High-intention latent guest mining method and system fusing ensemble learning and CNN deep learning

Abstract

The invention discloses a high-intention potential customer mining method and system fusing ensemble learning and CNN deep learning. The method comprises the following steps: acquiring original data of user ordering information, and processing and analyzing the original data; analyzing a business target, and performing feature engineering on the data to obtain a full-quantity feature set; respectively establishing a LightGBM ensemble learning model and a Random Forest ensemble learning model, obtaining evaluation indexes and total feature information output by the two ensemble learning models, and analyzing output results of the two ensemble learning models to obtain a final important feature set; establishing a CNN deep learning model for predicting the submersible guests; screening pseudo high potential users from historical data of point marketing; and filtering the pseudo high potential users by using the predicted potential guests output by the CNN deep learning model to obtain final potential guest information. According to the scheme, a large amount of data can be automatically processed, data assets are comprehensively mined and utilized, and the comprehensiveness and accuracy of decision making are improved.

Description

Technical Field

[0001] The present invention relates to the field of big data and AI technology, and in particular to a method and system for mining high-intent potential customers by integrating ensemble learning and CNN deep learning. Background Art

[0002] With the continuous development and iteration of telecom operators' products and technologies, various types of data are constantly accumulating and have become key production factors and strategic resources. At the same time, users' growing needs are also constantly generated. Telecom operators have developed more and more Internet-oriented and personalized business products. How to let users with needs know about the corresponding products can better provide users with satisfactory services. At present, the point marketing behavior of telecom operators is mainly based on the experience of business experts to segment indicators, or to screen users through stock tags, or to directly recommend all users, and then push the configured product information. This model has a very low utilization rate of various types of data information, and there are problems such as inaccurate positioning of target user groups, high computing costs and channel costs, and low marketing conversion rates. Therefore, making full use of various types of user data information to mine potential customers and realize accurate business operations has become an urgent problem to be solved. Summary of the invention

[0003] The purpose of the present invention is to provide a method and system for mining high-intent potential customers by integrating ensemble learning and CNN deep learning. To achieve the above purpose, the technical solution provided by the present invention is:

[0004] The first aspect of the present application provides a method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning, comprising the following steps:

[0005] Obtain the original data of user ordering information, and perform pre-processing and multi-dimensional analysis;

[0006] Analyze business goals, perform feature engineering on preprocessed and multi-dimensionally analyzed data to obtain a full feature set, and build a training set for the ensemble learning model based on the full feature set;

[0007] A LightGBM ensemble learning model and a Random Forest ensemble learning model were established respectively. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models were obtained. The output results of the two ensemble learning models were analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input were compared and verified. Based on this, the features were pre-processed to obtain the final set of important features. The training set and test set of the CNN deep learning model were constructed based on the final set of important features.

[0008] Establish a CNN deep learning model, input the training set of the CNN deep learning model to train the model and adjust the parameters, and verify it, and predict potential customers by inputting the test set;

[0009] Obtain historical data of point marketing, analyze and process it, and screen pseudo high-potential users; use the predicted potential customers output by the CNN deep learning model to filter pseudo high-potential users and obtain the final potential customer information.

[0010] To optimize the above technical solutions, the specific measures taken also include:

[0011] The obtaining of the original data of user ordering information includes: obtaining user side, product side, order side, points side and similar business ordering data; the preprocessing and multi-dimensional analysis include: data extraction, data filtering, data cleaning, data association and data verification, comprehensively considering the specific scenarios of data missing to process missing values, unifying the data format, and unique-hot encoding the qualitative data.

[0012] Furthermore, the business objectives are analyzed, feature engineering is performed on the data after preprocessing and multi-dimensional analysis to obtain a full feature set, and a training set of an integrated learning model is constructed based on the full feature set, including: first, extracting statistical features of users and products; then, based on user ordering behavior, constructing statistical features of time-sensitive behavioral interests associated with business objectives; then obtaining auxiliary features in cross-business data; finally, feature derivation and feature intersection are performed based on existing features to obtain a complete full feature set; the positive and negative sample calibers of the integrated learning model are determined based on the full feature set, and the integrated learning model training set is obtained based on the positive and negative sample calibers of the integrated learning model.

