on the collected customer data, fuse the preprocessed data, and output a standardized customer data set; S2. Based on the standardized customer dataset, we use Embedding and Multi-layer Perceptron to extract customer behavior features, fuse them into feature vector representations, and evaluate the feature importance of the feature vector representations to select highly relevant feature subsets. S3, input the high correlation feature subset into the swarm intelligence optimization module and output the initial network configuration parameters; S4. Build and initialize a customer demand prediction model based on the initial network configuration parameters; S5. Apply the backpropagation algorithm to perform local gradient updates on the customer demand forecasting model, evaluate the model's global performance indicators, and call the swarm intelligence optimization module based on the evaluation results to dynamically adjust the structure of the customer demand forecasting model. S6. Continue iteratively training the customer demand prediction model, with the goal of minimizing the potential customer demand prediction error, until a preset convergence criterion or a maximum number of training rounds is reached; S7. Input the new customer data collected in real time into the trained customer demand prediction neural network model to generate customer demand prediction results.

2. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1, characterized in that: The customer data includes structured data and unstructured data. The structured data includes basic customer information and transaction records, and the unstructured data includes customer communication texts, social media interaction content and browsing behavior records. The preprocessing of the structured data includes data format unification, missing value filling, outlier removal, numerical normalization and One-Hot encoding. The preprocessing of the unstructured data includes data cleaning, missing value filling, outlier removal and text encoding.

3. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1 is characterized in that: The customer behavior characteristics include historical purchase behavior characteristics, interactive behavior characteristics, interest preference characteristics and social behavior characteristics.

4. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1, characterized in that: The feature importance evaluation calculates the information gain and mutual information between feature vector representations, and takes the weighted sum of the information gain and mutual information as the importance score. The importance scores are sorted from high to low, and feature vector representations with importance scores greater than a threshold are screened out to form a high-correlation feature subset.

5. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1 is characterized in that: The swarm intelligence optimization module applies a fractional-order cuckoo search algorithm and an improved black hole optimization algorithm. The fractional-order cuckoo search algorithm performs a global search of network structure parameters, and the improved black hole optimization algorithm performs fine optimization of hyperparameters.

6. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1, characterized in that: The S3 specifically includes: S31, generating a feature subset matrix based on the high-correlation feature subset; S32, performing normalization coding on the neural network object to be optimized to generate an initial parameter population, wherein the neural network object to be optimized includes the number of hidden layers, the number of nodes in each layer, the connection mode coding, the learning rate, the regularization coefficient, the batch size, and the optimizer type coding; S33, randomly dividing the initial parameter population into a fast exploration group and a fine mining group according to a set ratio; S34. Define the performance evaluation index based on the feature subset matrix using the cross entropy loss function; S35. Apply the fractional-order cuckoo search algorithm within the rapid exploration group and the fine mining group, respectively, and update the individual step size according to the asynchronous adaptive mechanism: in, represents the step length of the i-th individual in the t-th generation, γ0 represents the initial step length factor, β represents the step length attenuation coefficient, r i represents a random number in the interval (0,1), and exp represents the natural exponential function; To perform an individual location update: in, represents the position of the i-th individual in the t+1 generation, represents the position of the i-th individual in the t-th generation, represents the step length of the i-th individual in the t-th generation, Levy(λ) represents the Levy flight step length vector, λ represents the stability index parameter of the Levy flight distribution, μ represents the fractional order influence weight, H represents the historical step length window, ω h represents the fractional weight coefficient, represents the position of the i-th individual in the th generation, represents the position of the i-th individual in the th-1 generation; When the maximum number of iterations is reached or the change in the optimal validation set loss of the group within w consecutive generations is less than the convergence judgment threshold, the fractional cuckoo search algorithm terminates the iteration; S36. Dynamically adjust the population size based on the cross entropy loss decrease rate of the current generation population; S37. After the cuckoo search termination condition is met, parameters that satisfy the cross entropy loss less than the loss screening threshold are selected from the optimized population to construct a high-quality parameter set; S38. Execute an improved black hole optimization algorithm based on a high-quality parameter set.

7. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 6 is characterized in that: The S38 specifically includes: S381. Extract the hyperparameter part from the high-quality parameter set and select the first k groups of individuals with the smallest cross entropy loss as the multi-black hole centers; S382. The remaining individuals in the high-quality parameter set except the black hole center are regarded as ordinary individuals. In the black hole optimization phase, each ordinary individual selects the nearest black hole based on the Euclidean distance and updates its position according to the adaptive attraction coefficient: in, represents the position of the i-th common individual in the t+1 generation, represents the position of the i-th common individual in the t-th generation, represents the attraction coefficient of the i-th individual in the t-th generation, represents the jth black hole center individual of the tth generation; The adaptive attraction coefficient is dynamically calculated and updated based on the distance between the current individual and the black hole: in, represents the attraction coefficient of the i-th individual in the t-th generation, ρ0 represents the initial attraction coefficient, κ represents the attraction contraction control coefficient, represents the distance from the ith individual of the tth generation to the nearest black hole, represents the maximum distance between individuals in the t-th generation population; S383. During the black hole optimization iteration process, if the performance of the validation set after the normal individual position is updated is better than that of the corresponding black hole individual, the corresponding black hole position is replaced by the normal individual; S384. When the black hole optimization reaches the maximum iteration number or meets the convergence condition, a black hole with the best verification performance is selected from the black hole set, and the initial network configuration parameters are output.

8. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1 is characterized in that: The S4 specifically includes: S41, extracting initial network configuration parameters, including the number of hidden layers, the number of nodes in each layer, the connection mode encoding, the learning rate, the regularization coefficient, the batch size, and the optimizer type encoding; S42. Initializing a customer demand prediction model based on the extracted parameters. The customer demand prediction model adopts a hierarchical structure of an encoding module, a hidden representation module, and a decoding module. The connection relationship between the modules is determined by the connection mode code. S43, the encoding module receives the input feature vector, performs linear transformation and activation operations, and obtains a preliminary hidden representation. The hidden representation module calculates the state of each hidden layer layer by layer, and the decoding module generates a potential customer demand prediction output based on the final hidden layer; S44. After the structure of the customer demand prediction model is completed, the highly correlated feature subsets are divided according to the batch size, the initial learning rate is set, the optimizer and regularization coefficient are applied, and the compilation initialization of the customer demand prediction model is completed; S45. To support dynamic topological changes during the training phase, a node growth mechanism is defined. When the local gradient norm of a layer is greater than the node growth trigger threshold during training, the number of newly added nodes is: Where, ΔN l Indicates the number of new nodes in the first layer, α g Indicates the node growth ratio coefficient, represents the second norm of the local gradient of the first layer, τ g Indicates the node growth trigger threshold; S46. Synchronously define a node pruning mechanism. When the sparsity of the node output of a certain layer is higher than the pruning threshold, remove the low-activity nodes and update the node connection matrix. The sparsity is measured using the L1 norm.

9. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1, characterized in that: The S5 specifically includes: S51, initialize the training cycle counter, set it to zero, configure the initial learning rate, regularization coefficient and batch size parameters, and divide the standardized customer data set into multiple small batch data sets; S52. For each training batch, extract an input feature vector from the small batch data set, input the feature vector into the customer demand prediction model, and generate a corresponding prediction output; S53. Based on the predicted output and the actual label data, the batch loss value is calculated, the batch loss value is used as training feedback, and the back propagation process is performed to calculate the local gradient of the model parameter set; S54. Apply the optimizer rule to locally update the model parameters based on the calculated local gradient to generate an updated parameter set after the current training batch; S55. After completing all batch training, perform inference based on the validation set data using the currently updated model parameters to calculate the corresponding validation set overall loss value. S56. When the decrease in the overall loss value of the validation set is less than the set threshold within a set number of consecutive training cycles, the swarm intelligence optimization module is started to execute the structure adaptive adjustment process; S57, in the structural adaptation stage, based on the current latest parameter set, the optimization parameter population is initialized as the input of the swarm intelligence optimization process; S58. In the swarm intelligence optimization process, the overall loss minimization of the verification set is used as the search goal, and the parameter population is iteratively searched and evolved to dynamically optimize the structure and parameter configuration of the customer demand prediction model.

10. The method for analyzing customer needs based on AI artificial intelligence CRM system data according to claim 1, characterized in that: The judgment condition of the preset convergence criterion is: if the decrease in the prediction error is less than the error convergence threshold, the training process is terminated.

Images