CN113592250A - Outbound robot distribution method - Google Patents
Outbound robot distribution method Download PDFInfo
- Publication number
- CN113592250A CN113592250A CN202110772699.3A CN202110772699A CN113592250A CN 113592250 A CN113592250 A CN 113592250A CN 202110772699 A CN202110772699 A CN 202110772699A CN 113592250 A CN113592250 A CN 113592250A
- Authority
- CN
- China
- Prior art keywords
- model
- robot
- neural network
- variables
- learning
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Pending
Links
- 238000000034 method Methods 0.000 title claims abstract description 63
- 238000013528 artificial neural network Methods 0.000 claims abstract description 32
- 238000000513 principal component analysis Methods 0.000 claims abstract description 24
- 230000001419 dependent effect Effects 0.000 claims abstract description 13
- 238000010801 machine learning Methods 0.000 claims abstract description 13
- 238000000611 regression analysis Methods 0.000 claims abstract description 3
- 230000008569 process Effects 0.000 claims description 14
- 238000012417 linear regression Methods 0.000 claims description 10
- YHXISWVBGDMDLQ-UHFFFAOYSA-N moclobemide Chemical compound C1=CC(Cl)=CC=C1C(=O)NCCN1CCOCC1 YHXISWVBGDMDLQ-UHFFFAOYSA-N 0.000 claims description 9
- 238000003062 neural network model Methods 0.000 claims description 9
- 238000004458 analytical method Methods 0.000 claims description 8
- 238000012937 correction Methods 0.000 claims description 8
- 239000011159 matrix material Substances 0.000 claims description 7
- 238000012546 transfer Methods 0.000 claims description 7
- 230000006870 function Effects 0.000 claims description 5
- 238000012545 processing Methods 0.000 claims description 5
- 238000004364 calculation method Methods 0.000 claims description 4
- 238000004519 manufacturing process Methods 0.000 claims description 4
- 238000012216 screening Methods 0.000 claims description 3
- 238000010626 work up procedure Methods 0.000 claims 1
- 238000012314 multivariate regression analysis Methods 0.000 description 6
- 238000012360 testing method Methods 0.000 description 6
- 238000004422 calculation algorithm Methods 0.000 description 5
- 230000000875 corresponding effect Effects 0.000 description 4
- 238000012549 training Methods 0.000 description 4
- 238000004891 communication Methods 0.000 description 3
- 238000010586 diagram Methods 0.000 description 3
- 238000012544 monitoring process Methods 0.000 description 3
- 230000000717 retained effect Effects 0.000 description 3
- 238000012353 t test Methods 0.000 description 3
- 230000000694 effects Effects 0.000 description 2
- 241001123248 Arma Species 0.000 description 1
- 230000002159 abnormal effect Effects 0.000 description 1
- 238000009825 accumulation Methods 0.000 description 1
- 238000005452 bending Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000001364 causal effect Effects 0.000 description 1
- 230000008859 change Effects 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 150000001875 compounds Chemical class 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000013461 design Methods 0.000 description 1
- 230000005611 electricity Effects 0.000 description 1
- 238000005516 engineering process Methods 0.000 description 1
- 230000004907 flux Effects 0.000 description 1
- 230000003993 interaction Effects 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 238000007620 mathematical function Methods 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 230000010355 oscillation Effects 0.000 description 1
- 230000007115 recruitment Effects 0.000 description 1
- 238000004088 simulation Methods 0.000 description 1
- 238000006467 substitution reaction Methods 0.000 description 1
Images
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/06—Resources, workflows, human or project management; Enterprise or organisation planning; Enterprise or organisation modelling
- G06Q10/063—Operations research, analysis or management
- G06Q10/0631—Resource planning, allocation, distributing or scheduling for enterprises or organisations
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
- G06F18/21—Design or setup of recognition systems or techniques; Extraction of features in feature space; Blind source separation
- G06F18/213—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods
- G06F18/2135—Feature extraction, e.g. by transforming the feature space; Summarisation; Mappings, e.g. subspace methods based on approximation criteria, e.g. principal component analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
- G06N3/084—Backpropagation, e.g. using gradient descent
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q10/00—Administration; Management
- G06Q10/04—Forecasting or optimisation specially adapted for administrative or management purposes, e.g. linear programming or "cutting stock problem"
Abstract
The invention relates to an outbound robot allocation method, which comprises the following steps: 1) the client calls and enters different calling skill groups through the selection of the routing module; 2) after entering the calling skill group, predicting a dependent variable, namely the time required by the robot to dial the task amount through principal component analysis and multiple regression analysis methods in sequence; 3) based on the prediction result in the step 2), autonomous machine learning is carried out on the prediction result by utilizing a method of combining a time series model and BP neural network weighting, an optimized learning rule is obtained, and then the robot seats are distributed. Compared with the prior art, the method has the advantages of reducing labor cost, improving prediction precision, having the autonomous learning capability of the neural network and the like.