[0013] Furthermore, the LightGBM ensemble learning model and the Random Forest ensemble learning model are established respectively, and after the training set of the ensemble learning model is input, the evaluation indicators and full feature information output by the two ensemble learning models are obtained, including: inputting the ensemble learning model training set into the LightGBM ensemble learning model and the Random Forest ensemble learning model for training and tuning respectively; the evaluation indicators include F1 value and recall rate; the full feature information includes full feature importance value, feature importance ranking and relevance.

[0014] Furthermore, the output results of the two integrated learning models are analyzed, and the evaluation indicators and full feature information obtained by the two integrated learning models under the same input are compared and verified, and the features are pre-processed accordingly, including: after obtaining the evaluation indicators and the full feature information, confirming whether the model and the full feature information are normal according to the evaluation indicators; comparing and verifying the feature importance rankings of the two integrated learning models, filtering invalid features according to the feature importance values, removing features with feature importance values ​​lower than a threshold, and filtering redundant features according to the correlation, and only retaining one of the features for the feature pairs with a correlation higher than the threshold, so as to obtain the important feature sets of the two integrated learning models; merging the important feature sets of the two integrated learning models to obtain the final important feature set.

[0015] Furthermore, constructing the training set and test set of the CNN deep learning model according to the final set of important features includes: determining the positive and negative sample calibers of the CNN deep learning model according to the final set of important features, and obtaining the training set and test set of the CNN deep learning model based on the positive and negative sample calibers of the CNN deep learning model.

[0016] Furthermore, the method of establishing a CNN deep learning model, inputting a training set of the CNN deep learning model for model training and parameter adjustment, and performing verification, and predicting potential customers by inputting a test set, includes: establishing a CNN deep learning potential customer mining model with multiple convolutional layers, inputting the training set of the CNN deep learning model into the CNN deep learning model for training and optimization, and then taking the test set of the CNN deep learning model and inputting it into the trained model to obtain a potential customer list and a potential customer intention probability predicted by the model.

[0017] The second aspect of the present application provides a high-intent potential customer mining system that integrates ensemble learning and CNN deep learning, including:

[0018] The data acquisition module is used to obtain the original data of user ordering information and perform preprocessing and multi-dimensional analysis;

[0019] The feature engineering module is used to analyze business goals, perform feature engineering on the data after preprocessing and multi-dimensional analysis, obtain the full feature set, and build the training set of the integrated learning model based on the full feature set;

[0020] The ensemble learning model construction module is used to respectively establish the LightGBM ensemble learning model and the Random Forest ensemble learning model. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models are obtained. The output results of the two ensemble learning models are analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input are compared and verified. Based on this, the features are pre-processed to obtain the final set of important features; the training set and test set of the CNN deep learning model are constructed based on the final set of important features;

[0021] The deep learning model building module is used to build a CNN deep learning model, input the training set of the CNN deep learning model to train the model and adjust the parameters, and perform verification, and predict potential customers by inputting the test set;

[0022] The potential customer mining module is used to obtain historical data of point marketing, analyze and process it, and screen pseudo high-potential users; the predicted potential customers output by the CNN deep learning model are used to filter pseudo high-potential users to obtain the final potential customer information.

[0023] The third aspect of the present application provides an electronic device, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for mining high-intent potential customers by integrating integrated learning and CNN deep learning as described in the first aspect of the present application is implemented.

[0024] The fourth aspect of the present application provides a computer-readable storage medium storing a computer program, wherein the computer program enables a computer to execute the high-intent potential customer mining method that integrates integrated learning and CNN deep learning as described in the first aspect of the present application.

[0025] Compared with the prior art, the present invention has the following beneficial effects:

[0026] The traditional marketing model based on expert experience has great bottlenecks in timeliness and efficiency. The communication time between business personnel and data analysts is long, and the efficiency of manual analysis and data acquisition is relatively low. The solution of the present invention can automatically process a large amount of data, reduce the need for manual analysis and screening, and improve work efficiency. At the same time, new data can be updated and processed in real time, so that the business team can quickly know and respond to market changes.

[0027] The present invention provides cross-business data integration to assist decision-making, and can comprehensively mine and utilize data assets. It not only covers user and business data, but also integrates other business data to achieve interoperability between various business data, characterize users from a broader perspective, and improve the comprehensiveness and accuracy of decision-making.

