A high-precision customer profiling method for logistics enterprises based on big data technology

Abstract

A high-precision logistics enterprise customer profiling method based on big data technology, which combines database information and multi-source heterogeneous crawlers to collect user data and pre-process the original data; based on the collected user data, it discovers implicit features, uses a density clustering algorithm to classify user groups, scores and classifies the classified users according to the acquired implicit features, and then divides the users into three categories according to the results; uses a convolutional neural network optimized by a salp swarm algorithm to classify user packages, and uses activation functions and modifies the network according to the scores of network filters to classify packages according to three dimensions and construct feature dimensions of user packages; completes user data labeling based on user packages and personal information, and generates high-precision customer portraits; from the perspective of logistics companies, it profiles logistics customers in combination with user packages to provide logistics companies with more accurate customer portraits.

Description

Technical Field

[0001] This invention belongs to the field of customer profiling technology, specifically relating to a high-precision customer profiling method for logistics companies based on big data technology. Background Technology

[0002] With the development of economic globalization and logistics informatization, my country's logistics companies have gradually expanded their service scope, shifting from early warehousing and transportation to comprehensive services. Simultaneously, corporate users have generated greater demand for logistics parks, undoubtedly bringing new challenges to the management and services of logistics companies. Having acquired a wealth of user business behavior information, logistics companies, in order to better serve users and develop corresponding precision marketing, business recommendations, and personalized service plans, need to build user profiles, analyze user business characteristics and preferences, optimize user management, formulate precise marketing methods and strategies, and proactively provide personalized services to users.

[0003] Building customer profiles is of great significance to enterprises. With a wealth of user business behavior information, and to better serve users, it's essential to develop targeted marketing, business recommendations, and personalized service plans. This requires creating user profiles for logistics companies, analyzing user business characteristics and preferences. Building customer profiles not only optimizes the user experience but also enables precise marketing. By analyzing and mining data such as each user's warehousing categories, storage volume, and processing volume, user profiles can be constructed. This data reveals the user's true needs, allowing for the design of services and products that better meet those needs, providing a better user experience, improving logistics efficiency, and creating greater economic value. Enterprises accumulate a large amount of business data in their daily operations. This data is relatively reliable, easy to obtain, and updated promptly, providing data support for building user profile models. By establishing data mining and analysis models based on this precise information, a complete user profile model system can be built, enabling a more effective and comprehensive understanding of users and achieving refined, intensive, and standardized marketing. Density clustering algorithms can identify clusters of arbitrary shapes and are effective in handling outlier data. Convolutional neural networks and conditional generative adversarial networks can better distinguish user information and more accurately fill in missing user information in text classification. This invention combines the above technologies and related improvements to create a more accurate customer profile for logistics companies by profiling logistics customers based on their packages. Summary of the Invention

[0004] The technical problem to be solved by this invention is to overcome the shortcomings of the existing technology and provide a high-precision customer profiling method for logistics enterprises based on big data technology. This method uses big data technology to process and extract user tags to complete a high-precision user profile. It involves collecting user data, preprocessing the collected raw data to mine implicit user characteristics, and dividing user groups through multi-dimensional combinations to achieve user classification. An improved density clustering algorithm is used to analyze the characteristics of each user category, summarizing user type tags. After selecting data tags, a high-precision user profile is generated, optimizing the enterprise's user experience while achieving precise marketing.

[0005] This invention provides a method for creating high-precision customer profiles for logistics companies based on big data technology, comprising the following steps:

[0006] Step S1. Collect and process user data. Combine existing information in the database with a multi-source heterogeneous crawler approach to collect user data. Preprocess the raw data and fill in missing data in the raw data using an adaptive conditional generative adversarial network. Delete erroneous data and duplicate data respectively.

[0007] Step S2. Discover latent features based on the collected user data, classify the user group using density clustering algorithm, score the classified users according to the acquired latent features, and then classify the users into advanced users, intermediate users, and general users based on the results.

[0008] Step S3. Use the convolutional neural network optimized by the tunic group algorithm to classify the user's packages. At the same time, use the activation function and modify the network according to the scores of the network filter to classify the packages according to three dimensions: weight, fragility, and urgency, and construct the feature dimensions of the user's packages.

[0009] Step S4. Based on the user's package and personal information, complete the user data tagging and generate a high-precision customer profile.