Description
Technical Field
The invention relates to the technical field of communication, in particular to an outbound robot allocation method.
Background
The call center is a comprehensive information service system realized by utilizing modern communication and computer technology, and can automatically and flexibly process a large amount of incoming and outgoing telephone services; in addition, the call center can provide functions of inquiring, summarizing, statistically analyzing and the like of all communication data of the enterprise, and assists enterprise decision-making. Currently, call centers are widely used in telecommunications, finance, government agencies, electricity, postal services, and other industries.
In the call center in the prior art, a telephone customer service system mainly carries out the operation of calling out and answering incoming calls in a manual seat mode; in the calling scene, the telephone customer service system transfers the connected telephone to the manual seat in the idle state through the dialing system. It can be seen that the size of the call center is limited by the number of human agents, and if more outgoing and incoming calls are to be handled, the call center needs to employ more customer service personnel; the call center is subject to a large number of recruitment, training and management problems and the salary cost for hiring people is gradually increased by simply relying on manual seats for telephone operation.
With the advent of voice robots, some enterprises begin to adopt a traditional prediction model method to perform allocation management on outbound robots, such as a single regression prediction model or a kalman filter prediction model, however, the traditional prediction model is generally established based on rules and mathematical functions, and has the following problems: 1. the mathematical logics of the back of different models are different, and the application scene is more limited; 2. based on a causal prediction model, the independence of equation establishment and local rules is taken as an assumed basis, and the long-term prediction bias is large; 3. the model which does not need a large amount of data to solve the problems of less historical data, sequence integrity and the like is only suitable for medium-short term prediction of exponential growth; 4. the model for predicting the future time situation of the system, which is only related to the present time and is not related to historical data, is not suitable for medium-long term prediction. The above problems lead to low accuracy of robot distribution results and failure to realize multi-turn dialogue interaction including complex content.
Disclosure of Invention
The invention aims to overcome the defects of the prior art and provide an outbound robot distribution method.
The purpose of the invention can be realized by the following technical scheme:
an outbound robot assignment method, the method comprising the steps of:
s1: and the client calls and enters different calling skill groups through the selection of the routing module.
S2: after entering the call-in skill group, the time required by the robot to dial the task amount is predicted through principal component analysis and multiple regression analysis methods in sequence. The concrete contents are as follows:
21) after entering an incoming call skill group, acquiring relevant variables of an outgoing call, extracting a group of irrelevant variables from the relevant variables by adopting a principal component analysis model, and taking the variables as principal components of the principal component analysis model; the related variables of the outbound are actual production data, including the working time of the robot, the actual processing list quantity and the dialing success quantity in different scenes. The irrelevant variables comprise dialing list quantity, dialing times, receiving quantity, missed receiving quantity, second hang-up quantity, success quantity, average call duration and CCOD channel number.
22) Judging the number of reserved principal components by checking a correlation coefficient matrix among variables, and screening the reserved principal components;
23) and performing predictive analysis on the principal components and the dependent variables screened out by the principal component analysis model by adopting multivariate linear regression analysis to obtain the dependent variables.
S3: and based on the prediction result of the step S2, performing autonomous machine learning on the prediction result by using a method of combining a time series model and BP neural network weighting to obtain an optimized learning rule, and further distributing the robot seats.
Further, the time series model adopts a time series ARIMA model. The structure of the time series ARIMA model is as follows:
in the formula, E (ε)sεt) 0, arbitrary s<t, wherein xiThe time sequence data is the specific dialing time length of the robot for completing the task amount; epsiloniIs a residual term; b is a delay operator; p is the autoregressive order; q is the moving average order; d is the difference order;
is a difference operator;
is an autoregressive coefficient polynomial, and theta (B) is a moving average coefficient polynomial.
