CN109242257B - 4G internet user complaint model based on key index correlation analysis - Google Patents

Abstract

The invention discloses a 4G internet user complaint model based on key index correlation analysis, which comprises the following steps: s1, finding out a plurality of factors influencing the complaints of the user by using a logistic regression model and establishing a decision tree model; s2 extracting user data, integrating and processing the data; s3, carrying out T test on the user data, and primarily finding out factors influencing complaints; s4, setting a training data set, and establishing a logistic regression model by using R language; optimizing by using a post-step logistic regression method according to a result returned after the primary model is established to obtain a final result and determine a final model; s5, carrying out chi-square inspection on the model; and S6, predicting the test data set by using the model, and performing cross statistics on the predicted result and the actual result. The invention provides a 4G internet user complaint model based on key index correlation analysis, which is used for solving the cause of complaint of a 4G internet user and carrying out preventive solution on potential complaint users in advance.

Description

4G internet user complaint model based on key index correlation analysis

Technical Field

The invention relates to the field of mobile communication, in particular to a 4G internet user complaint model based on key index correlation analysis.

Background

With the rapid development of the mobile internet, the requirements of users on the mobile network are higher and higher, the reduction of the complaint amount and the prevention of potential complaints of 4G users in advance are important tasks of network workers, the complaint amount of the 4G users can be reduced, and the internet surfing satisfaction of the 4G users is improved.

Disclosure of Invention

The invention aims to solve one or more defects and provides a 4G internet user complaint model based on key index correlation analysis.

In order to realize the purpose, the technical scheme is as follows:

A4G internet user complaint model building method based on key index correlation analysis comprises the following steps:

s1: exploring influence factors and conditions of the complaint users, finding out a plurality of factors influencing the complaints of the users by utilizing a logistic regression model, and establishing a decision tree model;

s2: extracting user data comprising the plurality of factors described in step S1, including raw complaint user data and non-complaint user data; integrating and processing data;

s3: performing T test on the original complaint user data and the non-complaint user data acquired in the step S2, comparing the difference between the complaint indexes and the non-complaint indexes, and primarily finding out factors influencing complaint;

s4: setting a training data set, and establishing a logistic regression model by using an R language; setting whether complaints are taken as dependent variables or not, setting values to be 0 and 1, optimizing by using a post-step logistic regression method according to a result returned after a primary model is established, obtaining a final result and determining a final model;

s5: performing chi-square test on the model to ensure that all variables of the model pass significance test and ensure that the whole model is significant;

s6: and predicting the test data set by using the model, and performing cross statistics on a predicted result and an actual result.

Further, in step S1, the several factors affecting the user complaint include an attach success rate, an attach delay, a default bearer success rate, a default bearer delay, a Tcp23 handshake success rate, and a Tcp23 handshake delay.

Further, the step S2 of sorting the data includes the following steps:

s2.1: removing percentage signs from all data success rates, reserving numbers between 0 and 100, and reserving 2 bits at decimal points;

s2.2: removing records with index missing number larger than 5;

s2.3: filling the missing values by adopting a K nearest neighbor method for the records with the number of the missing values of 1-5;

s2.4: 80% of complaint records and 80% of non-complaint records were randomly drawn for training the model, with the remaining 20% for prediction.

The final model is:

compared with the prior art, the invention has the beneficial effects that:

the invention provides a 4G internet user complaint model based on key index correlation analysis, which is used for solving the cause of complaint of a 4G internet user and carrying out preventive solution on potential complaint users in advance.

Drawings

FIG. 1 is a flow chart of the present invention.

Detailed Description

The drawings are for illustrative purposes only and are not to be construed as limiting the patent;

the invention is further illustrated below with reference to the figures and examples.