[0010] As a further technical solution of the present invention, a density index ω is introduced into the density clustering algorithm, the index being...

[0011]

[0012] Where M is the number of points in the dataset, K is the desired number of clusters, α ranges from 0 to 1, ω is the number of neighboring points used to calculate the local density of each point, and a threshold n is also used. c Therefore, the local density p i For, where d(x) i -x jk ) is x i and xjk The Euclidean distance between them, where j k It is the k-th point closest to point i;

[0013] Set the decision function γ to complete the clustering:

[0014]

[0015] Where, δ i The distance between samples is given by the data point x. i When the local density is maximized, δ i Take the maximum value, ensuring x i The definition of being selected as a cluster center

[0016]

[0017] Furthermore, the specific steps of the density clustering algorithm are as follows:

[0018] Step S21. Input user dataset S = {x1, x2, ..., xN}, use threshold nc and cluster number K to obtain neighboring points ω;

[0019] Step S22. Calculate the local density p of each sample point using a kernel-free method based on the obtained critical point ω. i ;

[0020] Step S23. Calculate the γ value for each sample point using the decision function γ;

[0021] Step S24. Calculate the distance δ between each selected sample point. i ;

[0022] Step S25. Sort the γ values ​​of each point in descending order and select the first k as cluster centers.

[0023] Step S26. Calculate the minimum distance from the non-cluster center to each center point. If the value is less than the threshold, the point is assigned to the same cluster; otherwise, the point is discarded. Process each point one by one, and finally determine the final assignment of each point to complete the clustering.

[0024] Furthermore, in step S1, the adaptive conditional generative adversarial network includes a generator, an auxiliary classifier, and a discriminator. The loss function of the generator is:

[0025]

[0026] The discriminator loss function is:

[0027]

[0028] Where m is the missing data index, m = 0 when the data is missing, otherwise m = 1; α and β are coefficients between 0 and 1, and C(x) is the output of the auxiliary classifier, (x R k -x k ) 2 To improve the conditional generative adversarial model, an auxiliary classifier C is added to improve the discriminator's satisfaction level. The classifier constrains the generator, a new loss function is obtained, and the discriminator and generator are trained again.

[0029] Furthermore, a weighted calculation method was used to calculate the missing rate of the sample data, increasing the weight of the top 5 data points. The calculation method is as follows:

[0030]

[0031] Where m is a loss flag, m=0 when data is lost, otherwise m=1.

[0032] Furthermore, the learning rate is adaptively adjusted using a loss function. Training is performed independently with several learning rates in each training epoch, and the learning rate that minimizes the training loss is used as the learning rate for each epoch. The relationship between the learning rates is as follows:

[0033] η t+1 =η t r n (5-1)

[0034] Where, r n The scaling factor is represented by a tree structure used to search through several state transitions to determine the learning rate η for different states. s The state transitions during training can be represented by a tree structure:

[0035]

[0036] Where η0 is the initial learning rate; the values ​​calculated on the nodes are sorted in descending order, and the labeled numbers represent the ranking of the training loss for each time. The state is moved to the next node marked as 1 using breadth-first search, which is then used as the learning rate at that moment.

[0037] Furthermore, the improved conditional generative adversarial network includes the following steps:

[0038] Step S11. First, initialize the model parameters, calculate the loss rate of sample data n and sort them in ascending order;

[0039] Step S12. Select the first λn data (0<λ<1) as preprocessed data and input them into the generator network for training to obtain sample data; train the discriminator by combining the generated sample data n1 with the real sample data n2, and update the parameters using an adaptive stochastic gradient descent method.

[0040] Step S13. Cluster the trained results using a clustering algorithm to obtain the data labels, and use these labeled data to train an auxiliary classifier;

[0041] Step S14. Continue training the generator and discriminator using the entire dataset, while using the pre-trained classifier to constrain the generator to meet the new objective function. Update the parameters again using stochastic gradient descent until the discriminator can no longer distinguish between the generator and the real data.

[0042] Step S15. Complete the training and obtain the improved conditional adversarial network.

[0043] Furthermore, the sea squirt swarm algorithm introduces a new computational factor, u, which is...

[0044]

[0045] p=λe -λt ;

[0046] Where p is the exponential probability value, and λ is the reciprocal of the average number of mutation operations;

[0047] The resulting final equation G(u) is:

[0048]

[0049] Since the control parameters c2 and c3 are uniformly distributed random numbers, by replacing one of the control parameters c3 with...