The process of autonomous machine learning by using the BP neural network model comprises the following steps:
a. initializing a network weight, and determining specific numerical values of an input data sample and an expected output data sample, wherein the input data sample is data which affects expected output and comprises a dialing list quantity, dialing times, a receiving quantity, a non-receiving quantity, a second hanging quantity, a success quantity, an average call duration and a CCOD channel number, and the expected output data sample is the dialing duration, namely the time required by the robot to complete a task quantity;
b. for all connection weights wij、vjiAnd all threshold values thetaj、γiCarry out assignment, wherein wijIs the connection value of the incident layer and the hidden layer, vjiThe connection value of the hidden layer and the output layer is obtained; randomly extracting a set of data input data pairs AiAnd the desired output data pair YiAnd endowing the neural network.
c. Calculating the input value s of each unit of the intermediate hidden layer according to the input data sample, the threshold values of all the units of the input layer and the intermediate hidden layerjThen, according to the transfer function, the output value b of each unit of the intermediate hidden layer is obtainedj：
d. According to bjThe weight between the middle layer and the output layer and the threshold value of each unit in the output layer are calculated to obtain the input value I of each unit in the output layertOutputting the actual output value c of each unit of the output layer through a transfer functiont；
e. Calculating the correction error of each unit of the output layer according to the expected output data
f. Root of herbaceous plantAccording to vji、
And bjCalculating a correction error
Then combining with a threshold value gammaiAnd calculating the next connection weight value and threshold value of the middle layer and the input layer.
g. According to
Obtaining the connection weight and the threshold value between the input layer and the middle layer in the next learning process through correction calculation, then randomly extracting a data sample to provide for the BP neural network, and learning again according to the method of the second step until all samples are completely learned;
h. randomly selecting a certain sample again, learning according to the step c, and if the global error of the BP neural network is smaller than the contrast error value which is set originally, indicating that the BP neural network is converged; if the learning times in the whole learning process reach the originally set learning times and the BP neural network is not converged, the network learning process is finished.
The specific content of step S3 is:
31) respectively calculating the error square sum of the time series model and the BP neural network model, and assigning the weight of each single prediction model according to the principle of minimum integral error square sum;
32) and acquiring a combined prediction model based on the assignment weights of the individual prediction models to further obtain a final prediction result.
The expression of the prediction result obtained by combining the prediction models is as follows:
in the formula, m is the number of the single-term prediction models, wjFor individual prediction modelsThe weight of the weight is calculated,
and (4) the prediction result of each single prediction model.
Weight w of each univariate predictive modeljThe expression of (a) is:
in the formula, ejThe sum of the squares of the errors for each of the univariate predictive models.
Compared with the prior art, the outbound robot allocation method provided by the invention at least comprises the following beneficial effects:
1) according to the invention, through the combination of principal component analysis, multivariate regression analysis, time sequence and BP neural network weighting, a deep machine learning algorithm is brought into distribution management, the application limitation of a single model is greatly reduced, the autonomous machine learning of the neural network is utilized, excessive dependence on experience is avoided, manpower is liberated, and unnecessary artificial monitoring control is reduced.
2) The method comprises the steps of firstly, utilizing a principal component analysis and multivariate regression analysis method to efficiently and comprehensively capture all relevant independent variables and determine appropriate coefficients, and then utilizing a time sequence and BP neural network combined prediction model to improve on the basis of the original prediction method.
3) Compared with the traditional single prediction model, the method can comprehensively capture relevant model variables, reduces the application limitation of the single model, and has the neural network autonomous learning capability, which are incomparable advantages compared with the traditional single prediction model.
Drawings
FIG. 1 is a schematic flow chart of an outbound robot assignment method in an embodiment;
FIG. 2 is a diagram of a principal component analysis model in an embodiment;
FIG. 3 is a flow chart of principal component analysis in the example;
FIG. 4 is a diagram of a BP neural network model prediction case in an embodiment.
Detailed Description
The invention is described in detail below with reference to the figures and specific embodiments. It is to be understood that the embodiments described are only a few embodiments of the present invention, and not all embodiments. All other embodiments, which can be obtained by a person skilled in the art without any inventive step based on the embodiments of the present invention, shall fall within the scope of protection of the present invention.
Examples
The invention relates to an outbound robot distribution method, which combines principal component analysis, multivariate regression analysis, time sequence and BP neural network weighting to bring a deep machine learning algorithm into robot distribution management, thereby greatly reducing the application limitation of a single model, avoiding excessive dependence on experience by using autonomous machine learning of a neural network, liberating manpower and reducing unnecessary artificial monitoring control.