Example 1

A method for establishing a complaint model of a 4G internet user based on key index correlation analysis, please refer to fig. 1, comprising the following steps:

s1: exploring influence factors and conditions of the complaint users, finding out a plurality of factors influencing the complaints of the users by utilizing a logistic regression model, and establishing a decision tree model;

and exploring influencing factors and conditions of the complaint users. T test is carried out on 15 indexes between a complaint group and a non-complaint group, and the significant difference of the success rate of attach, the time delay of attach, the success rate of default bearing, the time delay of default bearing, the success rate of Tcp23 handshake is found, and the time delay of Tcp23 handshake is found. Using a logistic regression model, 5 factors affecting user complaints were found: attach success rate, attach delay, default bearer delay, Tcp23 handshake success rate, Tcp23 handshake delay. And establishing a decision tree model according to the 5 factors to obtain the conditions of the user with high possibility of complaints:

1) tcp23 handshake success rate <100 and Tcp23 handshake delay <80 and attach delay < 178;

2) tcp23 handshake success rate <100 and Tcp23 handshake delay > 80;

3) tcp23 handshake success rate is 100 and attach delay is <374 and attach delay is 179 and default bearer delay is 191;

4) tcp23 handshake success rate is 100 and attach delay is < 179;

5) tcp23 handshake success rate is 100 and attach delay is 374.

The decision tree model has a prediction accuracy of 65.7%, but the resulting complaint conditions are less than satisfactory. Finally, analyzing whether a threshold exists in each index single factor analysis, so that the user is likely to complain beyond the threshold, and finally concluding as follows:

1) the user is likely to complain when the attach success rate is less than or equal to 60%;

2) the user is likely to complain when the attach delay is higher than or equal to 1500 ms;

3) when the default bearing success rate is less than or equal to 20%, the user is likely to complain;

4) when the default bearing delay is more than or equal to 1000ms, the user is likely to complain;

5) a user is likely to complain when the Tcp23 handshake success rate is 90% or less.

S2: extracting user data comprising the plurality of factors described in step S1, including raw complaint user data and non-complaint user data; integrating and processing data;

this example extracts 2336 copies of original complaint user data and 2993 copies of non-complaint users, for a total of 5329 records. The data are integrated into a table, which contains the following indices:

the method comprises the following steps of attach success rate, attach time delay, default bearing success rate, default bearing time delay, DNS success rate, DNS time delay, Tcp12 handshake success rate, Tcp12 handshake time delay, Tcp23 handshake success rate, Tcp23 handshake time delay, Get response success rate, Get response time delay, Post response success rate, Post response time delay and large packet (more than 500KB) download rate.

For modeling, the data is processed as follows:

1) all success rates are removed with percentiles, numbers between 0 and 100 are reserved, and decimal points are reserved with 2 bits. For example, 99.5% transforms to 99.50;

2) records with deletion index number greater than 5: firstly, calculating the index missing number of each record: if more than 5 (not containing 5) indexes of 15 indexes have missing values, direct elimination is considered, otherwise, the later modeling is very unfavorable. After kicking off, 4819 data remained (complaint 1999, non-complaint 2820);

3) for the records with the number of the missing values of 1-5, filling the missing values by adopting a K nearest neighbor method, so that the data to be modeled has no missing values and is convenient to model;

4) 80% of complaint records and 80% of non-complaint records were drawn randomly for training the model (total 3820 records), with the remaining 20% for prediction.

S3: performing T test on the original complaint user data and the non-complaint user data acquired in the step S2, comparing the difference between the complaint indexes and the non-complaint indexes, and primarily finding out factors influencing complaint;

the mathematical application of the P values in the T-test is as follows:

p value	Probability of coincidence	For invalid hypothesis	Statistical significance of
				P>0.05	The probability of occurrence of the accident is more than 5 percent	Fail to negate null hypothesis	Two groups of differences have no significant meaning
P<0.05	The probability of occurence is less than 5 percent	Invalid hypotheses may be negated	Two groups of differences have significant meaning
				P<0.01	The probability of occurence is less than 1%	Invalid hypotheses may be negated	The difference between the two is very significant

The difference between the complaint index and the non-complaint index is compared by using a T test method, and the factors influencing the complaint can be preliminarily found out. The data adopts 2336 parts of original complaint user data, 2993 parts of non-complaint users and 5329 records (containing missing values), each index is independently calculated, the records containing the missing values are automatically ignored, and the results are shown in the following table:

as can be seen from the above table, there are strong significant differences (99% confidence level) between the complaining user and the non-complaining user in the indexes of attach success rate, attach delay, default bearer delay, Tcp23 handshake success rate, and Tcp23 handshake delay. The default bearer success rate also differed significantly at the 95% confidence level and not at the 99% confidence level. Other indices were not significantly different.

