A method and apparatus for mesh refinement layering

Abstract

The application discloses a kind of grid refinement layering method and device, it is related to big data and AI new generation information technology field, the identification of the problem such as insufficient degree that fine operation of marketing system brings to business operation proposes the present scheme, including data acquisition module, grid refinement label module, user operation module and data storage module;Data acquisition module is used to collect data and carry out preliminary pretreatment;Grid refinement label module is based on the data collected by data acquisition module and carries out grid division;User operation module manages data and carries out data analysis;Data storage module is used to store data.The advantage is that different promotion marketing strategies are accurately formulated according to the demand of different grid unit groups and the contribution to enterprise, so as to reduce the sales cost of enterprise, improve product management efficiency and the competitive hard strength of enterprise.

Description

Technical Field

[0001] This invention relates to the field of big data and AI next-generation information technology, and in particular to a method and apparatus for fine-grained grid layering. Background Technology

[0002] As marketing systems have become more sophisticated, basic marketing management units have been refined to the grid level. This increased focus on refined operations and management has brought two main challenges to business management:

[0003] 1. The grid is not categorized according to business and attributes, making it difficult to provide intelligent management;

[0004] 2. When enterprises divide grid units, they often do so merely according to simple rules such as regions and areas, and then simply categorize the value generated by the grid into high, medium, and low levels. This results in a crude classification, where subjective stratification lacks scientific basis and reduces accuracy. The simplistic division of grid units leads to insufficient identification. Consequently, when enterprises implement refined strategies for their grid units, they cannot achieve targeting and differentiation, resulting in poor marketing effectiveness. Summary of the Invention

[0005] The purpose of this invention is to provide a method and apparatus for fine-grained mesh layering to solve the problems existing in the prior art.

[0006] The device for fine-grained grid layering according to the present invention includes a data acquisition module, a fine-grained grid labeling module, a user operation module, and a data storage module;

[0007] The data acquisition module is used to acquire data and perform preliminary data preprocessing.

[0008] The grid refinement labeling module divides the grid based on the data collected by the data acquisition module;

[0009] The user operations module manages data and performs data marketing and analysis;

[0010] The data storage module is used to store data.

[0011] The data acquisition module includes a data reporting unit and a data feature preprocessing unit. The data reporting unit cleans, transforms, and processes the data through a data request interface and stores the data in a basic database. The data feature preprocessing unit includes data integration, missing value processing, and outlier processing.

[0012] The grid-based refined tagging module includes a grid user basic tagging unit, a user churn warning tagging unit, and a grid hierarchical model tagging unit. The grid user basic tagging unit constructs behavioral tags according to product dimensions. The user churn warning tagging unit is used to predict whether users will churn from the platform and construct warning tags accordingly. The grid hierarchical model tagging unit constructs a grid hierarchical model based on a clustering algorithm.

[0013] The user operation module includes a grid user extraction unit, a recall management unit, and a grid operation management unit. The grid user extraction unit is used to extract users who meet the criteria. The recall management unit recalls users who are about to churn or have already churned. The grid operation management unit includes user behavior management and marketing activity analysis by operations personnel.

[0014] The data storage module includes, but is not limited to, tag storage units.

[0015] Behavioral tags include: registration behavior tags, activity behavior tags, and consumption behavior tags;

[0016] The process of predicting whether users will churn from the platform includes: extracting feature datasets, feature processing, constructing a fusion prediction model, generating a list of churned users, and storing it in a tag database.

[0017] The features include, but are not limited to: grid user age, user online time, grid cell, and customer star rating.

[0018] The aforementioned layering device is used to perform fine-grained mesh layering;

[0019] The feature processing method includes feature construction and feature selection; the feature construction involves encoding n category values ​​as integers between 0 and n-1 to establish a one-to-one mapping relationship, or transforming the features of each possible class value into a binary feature vector, where the feature vector of each class has only one 1 and the rest are 0; the feature selection is based on predictive ability and / or correlation with the target value.

[0020] The construction of the fusion prediction model includes the following steps: using a single-model learning algorithm to construct a single-model learning dataset; using the result of the single-model learning algorithm as the input to the fusion learning model; and constructing the fusion learning model.

[0021] The single-model learning algorithms mentioned include, but are not limited to, support vector machines, random forests, logistic regression, and neural networks.