[0050]

[0051] Where c2 is a random number between [0,1], and ub and lb are the upper and lower bounds of the individual in this dimension, respectively.

[0052] Furthermore, the leader's update method is as follows:

[0053]

[0054] Let S be the j-th dimension component of the leader individual i after iteration. j Let ub be the j-th dimension component of the current elite individual. j and lb jc1 and c2 are the upper and lower bounds of the individual j-dimensional dimension, respectively, and c2 is a random number between [0,1].

[0055] The follower update method is

[0056]

[0057] Where β is a random number between 0 and 1. An individual X randomly selected from the current population i (t) The j-th dimension component, S j (t), They are all different, f i f is the fitness of the current point. avr This represents the average fitness.

[0058] Furthermore, the improved steps of the tunicate group algorithm are as follows:

[0059] Step S31. Initialize parameters: initial position of the individual tunicate, maximum number of iterations t. max Population size N, individual dimension D;

[0060] Step S32. Calculate the fitness value of all individuals and the average fitness value. At the same time, sort the individuals according to their fitness values ​​and select the current best individual as the elite individual.

[0061] Step S33. Establish a tunicate chain. Sort the individuals in the population in descending order according to their fitness values ​​to establish a tunicate group chain. The individuals in the first half are leaders and the individuals in the second half are followers. Calculate the tunicate index probability value.

[0062] Step S34. Update the position of the elite individual and select the individual with the best current fitness as the elite individual S(t);

[0063] Step S35. Update the leader position and randomly select a leader individual. Update the leader position based on the calculated factor.

[0064] Step S36. Update follower positions and randomly select individuals from the population. Update followers;

[0065] Step S37. Determine the optimization termination condition: if the number of iterations has reached t. max Output the position of the elite individual, which is the optimal control parameter; otherwise, return to step S32.

[0066] The advantages of this invention lie in its ability to process existing data from logistics companies using big data technology, extract user tags to create high-precision user profiles, and collect usable user data such as usernames, user types, user shipping batches, and user sending and receiving times. The collected raw data is then preprocessed to uncover latent user characteristics. User groups are segmented through a multi-dimensional approach to achieve user classification. An improved density clustering algorithm is used to analyze the characteristics of each user category, summarizing user type tags. After selecting data tags, a high-precision user profile is generated, optimizing the user experience for businesses while achieving precise marketing. Attached Figure Description

[0067] Figure 1 This is a schematic diagram of the method flow of the present invention. Detailed Implementation

[0068] Please see Figure 1 This embodiment provides a high-precision customer profiling method for logistics enterprises based on big data technology, comprising the following steps:

[0069] Step 1: Collect and process user data. User data is collected by combining existing information from the enterprise database with multi-source heterogeneous web crawling. Simultaneously, the raw data is preprocessed, and an adaptive conditional generative adversarial network is used to fill in missing data. Erroneous and duplicate data are appropriately deleted, transforming the data into usable data and improving the accuracy of modeling.

[0070] Step 2: Based on the collected user data, uncover implicit features such as sending frequency, receiving frequency, and average spending. Use an improved density clustering algorithm to classify the user group. Score the classified users according to the obtained implicit features, and then classify the users into advanced users, intermediate users, and general users based on the results.

[0071] Step 3: An improved convolutional neural network optimized using the improved tunic group algorithm is used to classify users' packages. A new activation function is employed, and the network is modified appropriately based on the scores of the network filters to enable better classification of user packages. Packages are classified according to three dimensions: weight, fragility, and urgency, constructing the feature dimensions of user packages.

[0072] Step 4: Combine user package information and personal details to complete user data tagging and generate a high-precision customer profile. User characteristic dimensions can be constructed based on user features, allowing for user segmentation through multiple dimensions. After selecting data tags, the generated profile can be displayed visually.

[0073] In traditional density clustering algorithms, a cutoff distance d needs to be manually defined. c This can easily affect the accuracy of clustering results. Therefore, a new density index ω is introduced into the improved density clustering algorithm to eliminate the cutoff distance d. c To improve the accuracy of clustering, a new density index ω is introduced as follows:

[0074]

[0075] Where M is the number of points in the dataset, K is the desired number of clusters, α ranges from 0 to 1 and is used to fine-tune the metric, ω represents the number of neighboring points used to calculate the local density of each point, and a threshold n is also used. c Make sure ω is meaningful.