S4: setting a training data set, and establishing a logistic regression model by using an R language; setting whether complaints are taken as dependent variables or not, setting values to be 0 and 1, optimizing by using a post-step logistic regression method according to a result returned after a primary model is established, obtaining a final result and determining a final model;

a logistic regression model is built using the R language from the training data set (3820 records). Whether complaints are taken as dependent variables or not is judged, the values are only 0 and 1(0 is non-complaint and 1 is complaint), and 15 indexes are taken as independent variables. R the returned results after establishing the preliminary model are as follows:

Call:

glm (whether complaint-attach success rate + attach delay + default bearing success rate +

Default bearer delay + DNS success rate + DNS delay + Tcp12 handshake success rate + Tcp12 handshake delay + Tcp23 handshake success rate + Tcp23 handshake delay + Get response success rate + Get response delay + Post response success rate + Post response delay + big packet, greater than 500KB. download rate,

family＝"binomial",data＝train.dt)

Deviance Residuals:

Min 1Q Median 3Q Max

-3.2577-0.9793-0.9161 1.3126 2.3556

Coefficients:

big package, download rate of more than 500KB. -1.491 e-062.603 e-06-0.5730.566695

---

Signif.codes:0‘***’0.001‘**’0.01‘*’0.05‘.’0.1‘’1

(Dispersion parameter for binomial family taken to be 1)

Null deviance:5182.4 on 3819 degrees of freedom Residual deviance:5006.4 on 3804 degrees of freedom AIC:5038.4

Number of Fisher Scoring iterations:5

In the above results, the bands indicate that the coefficients are very significant, the bands indicate insignificant, and the bands do not indicate significant. The coefficients for the multiple indices in the results are not significant and further optimization of the model is required.

Optimizing the model by using a backward stepwise logistic regression method (removing a factor which has the least obvious influence in each step to ensure that the model is better than the previous model and only the factor with the obvious coefficient is finally kept), wherein the final result of the backward stepwise regression is as follows:

Call:

glm (whether complaint-attach success rate + attach delay + default bearer delay + Tcp12 handshake delay + Tcp23 handshake success rate + Tcp23 handshake delay, family ═ binomial ", data ═ train

Deviance Residuals:

Min 1Q Median 3Q Max -3.2489 -0.9789 -0.9198 1.3174 1.7080

Coefficients:

Signif.codes:0‘***’0.001‘**’0.01‘*’0.05‘.’0.1‘’1

(Dispersion parameter for binomial family taken to be 1)

Null deviance:5182.4 on 3819 degrees of freedom Residual deviance:5012.0 on 3813 degrees of freedom AIC:5026

Number of Fisher Scoring iterations:5

The model finally retains 6 metrics, but the factor of Tcp12 handshake delay is still not significant, and the model is re-modeled with the remaining 5 metrics in consideration of rejecting the metric:

Call:

glm (format is whether complaint is-attach success rate + attach delay + default bearer delay + Tcp23 handshake success rate + Tcp23 handshake delay, family is "binomial", data is train

Deviance Residuals:

Min 1Q Median 3Q Max -3.2410 -0.9790 -0.9201 1.3180 1.5413

Coefficients:

Signif.codes:0‘***’0.001‘**’0.01‘*’0.05‘.’0.1‘’1

(Dispersion parameter for binomial family taken to be 1)

Null deviance:5182.4 on 3819 degrees of freedom Residual deviance:5014.2 on 3814 degrees of freedom AIC:5026.2

Number of Fisher Scoring iterations:5

So far, all coefficients are remarkable, and relatively important variables are reserved. The model thus established is:

+0.0005773 default bearer delay-0.0666100 Tcp23 handshake success rate +0.0038896 Tcp23 handshake delay

When P calculated by the model is greater than 0.5, complaints are considered, otherwise complaints are not.