[0028] The present invention provides prediction of potential customers and also provides the probability of predicting potential customers, so that the business team can adopt different marketing strategies according to different probability values, effectively control marketing costs, and maximize the use of resources.

[0029] The method and system for mining high-intent potential customers by integrating integrated learning and CNN deep learning of the present invention have the following characteristics:

[0030] 1. High accuracy: This paper adopts the mainstream CNN deep learning model, combined with Random Forest and LightGBM

[0031] Two types of integrated learning algorithm models access rich feature data and comprehensively consider the user's multi-dimensional information; use the RandomForest integrated learning algorithm and the LightGBM integrated learning algorithm to pre-screen features and verify each other; the CNN deep learning model uses the screened feature data and deep learning technology to automatically learn and extract deep connections in the data, accurately capture the connections and potential needs between users' complex behaviors, and predict results with higher accuracy.

[0032] 2. High timeliness: The potential customer prediction method provided by the present invention only needs to input feature data, and can automatically learn user data, business data, and complex relationships between cross-business data in an automated and intelligent manner. It can quickly meet the prediction needs of large-scale data, respond to market changes in a timely manner, and solve the problems of timeliness and low efficiency of manual analysis.

[0033] 3. Scalability: The present invention has good scalability and can easily expand the scale and function of the model as the business develops and the amount of data increases. It can easily add and modify new features, data sources, algorithms, and integrate cross-business data to continuously improve the ability and effect of potential customer mining and meet the growing needs of precise business operations. BRIEF DESCRIPTION OF THE DRAWINGS

[0034] Figure 1 : Flow chart of the method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning of the present invention.

[0035] Figure 2 : The feature pre-processing flow chart of the LightGBM ensemble learning model and the Random Forest ensemble learning model of the present invention;

[0036] Figure 3 : Schematic diagram of the process of predicting potential customers using the CNN deep learning model of the present invention.

[0037] Figure 4 : A schematic diagram of the process of constructing a pseudo high-potential user portrait and filtering the pseudo high-potential user portrait of the present invention. DETAILED DESCRIPTION

[0038] The above contents of the present invention are further described in detail below in the form of embodiments, but this should not be understood as the scope of the above subject matter of the present invention being limited to the following embodiments, and all technologies realized based on the above contents of the present invention belong to the scope of the present invention.

[0039] Some of the terms in the present invention are explained as follows:

[0040] CNN (Convolutional Neural Network) is a deep learning model that consists of multiple layers of network structures, including convolutional layers, pooling layers, fully connected layers, etc. It learns and recognizes complex patterns in data by combining these layers. CNN has a high degree of fault tolerance. Through a large amount of training data and a complex network structure, it can automatically learn and adapt to various complex input data, and has a certain degree of robustness to noise and deformation; it also supports the processing and learning of large-scale data, and can perform efficient training and reasoning on high-performance computing platforms. In addition, CNN also has good interpretability and visualization capabilities, and can explain the network's learning process and decision-making basis through feature map visualization and gradient visualization.

[0041] LightGBM (Light Gradient Boosting Machine) is an efficient, distributed, high-performance gradient boosting ensemble learning algorithm suitable for machine learning tasks such as regression, classification, and ranking. It not only inherits the advantages of the gradient boosting framework, such as high accuracy, strong robustness to outliers, and strong interpretability, but also innovatively uses algorithms such as histograms and gradient-based one-sided sampling, with excellent training speed and prediction performance; and supports running on distributed computing systems such as Hadoop and Spark, and can make full use of the cluster's computing and storage resources to meet the needs of various large-scale data processing applications.

[0042] Random Forest is an ensemble learning algorithm based on decision trees. It builds multiple decision trees and combines the prediction results of these decision trees to provide more accurate and stable predictions. It is suitable for tasks such as classification, regression, feature selection, and anomaly detection. For the original data set, Random Forest automatically generates multiple sub-data sets to form multiple training sets, and randomly selects a part of the features on each sub-data set to build a decision tree, which can reduce overfitting and improve the generalization ability of the model.

[0043] With the diversification trend of business development today, the accumulation of relevant data has shown explosive growth, and data assets are becoming increasingly rich. At the same time, user needs are constantly rising, and product marketing strategies are gradually diversifying, making data-driven decision-making essential. This shift has brought new challenges to the in-depth analysis and application of user data. Therefore, it has become increasingly urgent to accurately explore users, target market positioning, and reduce marketing costs.

[0044] In this regard, the present invention proposes a method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning. Figure 1 As shown, the following steps are included:

[0045] Based on business understanding and needs, we obtain the original data of user ordering information, and perform pre-processing and multi-dimensional analysis;

[0046] Analyze business goals, perform feature engineering on preprocessed and multi-dimensionally analyzed data to obtain a full feature set, and build a training set for the ensemble learning model based on the full feature set;

[0047] A LightGBM ensemble learning model and a Random Forest ensemble learning model were established respectively. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models were obtained. The output results of the two ensemble learning models were analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input were compared and verified. Based on this, the features were pre-processed to obtain the final set of important features. The training set and test set of the CNN deep learning model were constructed based on the final set of important features.

[0048] Establish a CNN deep learning model, input the training set of the CNN deep learning model to train the model and adjust the parameters, and verify it, and predict potential customers by inputting the test set;

[0049] Obtain historical data of point marketing, analyze and process it, and screen pseudo high-potential users; use the predicted potential customers output by the CNN deep learning model to filter pseudo high-potential users and obtain the final potential customer information.

[0050] Based on business understanding and needs, the original data of user ordering information is obtained, including: obtaining user-side, product-side, order-side, points-side and similar business ordering data; preprocessing and multi-dimensional analysis include: data extraction, data filtering, data cleaning, data association and data verification, comprehensively considering the specific scenarios of data missing to handle missing values, unifying the data format, and unique-hot encoding qualitative data.

[0051] Analyze the business goals, perform feature engineering on the data after preprocessing and multi-dimensional analysis to obtain the full feature set, and build the training set of the integrated learning model based on the full feature set, including: first extracting the statistical features of users and products; then constructing the statistical features of time-sensitive behavioral interests associated with the business goals based on the user's ordering behavior; then obtaining auxiliary features from cross-business data; finally, performing feature derivation and feature intersection based on the existing features to obtain a complete full feature set; determine the positive and negative sample calibers of the integrated learning model based on the full feature set, and obtain the integrated learning model training set based on the positive and negative sample calibers of the integrated learning model.

[0052] like Figure 2 As shown in the figure, a LightGBM ensemble learning model and a Random Forest ensemble learning model are established respectively. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models are obtained, including: inputting the ensemble learning model training set into the LightGBM ensemble learning model and the Random Forest ensemble learning model for training and tuning respectively; the evaluation indicators include F1 value and recall rate; the full feature information includes the full feature importance value, feature importance ranking and relevance.

[0053] The output results of the two ensemble learning models are analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input are compared and verified, and the features are pre-processed accordingly, including: after obtaining the evaluation indicators and the full feature information, confirm whether the model and the full feature information are normal according to the evaluation indicators. If the F1 value is too low, it is necessary to check the ensemble learning model and the full feature information, adjust and modify the ensemble learning model, and then train it again; secondly, compare and verify the feature importance rankings of the two ensemble learning models. If the difference is not large, invalid features are filtered according to the feature importance values, and features with feature importance values ​​lower than the threshold are removed, and redundant features are filtered according to the correlation. For feature pairs with a correlation higher than the threshold, only one feature is retained to obtain the important feature sets of the two ensemble learning models; if the difference is too large, it is necessary to check the ensemble learning model structure, adjust the ensemble learning model parameters, and re-obtain the full feature information to ensure the correctness of the full feature information; merge the important feature sets of the two ensemble learning models to obtain the final important feature set.

[0054] Constructing a training set and a test set of a CNN deep learning model according to the final set of important features, including: determining the positive and negative sample calibers of the CNN deep learning model according to the final set of important features, and obtaining the training set and the test set of the CNN deep learning model based on the positive and negative sample calibers of the CNN deep learning model.

[0055] like Figure 3As shown, a CNN deep learning model is established, a training set of the CNN deep learning model is input for model training and parameter adjustment, and verification is performed, and potential customers are predicted by inputting a test set, including: establishing a CNN deep learning potential customer mining model with multiple convolutional layers, using the training set of the CNN deep learning model to input the CNN deep learning model for training and tuning, and then taking the test set of the CNN deep learning model and inputting it into the trained model to obtain a potential customer list and potential customer intention probability predicted by the model.