[0076] Therefore, the local density p i It can be defined by a new, kernelless approach:

[0077]

[0078] in It is x i and The Euclidean distance between them, where j k It is the k-th point closest to point i. The local density p can be directly calculated using this equation. i Instead of needing an experience-based parameter d c This greatly reduces human intervention in clustering algorithms and significantly improves the accuracy of clustering methods.

[0079] At the same time, a new decision function γ was set to achieve high-precision and robust clustering:

[0080]

[0081] Where δ i Represents the distance between samples, when data point x i When the local density is maximized, δ i Taking the maximum value is sufficient to guarantee x i The definition of a cluster center is as follows:

[0082]

[0083] The improved density clustering algorithm steps are as follows:

[0084] 1) Input user dataset S = {x1, x2, ..., xN}, using a threshold nc and the number of clusters K. Obtain the nearest neighbor ω from equation (2-1);

[0085] 3) Calculate the local density p of each sample point using a new nucleus-free method (2-2) based on the calculated critical point. i ;

[0086] 4) Calculate the distance δ between each selected sample point according to equation (2-3). i ;

[0087] 5) Calculate the γ value for each sample point using the new decision function γ(2-4);

[0088] 6) Sort the γ values ​​of each point in descending order and select the first k as cluster centers.

[0089] 7) Calculate the minimum distance from the non-cluster center to each center point. If the value is less than the threshold, the point is assigned to the same cluster; otherwise, the point is discarded. Process each point one by one, and finally determine the final assignment of each point to complete the clustering.

[0090] This adaptive conditional generative adversarial network (GAN) further improves the quality of the GAN by leveraging the implicit category information in incomplete data. An auxiliary classifier and discriminator are designed to help the generator produce results closer to real data while preserving the original category information. A new control term is introduced into the original loss function of the algorithm. The generator's loss function is as follows:

[0091]

[0092] The discriminator loss function is:

[0093]

[0094] Where m is the missing data index, m = 0 when the data is missing, and m = 1 otherwise; α and β are coefficients between 0 and 1, C(x) represents the output of the auxiliary classifier, and (xRk-xk)² represents the satisfaction level of the discriminator. Adding an auxiliary classifier C improves the conditional generative adversarial model by using the classifier to constrain the generator, resulting in a new loss function. The discriminator and generator are then trained again to enhance the model accuracy of the conditional generative adversarial network.

[0095] A new weighted calculation method was used for the missing rate of the sample data. Under a fixed missing rate, the missing data for each sample in the dataset is different. To facilitate the subsequent calculation of express delivery user data tags, the weight of the first 5 data points was increased. The calculation method is as follows:

[0096]

[0097] Where m is the loss flag, m=0 when data is lost, otherwise m=1.

[0098] The learning rate is adaptively adjusted using a loss function. Training is performed independently with multiple learning rates in each training epoch, and the learning rate that yields the minimum training loss is used as the learning rate for each epoch. The relationship between the learning rates can be expressed as follows:

[0099] η t+1 =η t r n (5-1)

[0100] Where r n This represents the scaling factor. Through multiple state transitions, a tree structure is used for searching to determine the learning rate η for different states. s The state transitions during training can be represented by a tree structure:

[0101]

[0102] Where η0 is the initial learning rate. The values ​​calculated at the nodes are sorted in descending order, and the labeled numbers represent the ranking of the training loss at each time. The state is moved to the next node marked with 1 using breadth-first search, which is then used as the learning rate at that moment.

[0103] The improved conditional generative adversarial network has the following steps:

[0104] (1) First, initialize the parameters of the model, calculate the loss rate of sample data n (4-1) and sort them in ascending order.

[0105] (2) Select the first λn data (0<λ<1) as preprocessing data and input them into the generator network to train according to formula (3-1) to obtain sample data; train the discriminator by combining the generated sample data n1 with the real sample data n2, and update the parameters using the adaptive stochastic gradient descent method.

[0106] (3) Cluster the training results using a clustering algorithm to obtain the data labels, and use the labeled data to train an auxiliary classifier.

[0107] (4) Continue training the generator and discriminator using the entire dataset, while using the pre-trained classifier to constrain the generator to satisfy the new objective function (3-2), and use stochastic gradient descent to update the parameters again until the discriminator can no longer distinguish the difference between the generator and the real data.