S5: performing chi-square test on the model to ensure that all variables of the model pass significance test and ensure that the whole model is significant;

the variables of the model pass the significance test and ensure that the whole model is significant, so that the model is correct and meaningful. Chi-square test was performed on the model, and the results are shown below:

Analysis of Deviance Table

Model:binomial,link:logit

response whether complaints are made

Terms added sequentially(first to last)

Signif.codes:0‘***’0.001‘**’0.01‘*’0.05‘.’0.1‘’1

The model passed the global significance test, indicating that a model consisting of the above variables is meaningful.

S6: and predicting the test data set by using the model, and performing cross statistics on a predicted result and an actual result.

The model is used to predict a test data set (956 records, complaint 396, non-complaint 560, total) to predict whether an unknown customer is likely to complain. When P calculated by the model is greater than 0.5, complaints are considered, otherwise complaints are not. And performing cross statistics on the predicted result and the actual result, wherein the results are shown in the following table:

	complaint user	Non-complaint users
			Predicting complaints	162	94
Prediction of non-complaints	234	466

Predicting the complaint accuracy rate: 162/(162+94) × 100% ═ 63.8%;

predicting the non-complaint accuracy rate: 466/(466+234) × 100% — 66.6%;

overall prediction accuracy: (466+162)/(466+162+234+94) × 100 ═ 65.7%;

the recall ratio is as follows: 162/(162+234) ═ 40.9%.

And (3) knotting: the model prediction capability is better.

It should be understood that the above-described embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Other variations and modifications will be apparent to persons skilled in the art in light of the above description. And are neither required nor exhaustive of all embodiments. Any modification, equivalent replacement, and improvement made within the spirit and principle of the present invention should be included in the protection scope of the claims of the present invention.

Claims (3)

1. A4G internet user complaint model based on key index correlation analysis is characterized by comprising the following steps:

s1: exploring influence factors and conditions of the complaint users, finding out a plurality of factors influencing the complaints of the users by utilizing a logistic regression model, and establishing a decision tree model; the plurality of factors influencing the user complaints comprise an attach success rate, an attach time delay, a default bearing success rate, a default bearing time delay, a Tcp23 handshake success rate and a Tcp23 handshake time delay;

s2: extracting user data comprising the plurality of factors described in step S1, including raw complaint user data and non-complaint user data; integrating and processing data;

s3: performing T test on the original complaint user data and the non-complaint user data acquired in the step S2, comparing the difference between the complaint indexes and the non-complaint indexes, and primarily finding out factors influencing complaint;

s4: setting a training data set, and establishing a logistic regression model by using an R language; setting whether complaints are taken as dependent variables or not, setting values to be 0 and 1, optimizing by using a post-step logistic regression method according to a result returned after a primary model is established, obtaining a final result and determining a final model;

s5: performing chi-square test on the model to ensure that all variables of the model pass significance test and ensure that the whole model is significant;

s6: and predicting the test data set by using the model, and performing cross statistics on a predicted result and an actual result.

2. The 4G internet user complaint model based on key index correlation analysis of claim 1, wherein the step S2 of collating data includes the following steps:

s2.1: removing percentage signs from all data success rates, reserving numbers between 0 and 100, and reserving 2 bits at decimal points;

s2.2: removing records with index missing number larger than 5;

s2.3: filling the missing values by adopting a K nearest neighbor method for the records with the number of the missing values of 1-5;

s2.4: 80% of complaint records and 80% of non-complaint records were randomly drawn for training the model, with the remaining 20% for prediction.

3. The 4G internet user complaint model based on key index correlation analysis as claimed in claim 1 or 2, wherein the final model is established as

Images