A broadband television user recommendation method and device based on information gain

Abstract

The present invention discloses a broadband TV user recommendation method and device based on information gain, which generalizes and extracts target customer attributes, obtains target customer attribute features and performs preprocessing, generates training samples, and performs dimension reduction processing on the attribute features in the training samples according to the correlation of the attribute features and information gain; then the training samples are input into a plurality of preset classification models, and cross-validation is performed to select the classification model with the best performance; finally, the parameters of the selected classification model are adjusted, and the classification model after parameter adjustment is used to predict potential customers, and potential customers whose prediction probability is greater than a set probability threshold and whose active days are greater than the active threshold are selected to form a recommended customer group. The present invention selects features according to the information gain level of each feature, filters noise data, and can better improve the prediction ability of the model.

Description

Technical Field

[0001] The present application belongs to the technical field of deep learning recommendation models, and in particular, relates to a method and device for recommending broadband TV users based on information gain. Background Art

[0002] Data mining refers to the process of discovering hidden information from a large amount of data, such as features (Pattern), trends (Trend) and correlations (Relationship). It can also be said to be extracting information or knowledge from data. By using complex data analysis tools to highlight the information structure under large data sets, the hidden potential relationships between these data are discovered. For customer consumption data, data mining technology can help companies better maintain customer relationships, discover the differences in customer group consumption needs and behavior patterns in multiple attributes and dimensions, and realize accurate potential customer mining. Data mining technology mainly includes the following aspects: clustering, classification, regression analysis, prediction, and association rules.

[0003] In data mining technology, several commonly used algorithms for customer mining are as follows: clustering algorithm, classification model and association rule mining. Clustering algorithm can find the differences in consumption behavior of different customer groups and help companies formulate accurate marketing strategies. Classification model can predict the trend of future customer consumption behavior. Association rule mining can find the correlation between customers and products and guide companies to cross-sell. Making full use of customer consumption information can improve customer loyalty and the quality of relationship management. At the same time, effectively allocate resources to maximize company profits and maintain competitiveness in the same industry.

[0004] However, most of the existing technologies select features relatively arbitrarily, which leads to overfitting of the model and low accuracy of the model prediction results. Summary of the invention

[0005] The purpose of this application is to provide a broadband TV user recommendation method and device based on information gain to overcome the problems of overfitting and low prediction accuracy.

[0006] In order to achieve the above purpose, the technical solution of this application is as follows:

[0007] A broadband television user recommendation method based on information gain, comprising:

[0008] Select users who have already applied for broadband as target users, conduct return visits to target users, and use target users who have applied for broadband TV services after the return visit as positive samples, and use target users who have not applied for broadband TV services after the return visit as negative samples;

[0009] Perform generalized extraction of target customer attributes, obtain target customer attribute features and perform preprocessing to generate training samples;

[0010] Perform dimension reduction processing on attribute features in training samples according to the relevance and information gain of attribute features;

[0011] Input the training samples into multiple preset classification models, perform cross-validation, and select the classification model with the best performance;

[0012] Adjust the parameters of the selected classification model;

[0013] The classification model with adjusted parameters is used to predict potential customers, and potential customers whose predicted probability is greater than the set probability threshold and whose active days are greater than the active threshold are selected to form the recommended customer group.

[0014] Furthermore, the dimension reduction process of the attribute features in the training samples according to the correlation and information gain of the attribute features includes:

[0015] Calculate the correlation of attribute features, filter out attribute features with high correlation, and remove attribute features whose correlation is greater than the correlation threshold;

[0016] The continuous attribute features are binned, and then the IV values ​​of the categorical attribute features and the continuous attribute features after binning are calculated, and the attribute features with IV values ​​within the preset range are retained.

[0017] Furthermore, the classification model with the best performance is selected as the extreme tree classification method.

[0018] Furthermore, the parameters of the selected classification model are adjusted, and the adjusted parameters include maximum depth and leaf node weights.

[0019] Furthermore, the parameter adjustment of the selected classification model includes:

[0020] Selecting a preset number of parameters with the aid of a learning curve function to generate a preset number of classification models;

[0021] The classification model is trained, and the classification model with the best parameters is selected using a cross-validation method.

[0022] The present application also proposes a broadband TV user recommendation device based on information gain, comprising a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the broadband TV user recommendation method based on information gain.

[0023] This application proposes a broadband TV user recommendation method and device based on information gain, which uses complex data analysis tools to highlight the information structure under large data sets and discover the hidden potential relationships between these data. Currently, telecom operators use big data technology to conduct customer acquisition analysis based on users' personal attributes, browsing behaviors, usage habits, etc., and predict future trends in customer consumption behaviors, thereby saving customer acquisition costs. It has the following beneficial effects:

[0024] 1. Combined with broad user characteristics. Utilizing the massive resources of operator data, we have extensively acquired the user's social attributes, basic attributes of consumption behavior, service attributes, family social attributes, and government and enterprise group attributes.

[0025] 2. Select features based on information gain. Selecting features based on the information gain level of each feature and filtering out noise data can improve the prediction ability of the model.

[0026] 3. It has strong replicability. It is also applicable to customer acquisition research in other directions and businesses. By changing the data sample and making simple parameter adjustments, it can become a new model with good efficiency.

[0027] 4. Multi-layer optimization model. By selecting a variety of classification models, select an algorithm with excellent performance and adaptability, adjust parameters with the help of learning curves, and select K-fold cross validation and other methods to optimize model performance. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] Figure 1 This is a flow chart of the broadband TV user recommendation method based on information gain in this application;

[0029] Figure 2 Schematic diagram of the performance indicators of each classification algorithm. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical solution and advantages of the present application more clearly understood, the present application is further described in detail below in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and are not used to limit the present application.

[0031] This application provides a broadband TV user recommendation method based on information gain, such as Figure 1 As shown, the service system applied to broadband TV user operators recommends users to broadband TV user operators. It includes:

[0032] Step S1, select users who have applied for broadband as target users, conduct return visits to the target users, and use the target users who have applied for broadband TV services after the return visit as positive samples, and use the target users who have not applied for broadband TV services after the return visit as negative samples.

[0033] Taking the mobile company as an example, users who have already applied for mobile broadband are selected as target users. After cleaning sensitive customers, outbound calls (recalls) are made, and the outbound call results are summarized and sorted. Then, target users who have applied for broadband TV services through all channels in the month are selected as positive samples, and users who have failed outbound marketing are selected as negative samples.

[0034] Step S2: generalize and extract target customer attributes, obtain target customer attribute features and perform preprocessing to generate training samples.

[0035] Generalized customer attributes are extracted from the mobile company's data delivery platform, and then used as attribute features by analyzing user demand for products, product characteristics, user purchasing power, stickiness with product providers, user basic information, etc.

[0036] For example, by analyzing the user's demand for the product, it is found that the user likes to watch movies and TV, and further obtains the data such as the number of days they watch broadband TV, traffic usage, traffic of Tencent / iQiyi and other related video apps, and whether Tencent / iQiyi and other related video apps have activated membership services as attribute features;

[0037] By analyzing product features, in the case of supporting multiple broadband TVs at the same time, the number of broadband TVs under the same broadband, the number of broadband and broadband TV users under the same contact number, whether broadband is integrated, and other data are obtained as attribute features; in the case of family products, information such as the number of family members, whether there are elderly people, whether there are children, and the activity data of broadband traffic and usage time are obtained as attribute features;

[0038] By analyzing the purchasing power of users, when they have a certain consumption capacity, we can obtain data such as actual cost, traffic cost, and preferential cost as attribute features;

[0039] By analyzing the stickiness with the product provider, in the case of close mobile contact, the data such as the number of times the user dials or receives 10086 calls and whether a WeChat account is opened are obtained as attribute features;

[0040] By analyzing the basic information of users, we can obtain data such as user age, family network, group users, whether they are locals, calling parties, and roaming traffic as attribute features.

[0041] It should be noted that which attribute features are extracted and how many are extracted are selected according to needs. In the embodiment of the present application, 162 attribute features are selected.

[0042] After selecting the attribute features, preprocessing is required. For example, standardize the attribute features, process the null values, take the average value for continuous variables, and set the discrete variables to 0. Uniformly scale the maximum and minimum values ​​for continuous variables. Eliminate the influence of dimension and the influence of the variable's own variation and numerical size, and reasonably improve the generalization performance of the model. In addition, since the size difference between the positive and negative samples is large and the positive and negative samples are unbalanced, SOMTE oversampling can be used, that is, deriving new positive samples based on the existing positive sample features; or downsampling can be performed at a certain ratio to eliminate the number of negative samples in the training samples so that the positive and negative samples present a suitable ratio.

[0043] Step S3: perform dimension reduction processing on the attribute features in the training samples according to the correlation and information gain of the attribute features.

[0044] Due to the multicollinearity of the data, there are correlations between the various features of the user. Multicollinearity will lead to spatial instability of the solution, resulting in weak generalization of the model; too many features will reduce the efficiency of the model in finding rules. In order to strengthen the extraction and synthesis of effective information and the rejection of useless information, this application uses the information gain of each feature to select features.

[0045] In a preferred embodiment, the attribute features in the training samples are subjected to dimension reduction processing according to the correlation and information gain of the attribute features, including:

[0046] Step S3.1, calculate the correlation of attribute features, screen attribute features with high correlation, and remove attribute features whose correlation is greater than the correlation threshold.

[0047] This application uses a custom function getCorrThre to calculate the correlation of attribute features, screen attribute features with high correlation, and then remove attribute features whose correlation is greater than the correlation threshold. For example, the received data type is the original DaTaFrame data set, and the default correlation threshold is 0.85; pandas is used to directly call corr() to calculate the correlation coefficient between features, and finally obtain the features with correlation greater than 0.85 for each continuous feature, and then remove them. In this way, high-correlation features are removed and non-high-correlation features are retained.

[0048] Step S3.2: bin the continuous attribute features, then calculate the IV values ​​of the categorical attribute features and the continuous attribute features after binning, and retain the attribute features whose IV values ​​are within a preset range.

[0049] This application also defines a custom function calIVAndTransform to calculate the IV value of continuous attribute features and categorical attribute features. In order to obtain the stability of the model, the continuous features need to be binned first.

[0050] This embodiment adopts the chi-square binning method. The chi-square binning is based on the chi-square distribution and chi-square value to determine whether a certain factor will affect the target variable. The binning is completed by a bottom-up continuous merging method. First, the continuous features are initialized, that is, the values ​​of the continuous variables are sorted and the initial discretization is performed; secondly, the discretized data is merged. The bin merging process is divided into two steps and repeated continuously: 1. Calculate the chi-square value of each adjacent box. 2. Merge adjacent boxes with low chi-square values ​​(according to the chi-square test principle, the lower the chi-square value, the more similar the distribution of the two boxes is, and they can be merged). The conditions for merging to stop are: 1. Until the chi-square value of all adjacent boxes is greater than or equal to the set chi-square value threshold. (Select an appropriate chi-square value threshold based on the degree of freedom and significance level; the degree of freedom can be determined based on the data as (R-1)x(C-1), because both adjacent boxes are calculated, so R=2; the number of features is the target feature and the feature to be transformed, so C=2, so the degree of freedom is 1. The significance level is recommended to be 0.1) 2. The number of boxes reaches the preset number. (In this implementation, the maximum number of bins is set to 10 by default)

[0051] Then, the IV values ​​of the categorical attribute features and the continuous attribute features after binning are calculated, and the IV values ​​of each feature are output and arranged from large to small according to the IV value. Generally speaking, IV values ​​< 0.03 are considered to have no predictive ability and IV values ​​> 0.5 are considered to be extremely high and suspicious, so features with IV values ​​between [0.1, 0.45] are selected.

[0052] It should be noted that dimensionality reduction can also be performed by calculating the WOE value of continuous attribute features and categorical attribute features. However, for scenarios where the number of samples in the groups varies greatly, the WOE value may not be able to well represent the predictive ability of this attribute feature, and the IV value is better for calculation.

[0053] This embodiment finally successfully reduces the dimensionality of 162 fields to 68 fields.

[0054] Step S4: input the training samples into multiple preset classification models, perform cross-validation, and select the classification model with the best performance.

[0055] This application intends to substitute the common classification models XGBoost, LogisticRegression, KNeighbors, SVC, MLP, DecisionTree, RandomForest, AdaBoost, and ExtraTrees into the training set and verification labels, adopt the cross-validation method, the number of folds of the cross-validation method can be set to 10, perform Cross_Val_Score calculation, and select the algorithm with excellent performance according to the score.

[0056] Figure 2 is the Cross_Val_Score value of each algorithm. After comparison, this application finally selected ExtraTrees (extreme tree classification method). This algorithm is very similar to random forest and random forest algorithm, and both are composed of many decision trees. However, there are still some differences between extreme trees and random forests: 1. Random forest uses the Bagging model, and all samples used by extraTree are randomly selected, but the features are randomly selected. Because the split is random, the results are better than those obtained by random forests to some extent. 2. Random forest obtains the best bifurcation attribute in a random subset, while extreme trees obtain bifurcation values ​​completely randomly, thereby bifurcating the decision tree.

[0057] Step S5: Adjust parameters of the selected classification model.

[0058] The purpose of this step is to further optimize the model by adjusting the parameters (maximum depth and leaf node weights) and then using the learning curve function to select parameters with better performance.

[0059] In a specific embodiment, adjusting parameters of the selected classification model includes:

[0060] Selecting a preset number of parameters with the aid of a learning curve function to generate a preset number of classification models;

[0061] The classification model is trained, and the classification model with the best parameters is selected using a cross-validation method.

[0062] Then use 10-fold cross validation to train the model: 1. Divide the data set into 10 equal parts; 2. Perform model training 10 times using 1 part of the data as test data and the rest as training data; 3. Obtain 10 models and use the ROC curve to select the one with better effect. Using 10-fold cross validation can reduce overfitting and alleviate the embarrassing situation of small amount of training data.

[0063] Step S6: Use the classification model after parameter adjustment to predict potential customers, and select potential customers whose prediction probability is greater than the set probability threshold and whose active days are greater than the active threshold to form a recommended customer group.

[0064] This application uses a parameter-adjusted classification model to predict potential customers, selects users with a predicted probability > 0.6 and broadband TV active days > 5 days last month to form a recommended customer group, removes sensitive customers according to the cleaning rules, and then enters the outbound call platform for outbound calls. If the customer is interested in applying, he can apply for the broadband TV VIP package service online. After using the technical solution of this application, the outbound call success rate has increased by nearly 9PP compared to before use.

[0065] In one embodiment, the present application also proposes a broadband TV user recommendation device based on information gain, including a processor and a memory storing a plurality of computer instructions, wherein the computer instructions, when executed by the processor, implement the steps of the broadband TV user recommendation method based on information gain.

[0066] The specific definition of the broadband TV user recommendation device based on information gain can be found in the definition of the broadband TV user recommendation method based on information gain above, which will not be repeated here. The above-mentioned broadband TV user recommendation device based on information gain can be implemented in whole or in part by software, hardware and a combination thereof. It can be embedded in or independent of the processor in the computer device in the form of hardware, or it can be stored in the memory of the computer device in the form of software, so that the processor can call and execute the above corresponding operations.

[0067] The memory and the processor are electrically connected directly or indirectly to achieve data transmission or interaction. For example, these elements can be electrically connected to each other through one or more communication buses or signal lines. The memory stores a computer program that can be run on the processor, and the processor implements the network topology layout method in the embodiment of the present invention by running the computer program stored in the memory.

[0068] The memory may be, but is not limited to, a random access memory (RAM), a read only memory (ROM), a programmable read-only memory (PROM), an erasable programmable read-only memory (EPROM), an electrically erasable read-only memory (EEPROM), etc. The memory is used to store a program, and the processor executes the program after receiving an execution instruction.

[0069] The processor may be an integrated circuit chip with data processing capabilities. The above-mentioned processor may be a general-purpose processor, including a central processing unit (CPU), a network processor (NP), etc. The methods, steps and logic block diagrams disclosed in the embodiments of the present invention may be implemented or executed. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0070] The above-mentioned embodiments only express several implementation methods of the present application, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for a person of ordinary skill in the art, several variations and improvements can be made without departing from the concept of the present application, and these all belong to the protection scope of the present application. Therefore, the protection scope of the patent of the present application shall be subject to the attached claims.

Claims

1. A broadband TV user recommendation method based on information gain, characterized in that: The broadband television user recommendation method based on information gain comprises: Select users who have already applied for broadband as target users, conduct return visits to target users, and use target users who have applied for broadband TV services after the return visit as positive samples, and use target users who have not applied for broadband TV services after the return visit as negative samples; Perform generalized extraction of target customer attributes, obtain target customer attribute features and perform preprocessing to generate training samples; Perform dimension reduction processing on attribute features in training samples according to the relevance and information gain of attribute features; Input the training samples into multiple preset classification models, perform cross-validation, and select the classification model with the best performance; Adjust the parameters of the selected classification model; Use the parameter-adjusted classification model to predict potential customers, and select potential customers whose prediction probability is greater than the set probability threshold and whose active days are greater than the active threshold to form a recommended customer group; The step of performing dimension reduction processing on the attribute features in the training samples according to the correlation and information gain of the attribute features includes: Calculate the correlation of attribute features, filter out attribute features with high correlation, and remove attribute features whose correlation is greater than the correlation threshold; Binning the continuous attribute features, then calculating the IV values ​​of the categorical attribute features and the continuous attribute features after binning, and retaining the attribute features whose IV values ​​are within the preset range; The training samples are input into a plurality of preset classification models, and cross-validation is performed to select the classification model with the best performance, including: Obtain common classification models, input training samples into common classification models, perform cross-validation, calculate model performance scores, and select the classification model with the best performance based on the scores; The step of adjusting parameters of the selected classification model includes: Selecting a preset number of parameters with the aid of a learning curve function to generate a preset number of classification models; The classification model is trained, and the classification model with the best parameters is selected using a cross-validation method.

2. The broadband TV user recommendation method based on information gain according to claim 1, characterized in that: The classification model with the best performance is selected as the extreme tree classification method.

3. The broadband TV user recommendation method based on information gain according to claim 2, characterized in that: The parameters of the selected classification model are adjusted, and the adjusted parameters include maximum depth and leaf node weights.

4. A broadband TV user recommendation device based on information gain, comprising a processor and a memory storing a plurality of computer instructions, characterized in that: When the computer instructions are executed by a processor, the steps of the method described in any one of claims 1 to 3 are implemented.

Images