[0108] (5) Complete the training to obtain an improved conditional adversarial network.

[0109] This improved tunic swarm algorithm introduces a new computational factor to update the leader's position, enhancing global search capabilities. The new computational factor u is as follows:

[0110]

[0111] p=λe λt (7-2)

[0112] Where p is the exponential probability value, and λ is the reciprocal of the average number of mutation operations.

[0113] The resulting final equation G(u) is:

[0114]

[0115] Furthermore, since the control parameters c2 and c3 are uniformly distributed random numbers, these parameters result in a long optimization convergence time. Therefore, the tunicate swarm algorithm is improved by replacing one of the control parameters c3 with the following formula:

[0116]

[0117] Where c2 is a random number between [0,1], and and are the upper and lower bounds of the individual in this dimension, respectively. By replacing another parameter in the original equation with the final equation and the calculation equation of c2, the number of parameters is reduced, making the search more targeted and significantly reducing the required time, while improving accuracy and precision.

[0118] The leader's update method is as follows:

[0119]

[0120] in Let S be the j-th dimension component of the leader individual i after iteration. j Let ub be the j-th dimension component of the current elite individual. j and lb j c1 and c2 are the upper and lower bounds of the individual's j-dimensional dimension, respectively, and c2 is a random number between [0,1].

[0121] The follower update method is as follows:

[0122]

[0123] Where β is a random number between 0 and 1 An individual X randomly selected from the current population i (t) The j-th dimension component, S j (t), They are all different, f i f is the fitness of the current point. avr This represents the average fitness.

[0124] The improved algorithm steps are as follows:

[0125] (1) Initialization parameters: initial position of the salps individual, maximum number of iterations t max Population size N, individual dimension D;

[0126] (2) Calculate the fitness value of all individuals and the average fitness value. At the same time, sort the individuals according to their fitness and select the current best individual as the elite individual.

[0127] (3) Establish a salps chain. Sort individuals in the population in descending order of fitness value to form a salps chain. The first half of the individuals are leaders, and the second half are followers. Calculate the salps index probability value.

[0128] (4) Update the position of the elite individual and select the individual with the best current fitness as the elite individual S(t);

[0129] (5) Update the leader position and randomly select a leader individual. Update the leader position according to formula (8-5) based on the calculated factors;

[0130] (6) Update the position of the follower and randomly select individuals from the population. Update the followers according to equation (8-6);

[0131] (7) Determine the termination condition for optimization. If the number of iterations has reached t. max Output the position of the elite individual, which is the optimal control parameter; otherwise, return to step (2).

[0132] An optimized convolutional neural network using an improved tunic group algorithm is used to classify customer packages more accurately, improving classification performance. An adjustment layer is introduced into the convolutional neural network so that the network weights can still adaptively change upon input data after training, enhancing the network's classification ability and robustness. The adjustment form of the weights after adding the adjustment layer is as follows:

[0133]

[0134] Where 'a' is the coefficient used to adjust the addition. It is the weight of the adjustment layer, w′ ij These are the weights before adjustment. The result obtained through the adjustment layer and the hidden layer is:

[0135]

[0136] Where z j Let a represent the input to the activation function f(·). j θ represents the output of the neuron in layer j. j This represents the bias vector of layer j.

[0137] A new activation function is proposed to improve upon the slow convergence speed of existing activation functions. The proposed activation function f(x) is as follows:

[0138]

[0139] The proposed activation function is combined with an existing S-function to avoid insufficient data information caused by the sparsity of customer data features. The combined function is as follows:

[0140]

[0141]

[0142] In the convolutional neural network constructed with the new function, the S-function is used after the convolutional layer, and the network structure with the newly designed f(x) as the activation function is used in the fully connected layer.

[0143] A new method has been added to evaluate the performance of convolutional filters based on features generated by pooling layers and to prune the network structure online. This prunes words in low-scoring filters to prevent excessive network growth that could lead to a decrease in learning convergence speed and capability. The pruning evaluation metric is defined as follows:

[0144]

[0145] Where i represents the index of the fully connected layer, m is the number of neurons in the output layer, and w ji w represents the weights from the fully connected layer to the output layer. ik denoted by , where is the weight of the fully connected layer, and n is the number of neurons in the fully connected layer.