As shown in fig. 1, the outbound robot assignment method of the present invention specifically includes the following steps:
step one, incoming a telephone line.
And the client calls and enters different calling skill groups through the selection of the routing module.
And step two, robot call distribution.
And after entering the call-in skill group, generating actual production data, such as the working time of the robot, the actual processing list quantity, the dialing success quantity in different scenes and the like, and eliminating abnormal values in the data for subsequent model processing. The subsequent model processing process is as follows: the relation between the dialing time of the robot list and variables such as the dialing list quantity, the number of channels of a virtual Call Center (CCOD) and the average call time is obtained by analyzing historical data through principal component analysis and multivariate regression analysis, the aim of accurately predicting the dialing time of the robot and reasonably distributing the work of the robot is achieved, autonomous machine learning is carried out by a method combining a time sequence and BP neural network weighting, and therefore optimal robot seat distribution is automatically obtained. Specifically, the method comprises the following steps:
1) firstly, a Principal Component Analysis (PCA) is used for extracting a group of direct and irrelevant variables (including dialing list quantity, dialing times, receiving quantity, non-receiving quantity, second hanging quantity, success quantity, average call duration and CCOD channel number) from a large number of relevant variables (a robot can generate a large amount of real production data in daily work), the variables are called principal components, and the criterion for judging the number of the principal components in the PCA is as follows:
i, judging the number of main components according to prior experience and theoretical knowledge;
II, judging the required principal component number according to a threshold value of an accumulation value of variable variance to be explained;
and III, judging the reserved principal component number by checking a correlation coefficient matrix of k x k between variables.
The invention can adopt one of the common principal component number determination methods to judge the reserved principal component book.
Most common are eigenvalue based methods. Each principal component is associated with an eigenvalue of a correlation coefficient matrix, the first principal component being associated with the largest eigenvalue, the second principal component being associated with the second largest eigenvalue, and so on. The Kaiser-Harris criterion suggests that principal components with eigenvalues greater than 1 are retained, and components with eigenvalues less than 1 account for less variance than is contained in a single variable. The Cattlel lithotripsy test draws a graph of the characteristic value and the principal component number. Such a pattern can clearly show the bending state of the pattern, and the main component above the maximum change of the pattern can be retained. And finally, simulation can be carried out, and the characteristic value to be extracted is judged according to a random data matrix with the same size as the initial matrix. If a certain eigenvalue based on real data is larger than the corresponding average eigenvalue of a set of random data matrix, then the principal component can be retained. This method is called parallel analysis. The principal component analysis model diagram is shown in fig. 2, and the flow chart is shown in fig. 3.
The goal of PCA is to replace a large number of correlated variables with a small set of uncorrelated variables while retaining as much information as possible of the original variables, these derived variables are called principal components, which are linear combinations of observed variables. If the first main component is:
PC1＝a1X1+a2X2+…+akXk
it is a weighted combination of k observed variables that is most explanatory of the variance of the initial set of variables. The second principal component is also a linear combination of the initial variables, with the interpretability of the variance being second and orthogonal (uncorrelated) to the first principal component. Each of the latter principal components maximizes its degree of interpretation of the variance while being orthogonal to all of the preceding principal components. Theoretically, you could choose the same principal component as the number of variables, but from a practical point of view, it is desirable to approximate the full set of variables with fewer principal components.
2) The time for predicting the number of N lists dialed by the robot is used as a dependent variable, the workload of all the robots can be reasonably arranged according to the time, and the daily task amount can be more reasonably processed. The independent variables (principal components) and dependent variables selected by the principal component analysis model are subjected to predictive analysis using multiple linear regression analysis. The algorithm steps of the multiple linear regression analysis are as follows:
setting the dependent variable as y, and screening out principal component x from the principal component analysis model in step 1)1,x2…xkWhen the independent variable is independent variable and the dependent variable is linear, the multiple linear regression model is:
y＝b0+b1x1+b2x2+…bkxk+e
wherein b is0Is a constant term, b1,b2,…bkAs a regression coefficient, b1Is x2,x3…xkAt the time of fixation, x1The effect of each increment of one unit on y, i.e. x1For the partial regression coefficient of y, the other same principles are carried out; the parameter statistics of the multiple regression model requires that the sum of squares of errors is minimum, the least square method is used for solving the parameters, taking the binary linear regression model as an example, and the standard equation set for solving the regression parameters is as follows:
multiple coefficient of determination R is observed in the test of multiple linear regression model2The ratio of the regression equation regression square sum in the total variation of the dependent variable is larger, the degree of fitting of the regression equation to each sample data point is stronger, and the relationship between all independent variables and the dependent variable is more close.