[0022] The construction of the mesh layered model includes the following steps:

[0023] S1: Start the centroid initialization sub-unit and randomly select a sample from the dataset as the initial cluster center;

[0024] S2: The clustering operation subunit calculates each sample x. i The shortest distance D(x) between the cluster center and the existing cluster center i );

[0025] S3: Calculate the probability that each sample will be selected as the next cluster center by calling the cluster center generation function.

[0026]

[0027] Calculate the cumulative probability value q for each sample. i

[0028]

[0029] For a pseudo-random array r generated at time [0,1], if q i If the sample x is greater than element r[i] in the array, then the sample x i If selected, then compare with the next one, and so on, until K cluster centers are selected;

[0030] S4: Call the clustering operation subunit to calculate x for each sample. i Calculate the distance between the current K cluster centers and assign it to the category corresponding to the cluster center with the shortest distance;

[0031] S5: Call the centroid update sub-unit, for each category.

[0032]

[0033] Recalculate its cluster centers c i ;

[0034] S6: Call the centroid comparison sub-unit, if c i If a change occurs, stop; otherwise, repeat steps S4 and S5 until the cluster center c is reached. i No longer changing;

[0035] S7: Output grid layer labels.

[0036] The tag storage unit uses a compressed bitmap storage method for storage; the 32-bit range of an integer is divided into 16-bit data blocks and 16-bit containers, each data block corresponds to the high 16 bits of the integer, and a container is used to store the low 16 bits of a value.

[0037] Two containers are used for storage: an array container and a bitmap container. The array container is used by default. When the capacity of the array container exceeds 4096 short integers, the stored data is automatically stored in the bitmap container.

[0038] The method and apparatus for fine-grained grid layering described in this invention have the advantages of providing analysis for operational decision-making and prediction with data mining as the core, solving the problem of management timeliness for achieving a "flat" management approach, and meeting the purpose of displaying and tracking the completion status of grid clustering functions.

[0039] By combining cluster analysis to identify key grid cells, and precisely formulating different marketing strategies based on the needs of different grid cell groups and their contribution to the enterprise, the company's sales costs are significantly reduced, broadband operation efficiency is improved, and the company's competitive strength is enhanced. The value and risk segmentation capabilities of grid cells are optimized by calculating the similarity between grid cells, which improves identification accuracy compared to traditional methods, reduces the influence of subjective factors, and ensures accuracy. The compressed bitmap storage method can significantly improve data storage efficiency. Attached Figure Description

[0040] Figure 1 This is a schematic diagram of a device for fine-grained mesh layering according to the present invention;

[0041] Figure 2 This is a schematic diagram of the fusion prediction model structure;

[0042] Figure 3 This is a schematic diagram of a mesh layered model structure. Detailed Implementation

[0043] A device for fine-grained grid layering includes a data acquisition module, a fine-grained grid labeling module, a user operation module, and a data storage module.

[0044] The data acquisition module is used to collect data and perform preliminary data preprocessing.

[0045] The grid refinement labeling module divides the grid based on the data collected by the data acquisition module.

[0046] The user operations module manages data and performs data marketing and analysis.

[0047] The data storage module is used to store data.

[0048] The data acquisition module includes a data reporting unit and a data feature preprocessing unit. The data reporting unit cleans, transforms, and processes the data through a data request interface and stores the data in a basic database. The data feature preprocessing unit includes data integration, missing value handling, and outlier handling. Data integration involves merging the datasets obtained in the previous step and processing redundant features such as homonyms and synonyms. Missing value handling methods include using the mean to replace continuous, skewed normally distributed feature data to maintain the mean; using the median to replace feature data with a long-tailed distribution to avoid the influence of outlier values; and filling discrete feature data with the highest frequency. Outlier handling includes deletion, average imputation, and imputation based on missing value model predictions.

[0049] The grid-based refined tagging module includes a grid user basic tag unit, a user churn warning tag unit, and a grid hierarchical model tag unit. The grid user basic tag unit constructs behavioral tags according to product dimensions. These behavioral tags include: registration behavior tags, activity behavior tags, and consumption behavior tags. Registration behavior tags include information such as user registration time, registration duration, registration device, registration location, city, and city level tag. Activity behavior tags include user activity index and user activity time periods. The user activity index p is used to measure user activity.

[0050]

[0051] Where, x i w is a metric for user activity over a period of N days. i This represents the weight value for the corresponding activity metric, and In this embodiment, m = 3, where x1 represents the number of offline days within the period, x2 represents the number of active days within the period, and x3 represents the difference between the last active date and the first active date within the period. The user activity index is measured by the following tags:

[0052]

[0053] Among them, 1 represents low activity, 2 represents moderate activity, 3 represents medium-high activity, and 4 represents high activity.

[0054] The consumption tag is a tag that evaluates a user's sensitivity to marketing activities by combining the ratio of the number of times a user participates in the activity to the total number of times they make a purchase.

[0055] The user churn warning tagging unit is used to predict whether users will churn from the platform and to generate warning tags accordingly. The purpose of this unit is to enhance user engagement and extend user lifecycle for users predicted to churn through operational measures.

[0056] The data storage module includes a MySQL database, a MariaDB database, a MongoDB database, or an HBase database. It also includes tag storage units, user behavior datasets, and tag datasets. Specifically, the user behavior dataset includes, but is not limited to, features such as grid user age and user online time.

[0057] The tag storage unit uses a compressed bitmap storage method. This method significantly improves query efficiency and storage and computation performance by employing a compressed bitmap algorithm to store tags. The 32-bit range of an integer is divided into 16-bit data blocks and 16-bit containers. Each data block corresponds to the high 16 bits of the integer, while a container stores the low 16 bits. Two containers are used for storage: an array container and a bitmap container. Each element in the array container is a two-byte short integer. Elements are arranged in descending order of memory size. The array container is used by default. When the array container's capacity exceeds 4096 short integers, the stored data is automatically moved to the bitmap container. This compressed bitmap storage method allows for fast retrieval of a specific value and minimizes memory waste. The features include, but are not limited to: grid user age, user's online time, grid cell, and customer star rating.

[0058] The process of predicting whether users will churn from the platform includes: extracting a feature dataset, feature processing, constructing a fusion prediction model, and generating a list of churned users and storing it in a tag database. The extracted feature dataset includes a sample dataset obtained by combining the aforementioned user behavior dataset and tag dataset, and associating them through unique user identifiers.

[0059] The feature processing method includes feature construction and feature selection. Feature construction involves encoding n category values ​​as integers between 0 and n-1 for discrete features, establishing a one-to-one mapping relationship, or transforming the feature vector of each possible class value into a binary feature vector, where each class's feature vector has only one 1 and all other positions are 0. For numerical features, processing methods include standardization, binarization, normalization, and discretization. In this embodiment, z-score standardization is used, ensuring that the processed data has a mean of 0 and a standard deviation of 1. Specifically:

[0060]

[0061] in, Let σ be the mean of feature x, and σ be the standard deviation of feature x.

[0062] For features with a high proportion of default values, they are either discarded or padded. For example, samples with a high proportion of default values ​​are discarded, qualitative features are padded with 0, and quantitative feature values ​​are padded with their mean. In particular, for feature values ​​with uneven numerical distribution, they are divided into intervals based on the effective information they contain. In this embodiment, for example, if the call duration over the past N days only concerns whether the online duration has reached a defined interval value, it is processed as 0 and 1 to represent not reaching the threshold and having reached the threshold, respectively.

[0063] Feature selection is based on the predictive power of features and / or their relevance to the target value. The stronger the predictive power of a feature, the higher its information contribution. Selection based on relevance to the target value specifically involves using machine learning methods to train and obtain the weight coefficients of each feature, removing features with lower weight coefficients.

[0064] The construction of the fusion prediction model includes the following steps: using a single-model learning algorithm to construct a single-model learning dataset; using the result of the single-model learning algorithm as the input to the fusion learning model; and constructing the fusion learning model.

[0065] Constructing a single-model learning model involves randomly dividing the dataset D into k uniformly sized sets {D1, D2, ..., D...}. k Each time, k-1 elements are selected as the training set, and the remaining set is used as the test set, thus obtaining K sets of training and test sets. j and d′ j Let each represent the j-th set D. j The training and test sets are partitioned from the given set. Among the known T primary learning algorithms, i.e., in algorithm {ξ1,ξ2,…,ξ}... T In}, the primary learner model This is the model obtained using the t-th learning algorithm. Secondly, for d... j Each sample x in i Assume the prediction result of the t-th model is Then, based on sample x i The generated secondary model training sample features are Z i =(Z i1 Z i2 ,…,Z iT The sample classification label is still the original label y. i The algorithm {ξ1,ξ2,…,ξ} T This includes, but is not limited to, support vector machines, random forests, logistic regression, and neural networks. Based on actual prediction results, some datasets with poor performance are discarded.

[0066] After the above k*T training and prediction operations, a secondary training set d′ is obtained. d′ is the dataset used to train the secondary model, and the secondary model y′ is Z1, Z2, ..., Z.T Functions related to y. It also includes training and predicting user churn probability using the LightGBM algorithm, and using the probability values ​​output by the model as the final output probability of the fusion layer. When the average output probability is greater than a set threshold, the user is considered a churned user.

[0067] The label unit of the grid hierarchical model is constructed based on a clustering algorithm to build the grid hierarchical model.

[0068] The user operation module includes a grid user extraction unit, a recall management unit, and a grid operation management unit. The grid user extraction unit is used to extract users who meet certain criteria. In this embodiment, users with tags indicating a network duration of less than or equal to 3 months, outstanding payments, and inactivity during the current month are extracted. These are typical high-risk new users, and the strategy must focus on retaining them. The recall management unit recalls users who are about to churn or have already churned. This module mainly uses grid user basic tags and user churn warning tags to recall churned users and users who are expected to churn in the near future. Specifically:

[0069] (1) The module obtains information on users who have churned and users who are about to churn, and notifies the operations staff;

[0070] (2) The grid operation personnel formulate recall strategies based on the profile information of the corresponding users. The recall strategies include, but are not limited to, SMS notifications, tariff discounts, and follow-up communication.

[0071] (3) Users who are successfully awakened will be included in the operation monitoring module and further maintained by the operation personnel.

[0072] The construction of the mesh layered model includes the following steps:

[0073] S1: Start the centroid initialization sub-unit and randomly select a sample from the dataset as the initial cluster center;

[0074] S2: The clustering operation subunit calculates each sample x. i The shortest distance D(x) between the cluster center and the existing cluster center i );

[0075] S3: Calculate the probability that each sample will be selected as the next cluster center by calling the cluster center generation function.

[0076]

[0077] Calculate the cumulative probability value q for each sample. i

[0078]

[0079] For a pseudo-random array r generated at time [0,1], if q i If the sample x is greater than element r[i] in the array, then the sample x i If selected, then compare with the next one, and so on, until K cluster centers are selected;

[0080] S4: Call the clustering operation subunit to calculate x for each sample. i Calculate the distance between the current K cluster centers and assign it to the category corresponding to the cluster center with the shortest distance;

[0081] S5: Call the centroid update sub-unit, for each category.

[0082]

[0083] Recalculate its cluster centers c i ;

[0084] S6: Call the centroid comparison sub-unit, if c i If a change occurs, stop; otherwise, repeat steps S4 and S5 until the cluster center c is reached. i No longer changing;

[0085] S7: Output grid layer labels.

[0086] The grid operation management unit includes user behavior management and marketing activity analysis for operations personnel. User behavior management is mainly divided into user behavior management and activity management, primarily managing users already included in marketing strategies. For example, regarding the reactivation of churned users, this module will statistically analyze and display the behavioral events of this group of users, including call duration, number of calls, recharge amounts, and plan usage. The marketing activity analysis for operations personnel fully showcases the work content and effectiveness of operations personnel, such as the number of user inquiries followed up, the number of users followed up, the number of users reactivated, and the reactivation ratio.

[0087] In this embodiment, three clustering sub-models are established: customer revenue, device dismantling status, and penetration rate. Finally, the three sub-models are cross-analyzed to identify high-value, high-risk and high-value, high-spatial-value grid cells, which are then labeled accordingly, providing strong support for subsequent precision marketing. The final output results are shown in Table 1.

[0088] Table 1

[0089]

[0090] This method enables grid digitization, safeguarding high-value grids and seizing high-space opportunities.

[0091] For those skilled in the art, various other corresponding changes and modifications can be made based on the technical solutions and concepts described above, and all such changes and modifications should fall within the protection scope of the claims of this invention.

Claims

1. A device for finely layering a grid, characterized in that, It includes a data acquisition module, a grid-based refined tagging module, a user operation module, and a data storage module; The data acquisition module is used to acquire data and perform preliminary data preprocessing. The grid refinement labeling module divides the grid based on the data collected by the data acquisition module; The user operations module manages data and performs data marketing and analysis; The data storage module is used to store data; The data acquisition module includes a data reporting unit and a data feature preprocessing unit. The data reporting unit cleans, transforms, and processes the data through a data request interface and stores the data in a basic database. The data feature preprocessing unit includes data integration, missing value processing, and outlier processing. The grid-based refined tagging module includes a grid user basic tagging unit, a user churn warning tagging unit, and a grid hierarchical model tagging unit. The grid user basic tagging unit constructs behavioral tags according to product dimensions. The user churn warning tagging unit is used to predict whether users will churn from the platform and construct warning tags accordingly. The grid hierarchical model tagging unit constructs a grid hierarchical model based on a clustering algorithm. The user operation module includes a grid user extraction unit, a recall management unit, and a grid operation management unit. The grid user extraction unit is used to extract users who meet the criteria. The recall management unit recalls users who are about to churn or have already churned. The grid operation management unit includes user behavior management and marketing activity analysis by operations personnel. Data storage modules include, but are not limited to, tag storage units; Behavioral tags include: registration behavior tags, activity behavior tags, and consumption behavior tags; The prediction of whether users will churn from the platform includes: extracting feature datasets, feature processing, constructing a fusion prediction model, generating a list of churned users, and storing it in a tag database; The construction of the mesh layered model includes the following steps: S1: Start the centroid initialization sub-unit and randomly select a sample from the dataset as the initial cluster center; S2: The clustering operation subunit will calculate each sample Shortest distance to existing cluster centers ; S3: Calculate the probability that each sample will be selected as the next cluster center by calling the cluster center generation function. ； Calculate the cumulative probability value for each sample. ； For pseudo-random arrays generated at time [0,1] ,like Greater than the elements in the array Then the sample If selected, then compare with the next one, and so on, until K cluster centers are selected; S4: Call the clustering operation subunit to calculate the clustering results for each sample. Calculate the distance between the current K cluster centers and assign it to the category corresponding to the cluster center with the shortest distance; S5: Call the centroid update sub-unit, for each category. ； Recalculate its cluster centers ; S6: Call the centroid comparison sub-unit, if If changes occur, stop; otherwise, repeat steps S4 and S5 until the cluster centers are reached. No longer changing; S7: Output grid layer labels.

2. The device for fine-grained mesh layering according to claim 1, characterized in that, The features include, but are not limited to: grid user age, user online time, grid cell, and customer star rating.

3. A method for fine-grained mesh layering, characterized in that, Mesh refinement layering is performed using the layering device as described in any one of claims 1-2; The feature processing includes feature construction and feature selection; the feature construction involves encoding n category values ​​as integers between 0 and n-1 to establish a one-to-one mapping relationship or transforming the features of each possible class value into a binary feature vector, where the feature vector of each class has only one 1 and the rest are 0; the feature selection is based on predictive ability and / or correlation with the target value.

4. The method for fine-grained mesh layering according to claim 3, characterized in that, The construction of the fusion prediction model includes the following steps: using a single-model learning algorithm to construct a single-model learning dataset; using the result of the single-model learning algorithm as the input to the fusion learning model; and constructing the fusion learning model.

5. The method for fine-grained mesh layering according to claim 4, characterized in that, The single-model learning algorithms mentioned include, but are not limited to, support vector machines, random forests, logistic regression, and neural networks.

6. The method for fine-grained mesh layering according to claim 3, characterized in that, The tag storage unit uses a compressed bitmap storage method for storage; the 32-bit range of an integer is divided into 16-bit data blocks and 16-bit containers, each data block corresponds to the high 16 bits of the integer, and a container is used to store the low 16 bits of a value.

7. The method for fine-grained mesh layering according to claim 6, characterized in that, Two containers are used for storage: an array container and a bitmap container. The array container is used by default. When the capacity of the array container exceeds 4096 short integers, the stored data is automatically stored in the bitmap container.

Images