[0146] Words in the low-scoring filter are removed, and the weights of surrounding neurons are updated to make the network model more accurate. The updated neuron parameters after pruning are as follows:

[0147]

[0148] Where v is the neuron with the smallest Euclidean distance from neuron s; k v and k s For the pruning parameter, Aw′ s Represents the adjusted parameters of the pruned neuron; w s w v The parameters for pruning neurons and neuron v before adjustment; w′ v These are the adjusted parameters for neuron v.

[0149] The steps for adjusting the initial weights and bias vectors of a convolutional neural network are as follows:

[0150] (1) Initialization parameters: Initialization parameters include the initial position of the salps individual and the maximum number of iterations t. max Population size N, individual dimension D, and initial weights ω and bias vector b of the improved convolutional neural network;

[0151] (2) Calculate the fitness value of all individuals and the average fitness value. At the same time, sort the individuals according to their fitness and select the current best individual as the elite individual.

[0152] (3) Establish a salps chain. Sort the individuals in the population in descending order according to their fitness values ​​to establish a salps group chain. The first half of the individuals are leaders and the second half are followers. Calculate the salps index probability value.

[0153] (4) Update the position of the elite individual and select the individual with the best current fitness as the elite individual S(t);

[0154] (5) Update the leader position and randomly select a leader individual. Update the leader position according to formula (7-5) based on the calculated factors;

[0155] (6) Check the feasibility of the new location, evaluate the new location, and update the memory;

[0156] (7) Determine whether the current number of iterations meets the condition. If not, return to step (2). Otherwise, terminate the iteration and output the current optimal solution ω and b.

[0157] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the specific embodiments described above. The specific embodiments and descriptions in the specification are merely for further illustrating the principles of the invention. Various changes and modifications can be made to the present invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the present invention as claimed. The scope of protection of the present invention is defined by the claims and their equivalents.

Claims

1. A high-precision logistics enterprise customer profiling method based on big data technology, characterized in that: The following steps are included: Step S1. Collect and process user data, collect user data by combining the existing information in the database and the comprehensive method of multi-source heterogeneous crawlers, pre-process the original data, fill the missing data in the original data through the adaptive conditional generative adversarial network, and delete the wrong data and duplicate data respectively; Step S2. Discover hidden features based on the collected user data, use density clustering algorithm to classify user groups, score and classify the classified users based on the acquired hidden features, and then classify the users into advanced users, intermediate users, and general users based on the obtained results; Step S3. Use the convolutional neural network optimized by the Salp Swarm algorithm to classify the user's packages, and use the activation function and modify the network according to the scores of the network filter to classify the packages according to the three dimensions of weight, fragility, and urgency, and construct the feature dimension of the user's package; Step S4. Complete user data labeling based on the user's package and user personal information to generate a high-precision customer portrait; A density index ω is introduced in the density clustering algorithm. The index is Where M is the number of points in the data set, K is the desired number of clusters, α ranges from 0 to 1, ω is the number of adjacent points used to calculate the local density of each point, and a threshold n is referenced. c , so the local density p i is, where d(x i -x jk ) is x i and x jk The Euclidean distance between k is the kth point closest to point i; Set the decision function γ to complete clustering: Among them, δ i is the distance between samples, when the data point x i When the local density is the largest, δ i Take the maximum value to ensure x i is selected as the cluster center and is defined as The specific steps of the density clustering algorithm are: Step S21. Input the user data set S = {x1, x2, ..., xN}, use the threshold nc, the number of clusters K, and obtain the adjacent point ω; Step S22. Calculate the local density p of each sample point using the obtained critical point ω in a kernel-free manner i ; Step S23. Calculate the γ value of each sample point through the decision function γ; Step S24. Calculate the distance δ of each sample point i ; Step S25. Arrange the γ values ​​of each point in descending order by using the descending order of each point, and select the first k as the cluster centers; Step S26. Calculate the minimum value of the distance from the non-cluster center point to each center point. If the value is less than the usage threshold, the point is assigned to the same cluster. Otherwise, the point is discarded. Each point is processed one by one, and finally the final affiliation of each point is determined to complete clustering.

2. According to claim 1, a high-precision logistics enterprise customer profiling method based on big data technology is characterized in that: In step S1, the adaptive conditional generative adversarial network includes a generator, an auxiliary classifier and a discriminator, and the loss function of the generator is: The discriminator loss function is: Where m is the data missing indicator. When the data is missing, m = 0, otherwise m = 1; β and β are coefficients between 0 and 1, C(x) is the output of the auxiliary classifier, (x R k -x k ) 2 is the satisfaction of the discriminator; an auxiliary classifier C is added to improve the conditional generative adversarial model, the generator is constrained by the classifier, a new loss function is obtained, and the discriminator and generator are trained again.