The validity of the whole regression equation is checked, that is, the significance of the whole regression equation is checked, usually, F test is adopted, and the calculation formula of F statistic is:
in the formula, n is an observed value included in the regression, and k is the number of observed variables. Looking up F distribution table according to given significance level alpha and degree of freedom (k, n-k-1) to obtain corresponding critical value FαIf F > FαThe regression equation has significant meaning and the regression effect is significant; otherwise it is not significant.
The variable x in the equation is also tested for validity, and the invention adopts t test (significance test of regression coefficient), wherein the t test is equivalent to the F test in the unitary linear regression, but the equivalence is not established in the multiple linear regression. t test is to test whether each regression coefficient in the regression model has significance, and the statistic t is calculated firstiThen, according to given significance level alpha, degree of freedom n-k-1, looking up t distribution table to obtain critical value tαOr tα/2,t＞tαOr tα/2The regression coefficient is significant, and the statistical quantity formula of t is as follows:
wherein, biIs a regression coefficient, SyIs the standard deviation of the equation, SbiFor the standard deviation of each coefficient, ij is the number of independent variable coefficients, namely the number of equation independent variables, and the equation independent variables comprise 8 dialing list quantity, dialing times, receiving quantity, missed flux, second hang-up quantity, success quantity, average call duration and CCOD channel number.
The multiple collinearity refers to a strong linear relationship between independent variables in a multiple linear regression equation, and if the relationship exceeds the linear relationship between the dependent variables and the independent variables, the stability of a regression model is damaged.
3) Based on the steps 1) and 2), the prediction of the dependent variable y is basically finished, but the invention hopes that the distribution scheme can be flexibly changed at any time according to the prediction result, so that when the historical data is accumulated to a certain extent, the time series + BP neural network combined prediction model is adopted for autonomous machine learning of the data.
3-1, time series ARIMA model
The time series ARIMA model has the following structure:
E(εsεt) 0, arbitrary s<t, wherein xiRepresenting time sequence data, namely the specific dialing time length of the robot for completing the task amount; epsiloniRepresenting a residual term; b represents a delay operator; p represents the autoregressive order; q represents the moving average order; d represents the order of differenceCounting;
representing a difference operator;
expressing an autoregressive coefficient polynomial; the specific expression is as follows:
θ (B) represents a moving average coefficient polynomial, and the specific expression is as follows:
θ(B)＝1-θ1B-θ2B2-…-θqBq
the model satisfies the ARIMA (p, d, q) structure of the model, namely a summation autoregressive moving average model, organically combines an ARMA (p, q) model with differential operation, has a short-term prediction function with higher precision, does not require data to have stronger structurality, only needs to search a rule from the data, and can better fit the data. By utilizing the time sequence ARIMA model and training the dialing duration model in a longer time period, corresponding rules can be automatically learned, for example, the number of lists with the same quantity on weekends and holidays can be reduced sharply, time consumption is reduced correspondingly, outbound of some scenes has certain time rules, and the outbound can also be automatically learned, so that early warning can be performed on key nodes, and personnel supervision and service better promotion are facilitated.
3-2, BP neural network model
The invention also utilizes the BP neural network model to learn the corresponding rule.
The result required by the BP neural network model is related to the time sequence and is different from the back principle of the time sequence ARIMA model, and the two models are weighted, so that the accuracy of the models is improved.
The BP neural network model is an algorithm model for exploratory learning, and the learning process specifically includes the following steps, and a prediction plan is shown in fig. 4:
firstly, initializing a network weight, and then determining specific numerical values of an input data sample and an expected output data sample.
The input data samples constitute factors that affect the desired output data, including dialing duration: dialing list quantity, dialing times, receiving quantity, missed receiving quantity, second hang-up quantity, success quantity, average call duration and CCOD channel number; the expected output data sample is the dialing duration (the time required for the robot to complete the task). I.e. input node 8 and output node 1.
For all connection weights wij、vjiAnd all threshold values thetaj、γiAssignment is made, with the value range being selected to be (-1,1), where wijIs the connection value of the incident layer and the hidden layer, vjiThe connection value of the hidden layer and the output layer is obtained; randomly extracting a set of data input data pairs AiAnd the desired output data pair YiAnd endowing the neural network.