[0056] The present invention is further described in detail below in conjunction with specific embodiments:

[0057] like Figure 1 As shown, the preferred implementation mode of the present invention is to first obtain and process all relevant raw data, perform feature engineering, and establish a LightGBM ensemble learning model and a Random Forest ensemble learning model to sort and screen all features; then establish a CNN deep learning model to predict potential customers, and finally obtain and process the historical data of point marketing, screen pseudo high-potential users and construct pseudo high-potential user portraits, use the CNN deep learning model to predict potential customers to filter pseudo high-potential users, and obtain the final potential customer list and potential customer intention probability.

[0058] Therefore, the method of the present invention specifically comprises three main parts, namely: Figure 2 As shown in the figure, the LightGBM ensemble learning model and the Random Forest ensemble learning model are established to pre-process all features. Figure 3 As shown in the figure, a CNN deep learning model is established to predict potential customers, and Figure 4 As shown, a pseudo high-potential user portrait is constructed and pseudo high-potential filtering is performed on potential customers predicted by the CNN deep learning model.

[0059] The working principle and working process of the present invention are as follows:

[0060] 1. Establish the LightGBM ensemble learning algorithm model and the Random Forest ensemble learning model to pre-process all features and verify each other;

[0061] The LightGBM ensemble learning model and the Random Forest ensemble learning model are used to pre-process all features, filter out invalid features and redundant features with high correlation, and verify the results of the two ensemble learning algorithm models to ensure the correctness of the feature pre-screening. The specific steps are as follows:

[0062] Step 1: Obtain multi-dimensional data on all user behavior, user preference, product, order, and points of user point subscriptions, as well as similar business subscriptions;

[0063] Step 2: Preprocess and analyze the above-mentioned full data; check the data format, type, average value and other statistical values, whether there are missing values ​​and other information, check whether there are any anomalies, and handle abnormal data according to specific scenarios, such as using the mean to fill missing age values, filtering out attributes with too many missing values, and performing one-hot encoding on qualitative data;

[0064] Step 3: Use the standardized data processed in step 2 to perform feature engineering; first add the relevant basic attributes to the features; then construct the user's time-effective behavior interest statistical features based on the user's order records; then obtain the auxiliary features in the cross-business data; finally, perform feature derivation and feature intersection based on the existing features; obtain the user's full feature set through several parts of features;

[0065] Step 4: Determine the positive and negative sample calibers based on the full set of user features obtained according to the business objectives, obtain the positive and negative sample data sets, analyze the amount and proportion of positive and negative sample data, determine whether there is a data imbalance problem, determine the final positive and negative samples and proportions, and obtain the training set of the integrated learning model;

[0066] Step 5: Establish the LightGBM ensemble learning model and the Random Forest ensemble learning model, input the training sets of the ensemble learning models for model training and tuning, and the two ensemble learning algorithm models output evaluation indicators such as F1 value and recall rate, as well as the importance ranking, importance value and correlation heat map of all features;

[0067] Step 6: First, confirm whether the model and data are normal based on the evaluation indicators. If the F1 value is too low, check the model and data, and train again after modification; secondly, compare and verify the two feature importance rankings. If the difference is not large, filter invalid features according to the feature importance value, remove features with importance values ​​lower than the threshold, and filter redundant features according to the correlation heat map. For feature pairs with correlation higher than the threshold, only one feature is retained to obtain two important feature sets; if the difference is too large, check the model structure, adjust the model parameters, and re-obtain the feature importance to ensure the correctness of the feature importance results; finally, merge the two important feature sets to obtain the final important feature set and reduce the feature dimension.

[0068] In some embodiments, the LightGBM ensemble learning model uses a histogram algorithm, a gradient-based unilateral sampling algorithm, and a mutually exclusive feature bundling algorithm, which greatly reduces memory consumption, has higher accuracy, supports parallel learning, and has a faster training speed. Its working principle is as follows:

[0069] The LightGBM ensemble learning model minimizes the objective function W(θ) as follows:

[0070] W(θ)=L(θ)+Ω(θ)

[0071] Where θ is a parameter, L is a loss function, and Ω is a regularization term. The specific expression is as follows:

[0072]

[0073] Where T is the number of leaves, is the L2 regularization term of the leaf fraction, γ is the complexity cost introduced by adding a new leaf node, and f t is the prediction function.

[0074] The prediction result of the Tth round of the LightGBM ensemble learning model = the prediction result of the previous T-1th round + the new prediction content, expressed as follows:

[0075]

[0076] in, is the prediction result of the tth round of the ensemble learning model, is the prediction result of the model in round t-1.

[0077] The LightGBM ensemble learning model obtains the optimal solution by minimizing the objective function and obtains the final prediction result by accumulating prediction results round by round.

[0078] In some embodiments, the Random Forest ensemble learning model uses decision trees as basic units, and a large number of decision trees are integrated to form a random forest. The model adopts the idea of ​​bagging, and n training samples are taken from the training set with replacement each time to form a new training set to train multiple sub-models.

[0079] For classification problems, the classification tree of the Random Forest ensemble learning model uses the Gini coefficient to determine the optimal feature. Under each condition of each feature A, the Gini coefficient Gini(D,A) of the training set D is:

[0080]

[0081] Among them, D1 and D2 are two parts divided by D.

[0082] 2. Establish a CNN deep learning model for potential customer prediction

[0083] A CNN deep learning model with 5 convolutional layers is established through the PyTorch framework. The activation function uses the ReLU function, and the pooling method of the pooling layer uses average pooling. The input of the CNN deep learning model is the important feature set after the pre-screening of the first step of the integrated model feature. The user label is output through the calculation of the input layer, convolution layer, excitation layer, pooling layer and fully connected layer. All its features come from the important features screened in the first part. It not only contains multi-dimensional data information, but also can fully explore user characteristics and improve prediction accuracy. In addition, the feature dimension is low, the model training speed is fast, and the risk of overfitting is reduced.

[0084] The specific steps of data processing in the CNN deep learning model include:

[0085] Step 1: First input the training data into the input layer

[0086] Step 2: Feature extraction is performed through local correlation and window sliding of the convolution layer. After layer-by-layer convolution, the maximum value of each convolution kernel is extracted to obtain the key features.

[0087] c=max(c1,c2,...,c n )

[0088] Where c is the key feature obtained, c n Represents the features extracted by the nth convolution kernel.

[0089] Step 3: Use the Relu function as the activation function to achieve nonlinear mapping. The Relu function can avoid the gradient vanishing problem and simplify the reverse derivative operation. The specific formula is as follows:

[0090] f(x)=max(0,x)

[0091] Where f(x) is the Relu function.

[0092] The output of using the Relu function as the activation function is:

[0093] y=max(0,ω Τ x+b)

[0094] Among them, y is the output result of using the Relu function as the activation function, x is the parameter, ω is the neuron weight, and b is the neuron bias.

[0095] Step 5: The pooling layer uses the maximum pooling method for pooling operation:

[0096] (c1,c2,...,c n ) pooling =maxpooling(c1,c2,...,c n )

[0097] Among them, (c1,c2,...,c n ) pooling Represents the pooling operation of the pooling layer, and maxpooling represents maximum pooling.

[0098] Step 6: Perform feature mapping through the fully connected layer to output the final potential customer label.

[0099] In some embodiments, the loss function used by the CNN deep learning model of the present invention is MSE mean square error, and the gradient update method is stochastic gradient descent.

[0100] 3. Build pseudo high potential user portraits and filter pseudo high potential customers predicted by CNN

[0101] Pseudo high potential user portraits can further filter and screen potential customers predicted by CNN, and further improve the quality of potential customers in terms of model prediction and mining capabilities, such as Figure 4 As shown, the specific steps are as follows:

[0102] Step 1: First, obtain historical points marketing data, including marketing time, marketing methods, marketing products, marketing results, and user marketing feedback, and perform preprocessing and analysis.

[0103] Step 2: Construct statistical features based on the number of marketing times and conversion effects, and use the CNN deep learning model to predict the potential customer list and potential customer intention probability. Screen out users with low total conversion rate, low marketing times and success ratio, users with non-target products in marketing results, and users with negative marketing feedback as pseudo high-potential users, and construct pseudo high-potential user portraits based on user characteristics.

[0104] Step 3: Use the predicted potential customers output by the CNN deep learning model to filter out pseudo high-potential users and obtain the final potential customer list and potential customer intention probability.

[0105] In another embodiment, the present invention proposes a high-intent potential customer mining system integrating ensemble learning and CNN deep learning, comprising:

[0106] The data acquisition module is used to obtain the original data of user ordering information based on business understanding and needs, and perform preprocessing and multi-dimensional analysis;

[0107] The feature engineering module is used to analyze business goals, perform feature engineering on the data after preprocessing and multi-dimensional analysis, obtain the full feature set, and build the training set of the integrated learning model based on the full feature set;

[0108] The ensemble learning model construction module is used to respectively establish the LightGBM ensemble learning model and the Random Forest ensemble learning model. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models are obtained. The output results of the two ensemble learning models are analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input are compared and verified. Based on this, the features are pre-processed to obtain the final set of important features; the training set and test set of the CNN deep learning model are constructed based on the final set of important features;

[0109] The deep learning model building module is used to build a CNN deep learning model, input the training set of the CNN deep learning model to train the model and adjust the parameters, and perform verification, and predict potential customers by inputting the test set;

[0110] The potential customer mining module is used to obtain historical data of point marketing, analyze and process it, and screen pseudo high-potential users; the predicted potential customers output by the CNN deep learning model are used to filter pseudo high-potential users to obtain the final potential customer information.

[0111] In another embodiment of the present invention, an electronic device is proposed, comprising: a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, the method for mining high-intent potential customers by integrating integrated learning with CNN deep learning as described above is implemented.

[0112] In another embodiment, the present invention proposes a computer-readable storage medium storing a computer program, which enables a computer to execute the high-intent potential customer mining method that integrates ensemble learning and CNN deep learning as described above.

[0113] In the embodiments disclosed in the present application, the computer storage medium can be a tangible medium that can contain or store a program for use by or in conjunction with an instruction execution system, device or equipment. The computer storage medium can include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the above. More specific examples of computer storage media can include electrical connections based on one or more lines, portable computer disks, hard disks, random access memories (RAM), read-only memories (ROM), erasable programmable read-only memories (EPROM or flash memory), optical fibers, portable compact disk read-only memories (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination of the above.

[0114] The above are only preferred embodiments of the present invention and are not intended to limit the present invention in any form. Any technician familiar with the profession, without departing from the scope of the technical solution of the present invention and based on the technical essence of the present invention, may make any simple modifications, equivalent substitutions and improvements to the above embodiments, which are still within the protection scope of the technical solution of the present invention.

Claims

1. A method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning, characterized in that: The following steps are involved: Obtain the original data of user ordering information, and perform pre-processing and multi-dimensional analysis; Analyze business goals, perform feature engineering on preprocessed and multi-dimensionally analyzed data to obtain a full feature set, and build a training set for the ensemble learning model based on the full feature set; A LightGBM ensemble learning model and a Random Forest ensemble learning model were established respectively. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models were obtained. The output results of the two ensemble learning models were analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input were compared and verified. Based on this, the features were pre-processed to obtain the final set of important features. The training set and test set of the CNN deep learning model were constructed based on the final set of important features. Establish a CNN deep learning model, input the training set of the CNN deep learning model to train the model and adjust the parameters, and verify it, and predict potential customers by inputting the test set; Obtain historical data of point marketing, analyze and process it, and screen pseudo high-potential users; use the predicted potential customers output by the CNN deep learning model to filter pseudo high-potential users and obtain the final potential customer information.

2. The method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning according to claim 1 is characterized in that: The obtaining of the original data of user ordering information includes: obtaining user side, product side, order side, points side and similar business ordering data; the preprocessing and multi-dimensional analysis include: data extraction, data filtering, data cleaning, data association and data verification, comprehensively considering the specific scenarios of data missing to process missing values, unifying the data format, and unique-hot encoding the qualitative data.

3. The method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning according to claim 1 is characterized in that: The business objectives are analyzed, feature engineering is performed on the data after preprocessing and multi-dimensional analysis to obtain a full feature set, and a training set of an integrated learning model is constructed based on the full feature set, including: first, extracting statistical features of users and products; then, based on user ordering behavior, constructing statistical features of time-sensitive behavior interests associated with the business objectives; then obtaining auxiliary features in cross-business data; finally, feature derivation and feature intersection are performed based on existing features to obtain a complete full feature set; the positive and negative sample calibers of the integrated learning model are determined based on the full feature set, and the integrated learning model training set is obtained based on the positive and negative sample calibers of the integrated learning model.

4. The method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning according to claim 1 is characterized in that: The method establishes a LightGBM ensemble learning model and a Random Forest ensemble learning model respectively, and after inputting the training set of the ensemble learning model, obtains the evaluation indicators and full feature information output by the two ensemble learning models, including: inputting the ensemble learning model training set into the LightGBM ensemble learning model and the Random Forest ensemble learning model respectively for training and tuning; the evaluation indicators include F1 value and recall rate; the full feature information includes full feature importance value, feature importance ranking and relevance.

5. The method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning according to claim 4 is characterized in that: The output results of the two integrated learning models are analyzed, and the evaluation indicators and full feature information obtained by the two integrated learning models under the same input are compared and verified, and the features are pre-processed accordingly, including: after obtaining the evaluation indicators and the full feature information, confirming whether the model and the full feature information are normal according to the evaluation indicators; comparing and verifying the feature importance rankings of the two integrated learning models, filtering invalid features according to the feature importance values, removing features with feature importance values ​​lower than a threshold, and filtering redundant features according to the correlation, retaining only one of the features for the feature pairs with a correlation higher than a threshold, thereby obtaining the important feature sets of the two integrated learning models; merging the important feature sets of the two integrated learning models to obtain the final important feature set.

6. The method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning according to claim 1 is characterized in that: The method of constructing a training set and a test set of a CNN deep learning model according to the final set of important features includes: determining the positive and negative sample calibers of the CNN deep learning model according to the final set of important features, and obtaining the training set and the test set of the CNN deep learning model based on the positive and negative sample calibers of the CNN deep learning model.

7. The method for mining high-intent potential customers by integrating ensemble learning and CNN deep learning according to claim 1, characterized in that: The method of establishing a CNN deep learning model, inputting a training set of the CNN deep learning model to perform model training and parameter adjustment, and performing verification, and predicting potential customers by inputting a test set, includes: establishing a CNN deep learning potential customer mining model with multiple convolutional layers, inputting the training set of the CNN deep learning model into the CNN deep learning model for training and optimization, and then taking the test set of the CNN deep learning model and inputting it into the trained model to obtain a potential customer list and a potential customer intention probability predicted by the model.

8. A high-intent potential customer mining system integrating ensemble learning and CNN deep learning, characterized in that: include: The data acquisition module is used to obtain the original data of user ordering information and perform preprocessing and multi-dimensional analysis; The feature engineering module is used to analyze business goals, perform feature engineering on the data after preprocessing and multi-dimensional analysis, obtain the full feature set, and build the training set of the integrated learning model based on the full feature set; The ensemble learning model construction module is used to respectively establish the LightGBM ensemble learning model and the Random Forest ensemble learning model. After inputting the training set of the ensemble learning model, the evaluation indicators and full feature information output by the two ensemble learning models are obtained. The output results of the two ensemble learning models are analyzed, and the evaluation indicators and full feature information obtained by the two ensemble learning models under the same input are compared and verified. Based on this, the features are pre-processed to obtain the final set of important features; the training set and test set of the CNN deep learning model are constructed based on the final set of important features; The deep learning model building module is used to build a CNN deep learning model, input the training set of the CNN deep learning model to train the model and adjust the parameters, and perform verification, and predict potential customers by inputting the test set; The potential customer mining module is used to obtain historical data of point marketing, analyze and process it, and screen pseudo high-potential users; the predicted potential customers output by the CNN deep learning model are used to filter pseudo high-potential users to obtain the final potential customer information.

9. An electronic device, characterized in that: include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein when the processor executes the computer program, the method for mining high-intent potential customers by integrating integrated learning and CNN deep learning as described in any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium storing a computer program, characterized in that: The computer program enables the computer to execute the high-intent potential customer mining method that integrates integrated learning and CNN deep learning as described in any one of claims 1 to 7.

Images