3. According to the high-precision logistics enterprise customer profiling method based on big data technology according to claim 1, it is characterized in that: The missing rate of sample data adopts a weighted calculation method to enhance the weight of the first 5 data. The calculation method is as follows: Wherein, m is a loss flag, when data is lost, m=0, otherwise m=1.

4. According to claim 2, a high-precision logistics enterprise customer profiling method based on big data technology is characterized in that: The learning rate is adaptively adjusted through the loss function. Training is performed independently at several learning rates in each training period, and the learning rate that obtains the minimum training loss is used as the learning rate for each period. The relationship between the learning rates is: or t+1 =the t r n (5-1), Among them, r n Represents the proportional factor, through several state transformations, using a tree structure to search and determine the learning rate η of different states s , the state transitions during training can be represented by a tree structure: Among them, η0 is the initial learning rate; the values ​​calculated on the nodes are sorted in descending order, the marked numbers represent the ranking of each training loss, and the state is moved to the next node marked as 1 through breadth-first selection as the learning rate at that moment.

5. According to claim 2, a high-precision logistics enterprise customer profiling method based on big data technology is characterized in that: The improved conditional generative adversarial network includes the following steps: Step S11. First, initialize the parameters of the model, calculate the loss rate of sample data n and arrange them in ascending order; Step S12. Select the first λn data (0<λ<1) as preprocessed data and input them into the generative network for training to obtain sample data; train the discriminator using the generated sample data n1 and the real sample data n2, and update the parameters using the adaptive stochastic gradient descent method; Step S13. Cluster the trained results by a clustering algorithm and obtain data labels, and train an auxiliary classifier by using the label data; Step S14. Continue to train the generator and discriminator using the entire data set, and use the pre-trained classifier to constrain the generator so that it satisfies the new objective function, and use the stochastic gradient descent method to update the parameters again until the discriminator cannot distinguish the difference between the generator and the real data; Step S15: Complete the training and obtain the improved conditional adversarial network.

6. According to the high-precision logistics enterprise customer profiling method based on big data technology as described in claim 1, it is also characterized in that: The salp swarm algorithm introduces a new calculation factor, the new calculation factor u is p=λe -λt ; Among them, p is the exponential probability value, and λ is the inverse of the average number of mutation operations; The resulting final equation G(u) is Since the control parameters c2 and c3 are uniformly distributed random numbers, by replacing one of the control parameters c3 with Among them, c2 is a random number between [0,1], ub and lb are the upper and lower bounds of the individual in this dimension respectively.

7. According to claim 6, a high-precision logistics enterprise customer profiling method based on big data technology is also characterized in that: The leader is updated as is the j-th dimension component of the leader individual i after iteration, S j is the j-th dimension component of the current elite individual, ub j and lb j are the upper and lower bounds of the dimension of individual j, respectively, and c2 is a random number between [0,1]; Followers are updated as Among them, β is a random number between 0 and 1. is an individual X randomly selected from the current population i (t) j-th dimension component, S j (t) Different from each other, i is the fitness of the current point, f avr is the average fitness.

8. According to claim 6, a high-precision logistics enterprise customer profiling method based on big data technology is characterized in that: The improved steps of the salp swarm algorithm are as follows: Step S31. Initialization parameters, initial position of individual salps, maximum number of iterations t max , population size N, individual dimension D; Step S32. Calculate the fitness values ​​of all individuals, and calculate the average value of the fitness, and sort the individuals according to the fitness, and select the current best individual as the elite individual; Step S33. Establish a salp chain, sort the individuals in the population in descending order according to their fitness values ​​to establish a salp group chain, with the first half of the individuals being leaders and the second half being followers, and calculate the probability value of the salp index; Step S34. Update the elite individual position and select the individual with the best current fitness as the elite individual S(t); Step S35. Update the leader position and randomly select the leader individual Update the leader position according to the calculation factor, Step S36. Update the follower position and randomly select individuals from the population Update followers; Step S37. Determine the optimization end condition. If the number of iterations has reached t max , output the elite individual position, i.e., the optimal control parameter; otherwise, return to step S32.

Images