Because the system is nonlinear, the selection of the initial value has great relation to whether the learning reaches the local minimum, whether convergence can be achieved and the length of the training time. The learning speed is influenced when the initial value is too large or too small, so the initial value of the weight value should be selected as a small number empirical value which is uniformly distributed, generally a random number with the initial weight value between (-1,1) is selected, and a random number with the initial weight value between [ -2.4/n,2.4/n ] can also be selected, wherein n is the number of the human input features. To avoid the adjustment direction of the weight value in each step being the same, the initial value should be set as a random number. In the design of a neural network, the network needs to be trained at several different learning rates (0.01-0.8), whether the selected learning rate is proper or not is judged by observing the descending rate of the error square sum after each training, if the descending rate is fast, the learning rate is proper, and if the oscillation phenomenon occurs, the learning rate is overlarge.
Secondly, calculating the input value s of each unit of the middle hidden layer according to the input data sample, the threshold values of all the units of the input layer and the middle hidden layerjThen, according to the transfer function, each intermediate hidden layer is obtainedOutput value b of unitj：
C according to bjCalculating the input value I of each unit in the output layer according to the weight between the middle layer and the output layer and the threshold value of each unit in the output layertThen outputting the actual output value c of each unit of the output layer through a transfer functiont
Fourthly, according to
Calculating correction errors of each unit of output layer
V according toji、
And bjCalculating a correction error
Then combining with a threshold value gammaiAnd calculating the next connection weight value and threshold value of the middle layer and the input layer.
Sixthly according to
And the connection weight and the threshold value of each unit are corrected and calculated to obtain the connection weight and the threshold value between the input layer and the middle layer in the next learning process. And then randomly extracting a data sample to provide for the BP neural network, and learning again according to the method of the second step until all the samples are learned.
And seventhly, randomly selecting a certain sample again, learning according to the second step, and if the global error of the BP neural network is smaller than the preset contrast error value, indicating that the BP neural network is converged. And if the learning times in the whole learning process reach the originally set learning times but the BP neural network is not converged, the network learning process is ended.
4) The time sequence and the BP neural network are good single prediction models, but each prediction model has the characteristics and the application limitation, and the reaction data information has certain difference, so the invention adopts a combined prediction model combining the single prediction models.
4-1, combined prediction model
A common form of the combined prediction model is a weighted average of the individual prediction models, so that the emphasis of the combined prediction model is on the determination of the weighting coefficients. The current commonly used methods comprise an arithmetic mean method, an optimal weight method and the like, and the invention adopts an inverse variance method in consideration of the conditions of real data and the like.
The basic principle is as follows:
firstly, calculating the error square sum e of each single prediction modeljThen, the weight of each single prediction model is assigned according to the principle that the sum of the squares of the overall errors is minimum, and the calculation formula is as follows:
in the formula, m is the number of the single prediction models. The combined prediction model then yields a prediction that can be expressed as:
in the formula (I), the compound is shown in the specification,
and (4) the prediction result of each single prediction model.
According to the invention, through the combination of principal component analysis, multivariate regression analysis, time sequence and BP neural network weighting, a deep machine learning algorithm is brought into distribution management, the application limitation of a single model is greatly reduced, the autonomous machine learning of the neural network is utilized, excessive dependence on experience is avoided, manpower is liberated, and unnecessary artificial monitoring control is reduced. The method comprises the steps of firstly, utilizing a principal component analysis and multivariate regression analysis method to efficiently and comprehensively capture all relevant independent variables and determine appropriate coefficients, and then utilizing a time sequence and BP neural network combined prediction model to improve on the basis of the original prediction method.
While the invention has been described with reference to specific embodiments, the invention is not limited thereto, and those skilled in the art can easily conceive of various equivalent modifications or substitutions within the technical scope of the invention. Therefore, the protection scope of the present invention shall be subject to the protection scope of the claims.
Claims (10)
1. An outbound robot assignment method, comprising the steps of:
1) the client calls and enters different calling skill groups through the selection of the routing module;
2) after entering the calling skill group, predicting the time required by the robot to dial the task amount through principal component analysis and multiple regression analysis methods in sequence;
3) based on the prediction result in the step 2), autonomous machine learning is carried out on the prediction result by utilizing a method of combining a time series model and BP neural network weighting, an optimized learning rule is obtained, and then the robot seats are distributed.
2. The outbound robot assignment method according to claim 1, wherein the specific contents of step 2) are:
21) after entering an incoming call skill group, acquiring relevant variables of an outgoing call, extracting a group of irrelevant variables from the relevant variables by adopting a principal component analysis model, and taking the variables as principal components of the principal component analysis model;
22) judging the number of reserved principal components by checking a correlation coefficient matrix among variables, and screening the reserved principal components;
23) and performing predictive analysis on the principal components and the dependent variables screened out by the principal component analysis model by adopting multivariate linear regression analysis to obtain the dependent variables.
3. The outbound robot distribution method according to claim 2, wherein the outbound related variables are actual production data including robot working duration, actual processing list amount, and dialing success amount in different scenes.
4. The outbound robot assignment method of claim 3, wherein the uncorrelated variables include dialed call order quantity, dialed times, call throughput, missed throughput, second hang-up quantity, work-up quantity, average call duration, and CCOD channel number.
5. The outbound robot assignment method of claim 4, wherein the time series model is a time series ARIMA model.
6. The outbound robot assignment method of claim 5, wherein the structure of the time series ARIMA model is:
in the formula, E (ε)sεt) 0, arbitrary s<t, wherein xiThe time sequence data is the specific dialing time length of the robot for completing the task amount; epsiloniIs a residual term; b is a delay operator; p is the autoregressive order; q is the moving average order; d is the difference order;
is a difference operator;
is an autoregressive coefficient polynomial, and theta (B) is a moving average coefficient polynomial.
7. The outbound robot assignment method of claim 4, wherein the process of autonomous machine learning using the BP neural network model comprises:
a. initializing a network weight, and determining specific numerical values of an input data sample and an expected output data sample, wherein the input data sample is data which affects expected output and comprises a dialing list quantity, dialing times, a receiving quantity, a non-receiving quantity, a second hanging quantity, a success quantity, an average call duration and a CCOD channel number, and the expected output data sample is the dialing duration, namely the time required by the robot to complete a task quantity;
b. for all connection weights wij、vjiAnd all threshold values thetaj、γiCarry out assignment, wherein wijIs the connection value of the incident layer and the hidden layer, vjiThe connection value of the hidden layer and the output layer is obtained; randomly extracting a set of data input data pairs AiAnd the desired output data pair YiEndowing a neural network;
c. calculating the input value s of each unit of the intermediate hidden layer according to the input data sample, the threshold values of all the units of the input layer and the intermediate hidden layerjThen, according to the transfer function, the output value b of each unit of the intermediate hidden layer is obtainedj：
d. According to bjThe weight between the middle layer and the output layer and the threshold value of each unit in the output layer are calculatedInput value I oftOutputting the actual output value c of each unit of the output layer through a transfer functiont；
e. Calculating the correction error of each unit of the output layer according to the expected output data
f. According to vji、
And bjCalculating a correction error
Then combining with a threshold value gammaiCalculating the connection weight and the threshold value of the next middle layer and the input layer;
g. according to
Obtaining the connection weight and the threshold value between the input layer and the middle layer in the next learning process through correction calculation, then randomly extracting a data sample to provide for the BP neural network, and learning again according to the method of the second step until all samples are completely learned;
h. randomly selecting a certain sample again, learning according to the step c, and if the global error of the BP neural network is smaller than the contrast error value which is set originally, indicating that the BP neural network is converged; if the learning times in the whole learning process reach the originally set learning times and the BP neural network is not converged, the network learning process is finished.
8. The outbound robot assignment method according to claim 1, wherein the specific contents of step 3) are:
31) respectively calculating the error square sum of the time series model and the BP neural network model, and assigning the weight of each single prediction model according to the principle of minimum integral error square sum;
32) and acquiring a combined prediction model based on the assignment weights of the individual prediction models to further obtain a final prediction result.
9. The outbound robot assignment method of claim 8, wherein the expression of the prediction result obtained by combining the prediction models is:
in the formula, m is the number of the single-term prediction models, wjFor the weight of each of the individual prediction models,
and (4) the prediction result of each single prediction model.
10. The outbound robot assignment method of claim 9, wherein the weight w of each individual predictive modeljThe expression of (a) is:
in the formula, ejThe sum of the squares of the errors for each of the univariate predictive models.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110772699.3A CN113592250A (en) | 2021-07-08 | 2021-07-08 | Outbound robot distribution method |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202110772699.3A CN113592250A (en) | 2021-07-08 | 2021-07-08 | Outbound robot distribution method |
Publications (1)
| Publication Number | Publication Date |
|---|---|
| CN113592250A true CN113592250A (en) | 2021-11-02 |
Family
ID=78246447
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202110772699.3A Pending CN113592250A (en) | 2021-07-08 | 2021-07-08 | Outbound robot distribution method |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN113592250A (en) |
Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6718330B1 (en) * | 1999-12-16 | 2004-04-06 | Ncr Corporation | Predictive internet automatic work distributor (Pre-IAWD) and proactive internet automatic work distributor (Pro-IAWD) |
| CN105898086A (en) * | 2016-03-31 | 2016-08-24 | 广东亿迅科技有限公司 | Method for predicting calling |
| CN110415013A (en) * | 2019-06-12 | 2019-11-05 | 河海大学 | A kind of combination forecasting method of net about vehicle trip requirements in short-term |
| CN110445942A (en) * | 2019-08-30 | 2019-11-12 | 上海浦东发展银行股份有限公司信用卡中心 | A kind of outer paging system of intelligence and intelligent calling-out method |
-
2021
- 2021-07-08 CN CN202110772699.3A patent/CN113592250A/en active Pending
Patent Citations (4)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US6718330B1 (en) * | 1999-12-16 | 2004-04-06 | Ncr Corporation | Predictive internet automatic work distributor (Pre-IAWD) and proactive internet automatic work distributor (Pro-IAWD) |
| CN105898086A (en) * | 2016-03-31 | 2016-08-24 | 广东亿迅科技有限公司 | Method for predicting calling |
| CN110415013A (en) * | 2019-06-12 | 2019-11-05 | 河海大学 | A kind of combination forecasting method of net about vehicle trip requirements in short-term |
| CN110445942A (en) * | 2019-08-30 | 2019-11-12 | 上海浦东发展银行股份有限公司信用卡中心 | A kind of outer paging system of intelligence and intelligent calling-out method |
Non-Patent Citations (1)
| Title |
|---|
| 傅惠 等: "基于卡尔曼滤波的路径行程时间预测方法", 微计算机信息, vol. 23, no. 8, pages 291 * |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN111428201B (en) | Prediction method for time series data based on empirical mode decomposition and feedforward neural network | |
| CN112200375A (en) | Prediction model generation method, prediction model generation device, and computer-readable medium | |
| CN111222553B (en) | Training data processing method and device of machine learning model and computer equipment | |
| CN111901134B (en) | Method and device for predicting network quality based on recurrent neural network model (RNN) | |
| WO2023273837A1 (en) | Model training method and apparatus, traffic prediction method and apparatus, traffic load balancing method and apparatus, and storage medium | |
| CN112787878B (en) | Network index prediction method and electronic equipment | |
| CN116050540A (en) | Self-adaptive federal edge learning method based on joint bi-dimensional user scheduling | |
| CN113592250A (en) | Outbound robot distribution method | |
| CN116452581B (en) | Intelligent voltage source state detection system and method based on machine vision | |
| CN1926805A (en) | System and method for quality condition analysis of access network supporting broad band telecommunication service | |
| CN116842440A (en) | Self-adaptive link switching method, system, equipment and medium based on context awareness | |
| CN114760215B (en) | Method and system for monitoring data transmission performance of computer network | |
| CN115115131A (en) | Multi-center power system fault prediction method and system based on transfer learning | |
| CN105338198A (en) | Method for computing availability of call center system | |
| CN111080142B (en) | Active service auxiliary judgment method based on power failure reporting | |
| US11475327B2 (en) | Apparatus and method for multivariate prediction of contact center metrics using machine learning | |
| CN110636246B (en) | Maintenance method of video communication cloud hardware terminal | |
| CN114254734A (en) | Traffic matrix modeling method supporting deterministic application | |
| Tabatabaeimehr et al. | Combining long-short term memory and reinforcement learning for improved autonomous network operation | |
| CN112580908A (en) | Wireless performance index evaluation method and device | |
| CN116011602B (en) | Government service management system and method based on Internet | |
| CN112637960B (en) | GNSS-based network node power control method | |
| CN112132190B (en) | Red tide sample data screening method and computer readable storage medium | |
| CN117014555B (en) | Identification method for call prediction of call center | |
| Bezruk et al. | Multicriterial optimization of communication means |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination |