CN104537252A - User state single-classification model training method and device - Google Patents

Abstract

The invention provides a user state single-classification model training method and device. The method includes the steps that at least two positive training samples belonging to the assigned user state class are obtained; each positive training sample comprises at least two items of user attribute information; the sample feature vector of each positive training sample is extracted according to all the user attribute information of each positive training sample; model parameters are estimated according to the sample feature vectors, and a probability density function model is generated according to the estimated model parameters; a user state single-classification model is generated, the user state single-classification model comprises the probability density function model for receiving the input feature vectors and working out function values and further comprises a classification judging model for working out classification results for representing whether the samples belong to the assigned user state class or not according to the obtained function values. The user state single-classification model training method and device are good in classification performance, small in influence of human factors and high in generalization ability.

Description

User Status list disaggregated model training method and device

Technical field

The present invention relates to technical field of computer information processing, particularly relate to a kind of User Status list disaggregated model training method and device.

Background technology

User Status is a kind of description with interim user property, and such as User Status can be student's state, child-bearing state, unmarried state etc.By detecting User Status, diversity service can be provided according to User Status, such as only to there is user's pushed information of specific user's state or providing service, or push different information respectively to the user and do not have with specific user's state or different services is provided.

There is a kind of method of fairly simple detection User Status at present, need the User Status of user's sets itself oneself and store, just can read the User Status of user's setting so when needed to reach the object detecting User Status.But the method for this detection User Status needs user to set User Status by hand, and need user to coordinate, complex operation, feasibility is low.

Also there is a kind of method detecting User Status at present, need to set up a kind of mathematical model of giving a mark in advance, then the behavioral data within the scope of recording user certain hour, the information that user is relevant to the User Status that need detect is found by analytical behavior data, adopt the mathematical model set up in advance to give a mark to each information relevant to User Status, the score value of multiple relevant information is added and obtains total score value.Just can judge whether place has a certain User Status to user by being compared with the total point threshold preset by this total score value.

But at present for detecting the mathematical model of User Status, need artificial setting marking rule, human factor impact is large.And detect User Status by the mode of marking, generalization ability is too weak, the User Status of potential user cannot be detected.Here so-called generalization ability (generalization ability) refers to the adaptive faculty of machine learning algorithm to fresh sample.

Summary of the invention

Based on this, be necessary for the current mathematical model human factor impact for detecting User Status large, and the problem that generalization ability is weak, a kind of User Status list disaggregated model training method and device are provided.

A kind of User Status list disaggregated model training method, described method comprises:

Obtain known at least two the positive training samples belonging to designated user state class; Each positive training sample has at least two customer attribute informations;

According to every customer attribute information of each positive training sample, extract the sampling feature vectors of each positive training sample;

Estimation model parameter is carried out according to described sampling feature vectors, and according to the model parameter generating probability density function model estimated;

Generate User Status list disaggregated model, described User Status list disaggregated model comprises the proper vector for receiving input and calculates the probability density estimation of functional value, also comprises the classification decision model whether belonging to the classification results of described designated user state class for calculating expression according to the functional value calculated.

A kind of User Status list disaggregated model trainer, described device comprises:

Positive training sample acquisition module, for obtaining known at least two the positive training samples belonging to designated user state class; Each positive training sample has at least two customer attribute informations;

Sampling feature vectors extraction module, for the every customer attribute information according to each positive training sample, extracts the sampling feature vectors of each positive training sample;

Model parameter estimation module, for carrying out estimation model parameter according to described sampling feature vectors, and according to the model parameter generating probability density function model estimated;

Training execution module, for generating User Status list disaggregated model, described User Status list disaggregated model comprises the proper vector for receiving input and calculates the probability density estimation of functional value, also comprises the classification decision model whether belonging to the classification results of described designated user state class for calculating expression according to the functional value calculated.

Above-mentioned User Status list disaggregated model training method and device, be different from the training that the both positive and negative training sample that adopts in conventional mode identification method carries out, but trained by the multiple positive training sample belonging to designated user state class and obtain.Like this relative to the disaggregated model adopting the training of positive and negative training sample to obtain, can avoid introducing the impact on classification performance that negative training sample causes, classification performance is better.And after the training of User Status list disaggregated model completes, can reflect the inherent law existed between customer attribute information, human factor impact is very little, and have good predictive ability for the example outside training sample, generalization ability is strong.

Accompanying drawing explanation

Fig. 1 is for realizing the cut-away view of the electronic equipment of User Status list disaggregated model training method in an embodiment;

Fig. 2 is the schematic flow sheet of User Status list disaggregated model training method in an embodiment;

Fig. 3 is the schematic diagram of homogeneous nucleus function in an embodiment;

Fig. 4 is the schematic diagram of normal state kernel function in an embodiment;

Fig. 5 is the sampling feature vectors distribution schematic diagram that in an embodiment, training sample concentrates all positive training samples;

Fig. 6 for finding a hypersphere to surround the schematic diagram of sampling feature vectors in the sampling feature vectors shown in Fig. 5 in an embodiment;

Fig. 7 utilizes hypersphere as shown in Figure 6 to carry out the schematic diagram of classifying in an embodiment;

Fig. 8 is for detecting the schematic flow sheet of the step of the User Status corresponding to user ID to be detected in an embodiment;

Fig. 9 is the schematic flow sheet of the step carrying out estimation model parameter in an embodiment according to sampling feature vectors;

Figure 10 is the schematic flow sheet obtaining the step of the span of model parameter in an embodiment;

Figure 11 is the schematic flow sheet of the step carrying out estimation model parameter in another embodiment according to sampling feature vectors;

Figure 12 is the schematic flow sheet of the step calculating auxiliary intermediate value in an embodiment;

Figure 13 is the structured flowchart of User Status list disaggregated model trainer in an embodiment;

Figure 14 is the structured flowchart of User Status list disaggregated model trainer in another embodiment;

Figure 15 is the structured flowchart of the model parameter estimation module in an embodiment in Figure 13;

Figure 16 is the structured flowchart of User Status list disaggregated model trainer in another embodiment;

Figure 17 is the structured flowchart of the model parameter estimation module in another embodiment in Figure 13;

Figure 18 is the structured flowchart of the auxiliary middle-value calculating module in an embodiment in Figure 17.

Embodiment

In order to make object of the present invention, technical scheme and advantage clearly understand, below in conjunction with drawings and Examples, the present invention is further elaborated.Should be appreciated that specific embodiment described herein only in order to explain the present invention, be not intended to limit the present invention.

As shown in Figure 1, in one embodiment, provide a kind of electronic equipment, this electronic equipment comprises the processor, internal memory, storage medium and the network interface that are connected by system bus.Wherein, the storage medium of this electronic equipment stores operating system, database, also stores a kind of User Status list disaggregated model trainer.This User Status list disaggregated model trainer is for realizing a kind of User Status list disaggregated model training method.The processor of this electronic equipment is configured to perform a kind of User Status list disaggregated model training method.This electronic equipment can be an independently equipment, or can be multiple can interconnected communication electronic equipment composition electronic equipment group, each functional module of User Status list disaggregated model trainer can be deployed on each electronic equipment in electronic equipment group respectively.This electronic equipment can be desk-top computer.

As shown in Figure 2, in one embodiment, a kind of User Status list disaggregated model training method is provided, for training generation one for detecting the User Status list disaggregated model of User Status.Single classification (One-Class-Classification) problem, also can be described as a classification problem, refers to the mark only knowing a certain class sample, judge whether to the data of unknown classification the problem belonging to such.Model then refers to mathematical model, and mathematical model is exactly in order to certain object, and the equation set up by letter, numeral and other mathematic sign or inequality etc. describe the feature of objective things and the mathematic(al) structure expression formula of inner link thereof.User Status list disaggregated model then refer to that training in advance obtains in order to judge whether the proper vector inputted belongs to a kind of mathematical model of designated user state.The electronic equipment that the present embodiment is applied in above-mentioned Fig. 1 in this way illustrates.The method specifically comprises the steps:

Step 202, obtains known at least two the positive training samples belonging to designated user state class; Each positive training sample has at least two customer attribute informations.

Particularly, obtain multiple positive training sample to form training sample set, and each positive training sample have at least two customer attribute informations respectively.In order to ensure the performance of training the User Status list disaggregated model obtained, customer attribute information preferably gets more than 10.Here only adopt positive training sample, and positive training sample refers to the known training sample belonging to designated user state class.

Designated user state is then predefined a kind of User Status, the present embodiment mainly for designated user state for child-bearing state is described, corresponding positive training sample is then the known set belonging to the various customer attribute informations of the user of child-bearing state.Being understandable that, can setting different designated user states according to actual needs, can be such as student's state, unmarried state etc.Every customer attribute information of each positive training sample is all relevant to designated user state.

Every customer attribute information of each positive training sample can take from age of user attribute, user's sex attribute, user's educational background attribute, user take in attribute and the behavioral data relevant to designated user state.Wherein relevant to designated user state behavioral data includes but not limited to the search of added group's quantity, the information content of being correlated with designated user state in social networks, the searching times of information of being correlated with designated user state, the number of clicks of webpage of being correlated with designated user state and the product of being correlated with designated user state of being correlated with designated user state, number of times of browsing, collect, place an order and strike a bargain.

For example, when designated user state is child-bearing state, then the corresponding behavioral data relevant to child-bearing state includes but not limited to: added with give birth to children relevant group's quantity, social networks in collect number of times etc. with place an order number of times, child-bearing Related product conclusion of the business number of times, child-bearing Related product of relevant information content of giving birth to children, the number of clicks with relevant webpage of giving birth to children, the enquirement number of times of being correlated with child-bearing of initiation, child-bearing relevant information searching times, child-bearing Related product number of visits, child-bearing Related product searching times, child-bearing Related product.

Similarly, when designated user state is student's state, then the corresponding behavioral data relevant to student's state includes but not limited to: added to discuss with study information content relevant with study in relevant group's quantity, social networks, the number of clicks of webpage of being correlated with study, initiation with the enquirement number of times learning to be correlated with, learn relevant information searching times, number of times etc. is searched for, browses, collects, places an order and struck a bargain to study article.

Step 204, according to every customer attribute information of each positive training sample, extracts the sampling feature vectors of each positive training sample.

In each customer attribute information of each positive training sample, the value of partial user attributes information is numeric data, in this case just can directly using the element that this numeric data is corresponding in corresponding sampling feature vectors, such as child-bearing Related product number of visits, child-bearing Related product searching times etc.

In each customer attribute information of each positive training sample, the value also having partial user attributes information is not numeric data, but there is the possible case of several limited quantity, in this case quantizes this this part customer attribute information with regard to needs.Specifically several possible cases of customer attribute information can be represented with different numerical value respectively, the overall element corresponding in corresponding sampling feature vectors of the numerical value then customer attribute information quantized.

Such as there are man and female's two kinds of situations in user's sex attribute, and can represent man and female's two kinds of possible cases with 1 and 2 respectively, then a sampling feature vectors can be [1,10,25 ... ], in this sampling feature vectors 1,10 and 25 etc. is all elements of this sampling feature vectors.Wherein in order, the element 1 in this sampling feature vectors represents that sex is man, and element 10 represents the relevant group's quantity that adds and give birth to children, and element 25 represents information content relevant to child-bearing in social networks, etc., by that analogy.

In one embodiment, the customer attribute information not being numeric data is quantized, can with preset length by 0 and 1 the numerical string formed represent customer attribute information, and each of numerical string is respectively as independently element in sampling feature vectors.Preferably in each numerical string, the quantity of 1 is 1.In the present embodiment, consider that several possible cases of the customer attribute information existence not being numeric data are the relations of equality, if only also overall as element corresponding in sampling feature vectors with being quantified as different numerical value, then can the difference of factor value size and the significance level run-off the straight of several possible cases of causing customer attribute information to exist, have influence on the accuracy that User Status list disaggregated model that training obtains carries out classifying.

Illustrate the quantification of the customer attribute information to categorical data.Such as there are man and female's two kinds of possible cases in user's sex attribute, and can represent man and female's two kinds of situations with 10 and 01 respectively, then a sampling feature vectors can be [1,0,10,25 ... ], wherein in order, the first two element 1 in this sampling feature vectors represents that together with 0 sex is man, and element 10 represents the relevant group's quantity that adds and give birth to children, and element 25 represents information content relevant to child-bearing in social networks, etc., by that analogy.

In one embodiment, after step 204, also comprise: each sampling feature vectors extracted is normalized.Consider that different customer attribute information dimensions, dimensional unit are different, directly training obtains User Status list disaggregated model, can have influence on the classification performance of User Status list disaggregated model, be necessary to be normalized.

In one embodiment, be normalized sampling feature vectors, the business that can obtain divided by the difference of greatest member and least member by the difference of each element in sampling feature vectors and least member is as each element in new sampling feature vectors.If such as sampling feature vectors [1,0,10,25 ... ], wherein greatest member is 25, and least member is 0, then the proper vector after normalization is [0.04,0,0.4,1 ... ], each element value in such proper vector, all between 0 ~ 1, conveniently calculates.

In one embodiment, sampling feature vectors is normalized, average and the standard deviation of each element in sampling feature vectors can be calculated, the difference then using each element in sampling feature vectors and this average again divided by the business of this standard deviation as each element in new sampling feature vectors.Certain normalized can also adopt current existing alternate manner, does not enumerate here.

Step 206, carrys out estimation model parameter according to sampling feature vectors, and according to the model parameter generating probability density function model estimated.

Particularly, probability density estimation is a part for User Status list disaggregated model, and for receiving the proper vector of input, and the proper vector calculating this input belongs to the probability of designated user state class.Model parameter is a parameter in probability density estimation, and the fundamental purpose of training obtains this model parameter.

In one embodiment, probability density estimation can adopt Parzen Window (Ba Ersen window) probability density function based on kernel function to generate probability density estimation.Concrete employing kernel function opens a window at each sampling feature vectors place, estimates the probability density at window place.For each sampling feature vectors, maximum to the distribution of self position contribution, from self-position more away from distribution contribution less.

Further, kernel function can choose homogeneous nucleus function and normal state kernel function.Wherein homogeneous nucleus function as shown in Figure 3, and homogeneous nucleus function also can be called rectangle kernel function; Then as shown in Figure 4, normal state kernel function also can be called gaussian kernel function to normal state kernel function.Position in the horizontal ordinate character pair space of kernel function, ordinate is the probability distribution of the proper vector of corresponding position in representation feature space then, for the present embodiment, expression is the probability that in feature space, the proper vector of corresponding position belongs to designated user state class.In gaussian kernel function, e is natural constant, and μ is mathematical expectation, and σ is standard deviation.

With reference to Fig. 5, suppose that training sample concentrates the sampling feature vectors distribution of all positive training samples as shown in Figure 5, so as shown in Figure 6, training obtains the process of probability density estimation, finds a hypersphere exactly and surround these sampling feature vectors in the sampling feature vectors shown in Fig. 5.With reference to Fig. 7, for a proper vector to be detected, if it is within the scope of this hypersphere encirclement, then illustrate that this proper vector belongs to designated user state class, such as proper vector 701; If proper vector to be detected not within the scope of this hypersphere encirclement, then illustrates that this proper vector does not belong to designated user state class, as proper vector 702.

In one embodiment, adopt normal state kernel function to concentrate the sampling feature vectors place of each positive training sample to open a window at training sample, set up Gauss model.Then probability density estimation can represent as shown in following formula (1):

Formula (1): $f\left(y\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\exp (-{(y-x_i)}^T{h}^{-2}(y-x_i));$

Probability density estimation represented by above-mentioned formula (1) be with natural constant be the truth of a matter, respectively with the exponential function that is index of the function between each sampling feature vectors and the proper vector of input and; The transposition that function between each sampling feature vectors and the proper vector of input is respectively the proper vector of input and the difference of corresponding sampling feature vectors is got negative, be multiplied by the negative second power of model parameter again, then be multiplied by the proper vector of input and the difference of corresponding sampling feature vectors.

Particularly, in formula (1), x _ifor training sample concentrates the sampling feature vectors of each positive training sample, i=1,2 ..., n represents sample sequence number.Y represents the proper vector of input, and function f (y) represents the probability density estimation of pre-training User Status list disaggregated model.

The exponential function that what in formula (1), function exp () represented is is the truth of a matter with natural constant e, (-(y-x _i) ^th ^-2(y-x _i)) be the index of this exponential function, be the function between each sampling feature vectors and the proper vector of input.Particularly, (-(y-x _i) ^th ^-2(y-x _i)) be the proper vector y of input and corresponding sampling feature vectors x _idifference (y-x _i) transposition (y-x _i) ^tget to bear and obtain-(y-x _i) ^t, the more negative second power being multiplied by model parameter h is-(y-x _i) ^th ^-2, then be multiplied by the proper vector of input and the difference (y-x of corresponding sampling feature vectors _i).By each sampling feature vectors x _icorresponding (-(y-x _i) ^th ^-2(y-x _i)) summation, obtain $f\left(y\right)=\underset{i=1}{\overset{n}{\mathrm{\Σ}}}\exp (-{(y-x_i)}^T{h}^{-2}(y-x_i)).$

As can be seen from above-mentioned formula (1), by training, the model parameter h estimated in formula (1) just can obtain probability density estimation f (y) of pre-training User Status list disaggregated model.

Step 208, generate User Status list disaggregated model, User Status list disaggregated model comprises the proper vector that receives input and calculates the probability density estimation of functional value, also comprises and calculates according to the functional value calculated the classification decision model whether expression belong to the classification results of designated user state class.

Particularly, the User Status list disaggregated model generated comprises probability density estimation and classification decision model, wherein probability density estimation is for receiving the proper vector of input and calculating functional value, classification decision model is then for calculating classification results according to the functional value calculated, and this classification results represents whether the proper vector of input belongs to designated user state class.If the proper vector of input belongs to designated user state class, illustrate that the proper vector of this input has designated user state, otherwise its User Status can not be judged.

Wherein, decision model of classifying can be expressed as shown in following formula (2):

Formula (2): $\mathrm{\&gamma;}=\left\{\begin{array}{ccc}t\arg \mathrm{et}& \mathrm{if}& f\left(y\right)\&GreaterEqual;\mathrm{\&theta;}\\ \mathrm{outlier}& \mathrm{if}& f\left(y\right)<\mathrm{\&theta;}\end{array}\right.;$

Wherein, in formula (2), y represents the proper vector of input, and be exactly proper vector to be detected when detecting User Status, function f (y) represents the probability density estimation of User Status list disaggregated model.The classification results that γ presentation class decision model exports, target presentation class result γ is for belonging to designated user state class, and outlier then represents and do not belong to designated user state class.θ represents pre-defined function value threshold value, and pre-defined function value threshold value θ is given in advance, can determine according to positive training sample.Pre-defined function value threshold value is given in advance, can determine according to positive training sample, the probability density estimation such as all positive training samples all being inputted pre-training User Status list disaggregated model obtains corresponding functional value, determines pre-defined function value threshold value according to wherein maximum functional value.

Above-mentioned User Status list disaggregated model training method, is different from the training that the both positive and negative training sample that adopts in conventional mode identification method carries out, but is trained by the multiple positive training sample belonging to designated user state class and obtain.Like this relative to the disaggregated model adopting the training of positive and negative training sample to obtain, can avoid introducing the impact on classification performance that negative training sample causes, classification performance is better.And after the training of User Status list disaggregated model completes, can reflect the inherent law existed between customer attribute information, human factor impact is very little, and have good predictive ability for the example outside training sample, generalization ability is strong.

As shown in Figure 8, in one embodiment, this User Status list disaggregated model training method also comprises the step of the User Status detected corresponding to user ID to be detected, specifically comprises the steps:

Step 802, obtains at least two customer attribute informations corresponding to user ID to be detected.

Particularly, user ID is the character string that unique identification goes out user identity, can comprise at least one in the characters such as numeral, symbol and letter.User ID to be detected is then the user ID of the User Status needing to determine its correspondence.Detect User Status in the present embodiment and refer to the whether corresponding designated user state of detection user ID to be detected.

The kind of the customer attribute information corresponding to user ID to be detected, it is the subset of the customer attribute information kind of each positive training sample that training sample is concentrated, even if some customer attribute information item numbers corresponding to user ID to be detected are like this less, as long as but wherein partial user attributes feature relatively significantly still can classify.

At least two customer attribute informations corresponding to user ID to be detected can take from age of user attribute, user's sex attribute, user's educational background attribute, user take in attribute and the behavioral data relevant to designated user state.Wherein relevant to designated user state behavioral data includes but not limited to the search of added group's quantity, the information content of being correlated with designated user state in social networks, the searching times of information of being correlated with designated user state, the number of clicks of webpage of being correlated with designated user state and the product of being correlated with designated user state of being correlated with designated user state, number of times of browsing, collect, place an order and strike a bargain.

Similarly, when designated user state is student's state, then the corresponding behavioral data relevant to student's state includes but not limited to: added to discuss with study information content relevant with study in relevant group's quantity, social networks, the number of clicks of webpage of being correlated with study, initiation with the enquirement number of times learning to be correlated with, learn relevant information searching times, number of times etc. is searched for, browses, collects, places an order and struck a bargain to study article.

Step 804, extracts proper vector to be detected according to the customer attribute information got.

Particularly, adopt the characteristic vector pickup mode adopted during training User Status list disaggregated model, to extract proper vector to be detected according to the customer attribute information got.Specifically, be the customer attribute information of numeric data for the customer attribute information intermediate value got, can directly using the element of this numeric data correspondence in corresponding proper vector to be detected.For the customer attribute information that the customer attribute information intermediate value got is not numeric data, then several possible cases of this customer attribute information can be represented with different numerical value respectively, the overall element corresponding in corresponding proper vector to be detected of the numerical value then this customer attribute information quantized.

In one embodiment, the customer attribute information not being numeric data is quantized, can with preset length by 0 and 1 the numerical string formed represent this customer attribute information, and each of numerical string is respectively as independently element in proper vector to be detected.Preferably in each numerical string, the quantity of 1 is 1.

In one embodiment, after step 804, also comprise: the proper vector to be detected extracted is normalized.Particularly, in one embodiment, be normalized proper vector to be detected, the business that can obtain divided by the difference of greatest member and least member by the difference of each element in proper vector to be detected and least member is as each element in new proper vector to be detected.In another embodiment, proper vector to be detected is normalized, average and the standard deviation of each element in proper vector to be detected can be calculated, then by the difference of each element in proper vector to be detected and this average again divided by the business of this standard deviation as each element in new proper vector to be detected.

Step 806, by proper vector input User Status list disaggregated model to be detected, exports the classification results representing and whether belong to designated user state class, to determine the User Status corresponding to user ID to be detected.

Particularly, by the probability density estimation of proper vector input User Status list disaggregated model to be detected, output function value, then by the classification decision model of the functional value of this output input User Status list disaggregated model, output category result.Such as according to the classification decision model of above-mentioned formula (2), export target and represent and belong to designated user state class, just can determine that the user corresponding to user ID has designated user state according to this; If export outlier, represent and do not belong to designated user state class, the user that now can provide corresponding to user ID does not have designated user state, or cannot determine its User Status.

In one embodiment, after step 806, also comprise: the User Status corresponding to the user ID to be detected determined carrys out pushed information.The information pushed can be advertising message, broadcast notice messages etc.Such as can for detecting that the user being in child-bearing state pushes and relevant advertising message of giving birth to children, such pushed information is more targeted, the validity that guarantee information is passed on.

In the present embodiment, some customer attribute informations corresponding to given user ID, just can determine the User Status corresponding to user ID accordingly, basic without the need to manpower intervention, human factor impact is little, and classification accuracy is compared and adopted the accuracy rate of the mathematical model classification of artificial setting marking rule to want high.

As shown in Figure 9, in one embodiment, carry out the step of estimation model parameter in step 206 according to sampling feature vectors, specifically comprise the steps:

Step 902, is divided into the positive training sample of the first kind and the positive training sample of Equations of The Second Kind by all positive training samples.

Here all positive training samples refer to all positive training sample that training sample is concentrated.Particularly, the positive training sample of the first kind is used to the positive training sample of training, and the positive training sample of Equations of The Second Kind is then for checking or being called test, specifically for testing to the probability density estimation drawn according to the positive training sample of the first kind.The positive training sample of major part in all positive training sample training sample can concentrated is as the positive training sample of the first kind, and remaining positive training sample is then as the positive training sample of Equations of The Second Kind.In such as all positive training samples, 2/3rds as the positive training sample of the first kind, remaining three/first as Equations of The Second Kind training sample.

Step 904, gets the candidate parameter value of predetermined number in the span of model parameter.

Model parameter has span, and this span can rule of thumb set, and also can determine by calculating, can provide the step of the span calculating this model parameter below.In the span of model parameter, a candidate parameter value can be chosen every a fixed step size, form the set of candidate parameter value.Candidate parameter value is the model parameter of hypothesis.

Step 906, generates candidate user state list disaggregated model respectively according to the positive training sample of the first kind and each candidate parameter value, to carry out classification to the positive training sample of Equations of The Second Kind and statistical classification accuracy rate.

Particularly, substituted in above-mentioned formula (1) by the sampling feature vectors of each candidate parameter value and every part of positive training sample of the first kind and obtain corresponding candidate probability density function model, this candidate probability density function model refers to the probability density estimation under the hypothesis of candidate parameter value.

The positive training sample of Equations of The Second Kind is used for checking the classification accuracy of candidate probability density function model, specifically the sampling feature vectors of positive for each Equations of The Second Kind training sample is inputted candidate probability density function model respectively and obtain corresponding functional value, corresponding functional value is inputted the classification decision model represented by above-mentioned formula (2), by the computing of this classification decision model, the relatively size of candidate functions value and pre-defined function value threshold value, output category result.

Corresponding each candidate parameter value calculates classification accuracy respectively, and this classification accuracy is that the positive training sample of Equations of The Second Kind is classified into the quantity of designated user state class and the ratio of the positive training sample sum of Equations of The Second Kind.

Step 908, using candidate parameter value the highest for corresponding classification accuracy as the model parameter estimated.

Particularly, the size of match stop accuracy rate, classification accuracy is higher, and classification performance is better, represent the probability density estimation of corresponding candidate probability density function model close to optimum, using candidate parameter value the highest for the accuracy rate of correspondence as the model parameter estimated.The model parameter this estimated brings probability density estimation function into, thus obtains probability density estimation.

In the present embodiment, by positive training sample being divided into the positive training sample of the first kind and the positive training sample of Equations of The Second Kind, last class is used for training, and a rear class is used for inspection, thus show that algorithm easily realizes close to optimum model parameter.

In one embodiment, step 902 specifically comprises: all positive training samples are divided into default number, in turn using a copy of it as the positive training sample of Equations of The Second Kind, and using remaining positive training sample as the positive training sample of the first kind.For example, the all positive training sample that training sample is concentrated can be divided into 10 parts, each using a copy of it as the positive training sample of Equations of The Second Kind, and remaining 9 parts as the positive training sample of the first kind, then estimate model parameter by continuing to perform step 904 ~ step 908.Like this by crosscheck, the more close optimum model parameter of the model parameter estimated, to obtain close to optimum probability density estimation.

As shown in Figure 10, in one embodiment, this User Status list disaggregated model training method also comprises the step of the span obtaining model parameter, specifically comprises the following steps:

Step 1002, calculates the Mean Matrix of the sampling feature vectors of all positive training samples.

Particularly, according to following formula (3) computation of mean values matrix E:

Formula (3): $E=\frac{1}{n}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}{x}_i;$

In formula (3), n is positive training sample sum, x _ifor training sample concentrates the sampling feature vectors of each positive training sample.Formula (3) represents all sampling feature vectors summations again divided by positive training sample sum.

Step 1004, calculates variance matrix according to Mean Matrix.

Particularly, variance matrix C is calculated according to following formula (4):

Formula (4): $C=\frac{1}{n-1}\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}(E-x_i){(E-x_i)}^T;$

In formula (4), n is positive training sample sum, x _ifor training sample concentrates the sampling feature vectors of each positive training sample, E is the Mean Matrix calculated in step 1002.Formula (4) expression corresponds respectively to each sampling feature vectors, computation of mean values matrix E deducts the product of the difference of this sampling feature vectors and the transposition of this difference, then by after corresponding to product summation that each sampling feature vectors calculates, divided by positive training sample sum with 1 difference.

Step 1006, substitutes into model parameter function to obtain the span of model parameter by variance matrix; Model parameter function is extract square root after the inverse of the dimension of sampling feature vectors is multiplied by the mark of variance matrix, then is multiplied by and adds up to the truth of a matter with positive training sample and get with the business of parametric variable and dimension the exponential function born as index; Parametric variable has default span.

Particularly, bring the variance matrix calculated in step 1004 into model parameter function, wherein model parameter function representation is following formula (5):

$h\left(n\right)={\left(\frac{1}{m}\mathrm{tr}\left(c\right)\right)}^{\frac{1}{2}}{n}^{\frac{-\mathrm{\&alpha;}}{m}},0<\mathrm{\&alpha;}<0.5$

In formula (5), m represents the dimension of sampling feature vectors, and tr (C) represents the mark of variance matrix C, matrix trace refer to element on this diagonal of a matrix and; N is positive training sample sum, and α is parametric variable.Then model parameter function is that mark tr (C) that the inverse of the dimension m of sampling feature vectors is multiplied by variance matrix C extracts square root afterwards and is be multiplied by again and get with the business of parametric variable α and dimension m the exponential function born as index with positive training sample sum n for the truth of a matter be obtain it is 0 < α < 0.5 that parametric variable has default span.

In the present embodiment, the span of first Confirming model parameter, adopt step 902 ~ step 908 according to the model parameter in sampling feature vectors estimated probability density function model again, preferably model parameter can be determined quickly, the classification performance of further optimizing user state list disaggregated model.

As shown in figure 11, in one embodiment, carry out the step of estimation model parameter in step 206 according to sampling feature vectors, specifically comprise the steps:

Step 1102, calculates Euclidean distance value between any two in the sampling feature vectors of all positive training samples.

If sampling feature vectors integrates as X={x ₁, x ₂..., x _n, wherein { x _i| x _i1, x _i2..., x _imrepresent a sampling feature vectors.Integrate in as X at sampling feature vectors and ask for each sampling feature vectors x _iwith each sampling feature vectors x _jeuclidean distance value be D _ij(i ≠ j).

Step 1104, filters out the minimum Eustachian distance value that each sampling feature vectors is corresponding.

Particularly, from Euclidean distance value D _ijthe each sampling feature vectors x of middle correspondence _ifiltering out minimum Eustachian distance value is E _i=min (D _ij).

Step 1106, the square root calculating the maximal value in all minimum Eustachian distance values is using as first candidate family parameter.

Particularly, by each sampling feature vectors x corresponding in step 1104 _ithe minimum Eustachian distance value E filtered out _iin maximal value max (E _i) square root sqrt (max (E _i)) using as first candidate family parameter h ₁.Wherein sqrt () refers to and asks for subduplicate function.Here first and following second etc. describe is existence order.First candidate family parameter is the initial value of iterative computation.

Step 1108, adopts the step calculating auxiliary intermediate value to calculate first auxiliary intermediate value.

Auxiliary intermediate value is used for aided solving candidate family parameter, finally to calculate model parameter.First auxiliary intermediate value is expressed as F ₁.

Step 1110, calculate first candidate family parameter square to be multiplied by positive training sample sum and after the dimension being multiplied by sampling feature vectors adds first candidate family parameter, again divided by after the product of positive training sample sum and dimension, then extract square root, to obtain second candidate family parameter.

Particularly, second candidate family parameter represent first candidate family parameter h ₁square to be multiplied by positive training sample sum n and the dimension m being multiplied by sampling feature vectors adds first auxiliary intermediate value F ₁after, again divided by after the product m of positive training sample sum n and dimension, then extraction of square root obtains to obtain second candidate family parameter h ₂.

Step 1112, adopts the step calculating auxiliary intermediate value to calculate second auxiliary intermediate value.Second auxiliary intermediate value is expressed as F ₂.

Step 1114, adopts candidate family parameter middle-value calculating step to calculate current candidate family parameter intermediate value.

Calculate acquisition first candidate family parameter h ₁, second candidate family parameter h ₂, first auxiliary intermediate value represent F ₁and second auxiliary intermediate value F ₂after, these iteration initial values just can be utilized to carry out iterative computation, to estimate model parameter.H ₁, h ₂, F ₁and F ₂subscript are all sequence numbers.

From sequence number is the 3rd, following formula (6) is adopted to calculate current candidate family parameter intermediate value h _middle:

Formula (6): $h_{\mathrm{middle}}=\frac{(h_{q-1}\&times;h_{q-1}\&times;h_{q-2}\&times;h_{q-2})\&times;(F_{q-1}-F_{q-2})}{F_{q-1}\&times;h_{q-1}\&times;h_{q-1}-F_{q-2}\&times;h_{q-2}\&times;h_{q-2}};$

What represent in formula (6) is candidate family parameter middle-value calculating step, is specially: relative to current candidate family parameter h _qwith current auxiliary intermediate value F _q, calculate previous auxiliary intermediate value F _q-1with previous auxiliary intermediate value F again _q-2the first difference, then be multiplied by previous candidate family parameter h by the first difference _q-1square, then be multiplied by again previous candidate family parameter h _q-2square, then divided by the second difference, this second difference is previous auxiliary intermediate value F _q-1with previous candidate family parameter h _q-1square product, then deduct again previous auxiliary intermediate value F _q-2with previous candidate family parameter h again _q-2square product.Wherein, previous and previous is again relative to the sequence number of candidate family parameter and auxiliary intermediate value, for current sequence number q, is previously expressed as q-1, more previous, is expressed as q-2.

Step 1116, more current candidate family parameter intermediate value and the size of 0; If be less than 0, perform step 1118, if be greater than 0, perform step 1120.

Particularly, more current candidate family parameter intermediate value h _middlewith 0 size.

Step 1118, determines current candidate family parameter according to previous candidate family parameter and previous auxiliary intermediate value.

If current candidate family parameter intermediate value h _middle< 0, then current candidate family parameter is represent that the dimension m that the total n of positive training sample is multiplied by sampling feature vectors is multiplied by previous candidate family parameter h again _q-1square after add previous auxiliary intermediate value h _q-1after, divided by after positive training sample sum n and the product of the dimension m of sampling feature vectors, extract square root.

Step 1120, determines current candidate family parameter according to current candidate family parameter intermediate value.

Particularly, if current candidate family parameter intermediate value h _middle> 0, then current candidate family parameter is h _q=sqrt (h _middle), the current candidate family parameter intermediate value h that expression will calculate according to formula (6) _middleextraction of square root.

Step 1122, judges whether current auxiliary intermediate value and current candidate family parameter meet respective stopping criterion for iteration simultaneously.If so, then perform step 1124, then perform step 1126 if not.

Particularly, current here candidate family parameter h _qstopping criterion for iteration be: | 1-h _q/ h _q-1| >threshold1, wherein threshold1 is first threshold.First threshold threshold1 can be 10 ^-3, or can be 10 ^-3neighbouring value, such as [0.8*10 ^-3, 1.2*10 ^-3] value within the scope of this.| 1-h _q/ h _q-1| >threshold1 represents current candidate family parameter h _qdivided by previous candidate family parameter h _q-1business and 1 the absolute value of difference | 1-h _q/ h _q-1| be greater than first threshold threshold1.

Here current auxiliary intermediate value F _qstopping criterion for iteration be: | 1-F _q/ F _q-1| >threshold2, and | F _q| >threshold3, represents current auxiliary intermediate value F _qdivided by previous auxiliary intermediate value F _q-1business and 1 the absolute value of difference | 1-F _q/ F _q-1| be greater than Second Threshold threshold2, and current auxiliary intermediate value F _qabsolute value be greater than the 3rd threshold value threshold3.Here Second Threshold threshold2 can be 10 ^-4, or can be 10 ^-4neighbouring value, such as [0.8*10 ^-4, 1.2*10 ^-4] value within the scope of this.3rd threshold value threshold3 can be 10 ^-70, or can be 10 ^-70neighbouring value, such as [0.8*10 ^-70, 1.2*10 ^-70] value within the scope of this.

Step 1124, using current candidate family parameter as the model parameter in probability density estimation.

Particularly, as | 1-h _q/ h _q-1| >threshold1, | 1-F _q/ F _q-1| >threshold2 and | F _q| when >threshold3 tri-stopping criterion for iteration all meet, stop iteration, by current candidate family parameter h _qas the model parameter h in probability density estimation, obtain the probability density estimation trained.

Step 1126, returns step 1114 to continue to adopt candidate family parameter middle-value calculating step to calculate next candidate family parameter intermediate value and to determine next candidate family parameter.When | 1-h _q/ h _q-1| >threshold1, | 1-F _q/ F _q-1| >threshold2 and | F _q| when as long as >threshold3 tri-stopping criterion for iteration have one to be false, just return step 1114 and continue iteration, until the auxiliary intermediate value of this next one and this next candidate family parameter meet respective stopping criterion for iteration simultaneously.

In the present embodiment, carry out the model parameter h in calculating probability density function model by iteration convergence method, approach optimum model parameter h gradually, also can calculate preferably model parameter h.

As shown in figure 12, in one embodiment, the step calculating auxiliary intermediate value specifically comprises the steps:

Step 1202, calculate each element in the first intermediary matrix, wherein in the first intermediary matrix, the element at the unequal place of ranks sequence number is set to the Euclidean distance value between corresponding sampling feature vectors, and in the first intermediary matrix, the element at the equal place of ranks sequence number is set to the first preset positive value.

Particularly, the first intermediary matrix DD be line number, columns be positive training sample sum n matrix.When calculating the element of this first intermediary matrix DD, its line order number is i, is corresponding with the sample sequence number of sampling feature vectors, and row sequence number is j.If i ≠ j, then make DD _ij=D _ij, D _ijfor each sampling feature vectors x _iwith each sampling feature vectors x _jeuclidean distance value; If i=j, then make DD _ij=value1.Value1 is the first preset positive value, is a larger numerical value, desirable 10 ⁷⁰or get [0.8*10 ⁷⁰, 1.2*10 ⁷⁰] value within the scope of this.

Step 1204, calculate the difference of the minimum Eustachian distance value corresponding with corresponding sampling feature vectors of each element in the first intermediary matrix, again divided by current candidate family parameter square two times, to generate the element value intermediate value corresponding with each element in the first intermediary matrix.

Particularly, defining element value intermediate value is represent each element DD in the first intermediary matrix DD _ijthe minimum Eustachian distance value E corresponding with corresponding sampling feature vectors _idifference, then divided by current candidate family parameter h _qsquare two times.

Step 1206, structure ranks number consistent with the ranks number of the first intermediary matrix complete zero the second intermediary matrix, if the element value intermediate value corresponding to the element in the first intermediary matrix is less than the second preset positive value, to be then set to element corresponding in the second intermediary matrix with natural constant be the truth of a matter, get with corresponding element value intermediate value the numerical value born as power.

Particularly, the second intermediary matrix P=zeros (n, n) that ranks number is complete zero of positive training sample sum n is constructed.If Y _ij<value2, then P _ij=exp (-Y _ij), if represent the element value intermediate value Y corresponding to element in the first intermediary matrix _ijbe less than the second preset positive value value2, then by element P corresponding in the second intermediary matrix P _ijbe set to exp (-Y _ij), exp (-Y _ij) be take natural constant as the truth of a matter, with corresponding element value intermediate value Y _ijget the numerical value born as power.Here the numerical value in value2 desirable 16 ~ 24, preferably gets 20.

Step 1208, the element of often being gone by the second intermediary matrix is added and obtains adding and value of corresponding often row; If this adds and value equals 0, then calculating the first auxiliary parameter corresponding to this row is the 3rd preset positive value; If this adds and value is not equal to 0, then calculating the first auxiliary parameter corresponding to this row is the inverse that this adds and is worth.

Particularly, calculate if PP _i=0, then FU _i=value3; If PP _i≠ 0, then wherein, PP _ifor the addition of the element of the second intermediary matrix i-th row is obtained capable the adding and value of corresponding i.FU _iit is the first auxiliary parameter that the i-th row is corresponding.Value3 is the 3rd preset positive value, desirable 1.7977 × 10 ³⁰⁸or get [0.8*1.7977 × 10 ³⁰⁸, 1.2*1.7977 × 10 ³⁰⁸] numerical value within the scope of this.

Step 1210, sues for peace after the element multiplication of every row corresponding position in each element of the first intermediary matrix and the second intermediary matrix, to obtain the second auxiliary parameter of corresponding often row.

Particularly, calculate wherein FF _ibe the second auxiliary parameter that the i-th row is corresponding, DD _ijbe the element of the i-th row in the first intermediary matrix DD, jth row, P _ijit is the element of the i-th row in the second intermediary matrix P, jth row.

Step 1212, summation after respectively the first corresponding for often row auxiliary parameter being multiplied with the second auxiliary parameter, again divided by current candidate family parameter square after, then deduct the product of positive training sample sum and the dimension of sampling feature vectors, using as current auxiliary intermediate value.

Particularly, calculate wherein F _qfor current auxiliary intermediate value.Calculate current auxiliary intermediate value F _qtime, respectively by the first corresponding for often row auxiliary parameter FU _iwith the second auxiliary parameter FF _isue for peace after being multiplied, then divided by current candidate family parameter h _qsquare after, then deduct the product of dimension m of positive training sample sum n and sampling feature vectors, obtain current auxiliary intermediate value F _q.

In the present embodiment, provide the step calculating auxiliary intermediate value, for when being carried out the model parameter h in calculating probability density function model by iteration convergence method, provide the auxiliary intermediate value needed for iterative computation.

As shown in figure 13, in one embodiment, provide a kind of User Status list disaggregated model trainer 1300, there is the function of the User Status list disaggregated model training method realizing each embodiment above-mentioned.This User Status list disaggregated model trainer 1300 comprises: positive training sample acquisition module 1310, sampling feature vectors extraction module 1320, model parameter estimation module 1330 and training execution module 1340.

Positive training sample acquisition module 1310, for obtaining known at least two the positive training samples belonging to designated user state class; Each positive training sample has at least two customer attribute informations.

Particularly, positive training sample acquisition module 1310 is for obtaining multiple positive training sample to form training sample set, and each positive training sample has at least two customer attribute informations respectively.In order to ensure the performance of training the User Status list disaggregated model obtained, customer attribute information preferably gets more than 10.Here only adopt positive training sample, and positive training sample refers to the known training sample belonging to designated user state class.

Designated user state is then predefined a kind of User Status, the present embodiment mainly for designated user state for child-bearing state is described, corresponding positive training sample is then the known set belonging to the various customer attribute informations of the user of child-bearing state.Being understandable that, can setting different designated user states according to actual needs, can be such as student's state, unmarried state etc.Every customer attribute information of each positive training sample is all relevant to designated user state.

Every customer attribute information of each positive training sample can take from age of user attribute, user's sex attribute, user's educational background attribute, user take in attribute and the behavioral data relevant to designated user state.Wherein relevant to designated user state behavioral data includes but not limited to the search of added group's quantity, the information content of being correlated with designated user state in social networks, the searching times of information of being correlated with designated user state, the number of clicks of webpage of being correlated with designated user state and the product of being correlated with designated user state of being correlated with designated user state, number of times of browsing, collect, place an order and strike a bargain.

Similarly, when designated user state is student's state, then the corresponding behavioral data relevant to student's state includes but not limited to: added to discuss with study information content relevant with study in relevant group's quantity, social networks, the number of clicks of webpage of being correlated with study, initiation with the enquirement number of times learning to be correlated with, learn relevant information searching times, number of times etc. is searched for, browses, collects, places an order and struck a bargain to study article.

Sampling feature vectors extraction module 1320, for the every customer attribute information according to each positive training sample, extracts the sampling feature vectors of each positive training sample.

In each customer attribute information of each positive training sample, the value of partial user attributes information is numeric data, in this case sampling feature vectors extraction module 1320 may be used for directly using the element that this numeric data is corresponding in corresponding sampling feature vectors, such as child-bearing Related product number of visits, child-bearing Related product searching times etc.

In each customer attribute information of each positive training sample, the value also having partial user attributes information is not numeric data, but there is the possible case of several limited quantity, sampling feature vectors extraction module 1320 may be used for quantizing this this part customer attribute information in this case.Specifically several possible cases of customer attribute information can be represented with different numerical value respectively, the overall element corresponding in corresponding sampling feature vectors of the numerical value then customer attribute information quantized.

In one embodiment, the customer attribute information not being numeric data is quantized, can with preset length by 0 and 1 the numerical string formed represent customer attribute information, and each of numerical string is respectively as independently element in sampling feature vectors.Preferably in each numerical string, the quantity of 1 is 1.In the present embodiment, consider that several possible cases of the customer attribute information existence not being numeric data are the relations of equality, if only also overall as element corresponding in sampling feature vectors with being quantified as different numerical value, then can the difference of factor value size and the significance level run-off the straight of several possible cases of causing customer attribute information to exist, have influence on the accuracy that User Status list disaggregated model that training obtains carries out classifying.

In one embodiment, each sampling feature vectors that sampling feature vectors extraction module 1320 can also be used for extracting is normalized.Consider that different customer attribute information dimensions, dimensional unit are different, directly training obtains User Status list disaggregated model, can have influence on the classification performance of User Status list disaggregated model, be necessary to be normalized.

In one embodiment, be normalized sampling feature vectors, the business that can obtain divided by the difference of greatest member and least member by the difference of each element in sampling feature vectors and least member is as each element in new sampling feature vectors.If such as sampling feature vectors [1,0,10,25 ... ], wherein greatest member is 25, and least member is 0, then the proper vector after normalization is [0.04,0,0.4,1 ... ], each element value in such proper vector, all between 0 ~ 1, conveniently calculates.

In one embodiment, sampling feature vectors is normalized, average and the standard deviation of each element in sampling feature vectors can be calculated, the difference then using each element in sampling feature vectors and this average again divided by the business of this standard deviation as each element in new sampling feature vectors.Certain normalized can also adopt current existing alternate manner, does not enumerate here.

Model parameter estimation module 1330, for carrying out estimation model parameter according to sampling feature vectors, and according to the model parameter generating probability density function model estimated.

Particularly, probability density estimation is a part for User Status list disaggregated model, and for receiving the proper vector of input, and the proper vector calculating this input belongs to the probability of designated user state class.Model parameter is a parameter in probability density estimation, and the fundamental purpose of training obtains this model parameter.

In one embodiment, probability density estimation can adopt the Parzen Window probability density function based on kernel function to generate probability density estimation.Concrete employing kernel function opens a window at each sampling feature vectors place, estimates the probability density at window place.For each sampling feature vectors, maximum to the distribution of self position contribution, from self-position more away from distribution contribution less.

Further, kernel function can choose homogeneous nucleus function and normal state kernel function.Wherein homogeneous nucleus function as shown in Figure 3, and homogeneous nucleus function also can be called rectangle kernel function; Then as shown in Figure 4, normal state kernel function also can be called gaussian kernel function to normal state kernel function.Position in the horizontal ordinate character pair space of kernel function, ordinate is the probability distribution of the proper vector of corresponding position in representation feature space then, for the present embodiment, expression is the probability that in feature space, the proper vector of corresponding position belongs to designated user state class.In gaussian kernel function, e is natural constant, and μ is mathematical expectation, and σ is standard deviation.

With reference to Fig. 5, suppose that training sample concentrates the sampling feature vectors distribution of all positive training samples as shown in Figure 5, so as shown in Figure 6, training obtains the process of probability density estimation, finds a hypersphere exactly and surround these sampling feature vectors in the sampling feature vectors shown in Fig. 5.With reference to Fig. 7, for a proper vector to be detected, if it is within the scope of this hypersphere encirclement, then illustrate that this proper vector belongs to designated user state class, such as proper vector 701; If proper vector to be detected not within the scope of this hypersphere encirclement, then illustrates that this proper vector does not belong to designated user state class, as proper vector 702.

In one embodiment, adopt normal state kernel function to concentrate the sampling feature vectors place of each positive training sample to open a window at training sample, set up Gauss model.Then probability density estimation can represent as shown in following formula (1):

Formula (1): $f\left(y\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\exp (-{(y-x_i)}^T{h}^{-2}(y-x_i));$

Probability density estimation represented by above-mentioned formula (1) be with natural constant be the truth of a matter, respectively with the exponential function that is index of the function between each sampling feature vectors and the proper vector of input and; The transposition that function between each sampling feature vectors and the proper vector of input is respectively the proper vector of input and the difference of corresponding sampling feature vectors is got negative, be multiplied by the negative second power of model parameter again, then be multiplied by the proper vector of input and the difference of corresponding sampling feature vectors.

Particularly, in formula (1), x _ifor training sample concentrates the sampling feature vectors of each positive training sample, i=1,2 ..., n represents sample sequence number.Y represents the proper vector of input, and function f (y) represents the probability density estimation of pre-training User Status list disaggregated model.

The exponential function that what in formula (1), function exp () represented is is the truth of a matter with natural constant e, (-(y-x _i) ^th ^-2(y-x _i)) be the index of this exponential function, be the function between each sampling feature vectors and the proper vector of input.Particularly, (-(y-x _i) ^th ^-2(y-x _i)) be the proper vector y of input and corresponding sampling feature vectors x _idifference (y-x _i) transposition (y-x _i) ^tget to bear and obtain-(y-x _i) ^t, the more negative second power being multiplied by model parameter h is-(y-x _i) ^th ^-2, then be multiplied by the proper vector of input and the difference (y-x of corresponding sampling feature vectors _i).By each sampling feature vectors x _icorresponding (-(y-x _i) ^th ^-2(y-x _i)) summation, obtain $f\left(y\right)=\underset{i=1}{\overset{n}{\mathrm{\&Sigma;}}}\exp (-{(y-x_i)}^T{h}^{-2}(y-x_i)).$

As can be seen from above-mentioned formula (1), by training, the model parameter h estimated in formula (1) just can obtain probability density estimation f (y) of pre-training User Status list disaggregated model.

Training execution module 1340, for generating User Status list disaggregated model, User Status list disaggregated model comprises the proper vector for receiving input and calculates the probability density estimation of functional value, also comprises the classification decision model whether belonging to the classification results of designated user state class for calculating expression according to the functional value calculated.

Particularly, the User Status list disaggregated model that training execution module 1340 generates comprises probability density estimation and classification decision model, wherein probability density estimation is for receiving the proper vector of input and calculating functional value, classification decision model is then for calculating classification results according to the functional value calculated, and this classification results represents whether the proper vector of input belongs to designated user state class.If the proper vector of input belongs to designated user state class, illustrate that the proper vector of this input has designated user state, otherwise its User Status can not be judged.

Wherein, decision model of classifying can be expressed as shown in following formula (2):

Formula (2): $\mathrm{\&gamma;}=\left\{\begin{array}{ccc}t\arg \mathrm{et}& \mathrm{if}& f\left(y\right)\&GreaterEqual;\mathrm{\&theta;}\\ \mathrm{outlier}& \mathrm{if}& f\left(y\right)<\mathrm{\&theta;}\end{array}\right.;$

Wherein, in formula (2), y represents the proper vector of input, and be exactly proper vector to be detected when detecting User Status, function f (y) represents the probability density estimation of User Status list disaggregated model.The classification results that γ presentation class decision model exports, target presentation class result γ is for belonging to designated user state class, and outlier then represents and do not belong to designated user state class.θ represents pre-defined function value threshold value, and pre-defined function value threshold value θ is given in advance, can determine according to positive training sample.Pre-defined function value threshold value is given in advance, can determine according to positive training sample, the probability density estimation such as all positive training samples all being inputted pre-training User Status list disaggregated model obtains corresponding functional value, determines pre-defined function value threshold value according to wherein maximum functional value.

Above-mentioned User Status list disaggregated model trainer 1300, is different from the training that the both positive and negative training sample that adopts in conventional mode identification method carries out, but is trained by the multiple positive training sample belonging to designated user state class and obtain.Like this relative to the disaggregated model adopting the training of positive and negative training sample to obtain, can avoid introducing the impact on classification performance that negative training sample causes, classification performance is better.And after the training of User Status list disaggregated model completes, can reflect the inherent law existed between customer attribute information, human factor impact is very little, and have good predictive ability for the example outside training sample, generalization ability is strong.

As shown in figure 14, in one embodiment, this User Status list disaggregated model trainer 1300 also comprises: customer attribute information acquisition module 1350, characteristic vector pickup module 1360 to be detected and sort module 1370.

Customer attribute information acquisition module 1350, for obtaining at least two customer attribute informations corresponding to user ID to be detected.

Particularly, user ID is the character string that unique identification goes out user identity, can comprise at least one in the characters such as numeral, symbol and letter.User ID to be detected is then the user ID of the User Status needing to determine its correspondence.

The kind of the customer attribute information corresponding to user ID to be detected, it is the subset of the customer attribute information kind of each positive training sample that training sample is concentrated, even if some customer attribute information item numbers corresponding to user ID to be detected are like this less, as long as but wherein partial user attributes feature relatively significantly still can classify.

At least two customer attribute informations corresponding to user ID to be detected can take from age of user attribute, user's sex attribute, user's educational background attribute, user take in attribute and the behavioral data relevant to designated user state.Wherein relevant to designated user state behavioral data includes but not limited to the search of added group's quantity, the information content of being correlated with designated user state in social networks, the searching times of information of being correlated with designated user state, the number of clicks of webpage of being correlated with designated user state and the product of being correlated with designated user state of being correlated with designated user state, number of times of browsing, collect, place an order and strike a bargain.

Similarly, when designated user state is student's state, then the corresponding behavioral data relevant to student's state includes but not limited to: added to discuss with study information content relevant with study in relevant group's quantity, social networks, the number of clicks of webpage of being correlated with study, initiation with the enquirement number of times learning to be correlated with, learn relevant information searching times, number of times etc. is searched for, browses, collects, places an order and struck a bargain to study article.

Characteristic vector pickup module 1360 to be detected, for extracting proper vector to be detected according to the customer attribute information got.

Particularly, adopt the characteristic vector pickup mode adopted during training User Status list disaggregated model, to extract proper vector to be detected according to the customer attribute information got.Specifically, be the customer attribute information of numeric data for the customer attribute information intermediate value got, can directly using the element of this numeric data correspondence in corresponding proper vector to be detected.For the customer attribute information that the customer attribute information intermediate value got is not numeric data, then several possible cases of this customer attribute information can be represented with different numerical value respectively, the overall element corresponding in corresponding proper vector to be detected of the numerical value then this customer attribute information quantized.

In one embodiment, the customer attribute information not being numeric data is quantized, can with preset length by 0 and 1 the numerical string formed represent this customer attribute information, and each of numerical string is respectively as independently element in proper vector to be detected.Preferably in each numerical string, the quantity of 1 is 1.

In one embodiment, characteristic vector pickup module 1360 to be detected is also for being normalized the proper vector to be detected extracted.Particularly, be normalized proper vector to be detected, the business that can obtain divided by the difference of greatest member and least member by the difference of each element in proper vector to be detected and least member is as each element in new proper vector to be detected.Proper vector to be detected is normalized, average and the standard deviation of each element in proper vector to be detected can be calculated, then by the difference of each element in proper vector to be detected and this average again divided by the business of this standard deviation as each element in new proper vector to be detected.

Sort module 1370, for by proper vector input User Status list disaggregated model to be detected, exports the classification results representing and whether belong to designated user state class, to determine the User Status corresponding to user ID to be detected.

Particularly, by the probability density estimation of proper vector input User Status list disaggregated model to be detected, output function value, then by the classification decision model of the functional value of this output input User Status list disaggregated model, output category result.Such as according to the classification decision model of above-mentioned formula (2), export target and represent and belong to designated user state class, just can determine that the user corresponding to user ID has designated user state according to this; If export outlier, represent and do not belong to designated user state class, the user that now can provide corresponding to user ID does not have designated user state, or cannot determine its User Status.

In one embodiment, this User Status list disaggregated model trainer 1300 also comprises pushing module (not shown), carrys out pushed information for the User Status corresponding to the user ID to be detected determined.The information pushed can be advertising message, broadcast notice messages etc.Such as can for detecting that the user being in child-bearing state pushes and relevant advertising message of giving birth to children, such pushed information is more targeted, the validity that guarantee information is passed on.

In the present embodiment, some customer attribute informations corresponding to given user ID, just can determine the User Status corresponding to user ID accordingly, basic without the need to manpower intervention, human factor impact is little, and classification accuracy is compared and adopted the accuracy rate of the mathematical model classification of artificial setting marking rule to want high.

As shown in figure 15, in one embodiment, model parameter estimation module 1330 comprises: sample divides module 1331, candidate parameter value chooses module 1332, statistic of classification module 1333 and model parameter determination module 1334.

Sample divides module 1331, for all positive training samples are divided into the positive training sample of the first kind and the positive training sample of Equations of The Second Kind.

Here all positive training samples refer to all positive training sample that training sample is concentrated.Particularly, the positive training sample of the first kind is used to the positive training sample of training, and the positive training sample of Equations of The Second Kind is then for checking or being called test, specifically for testing to the probability density estimation drawn according to the positive training sample of the first kind.The positive training sample of major part in all positive training sample training sample can concentrated is as the positive training sample of the first kind, and remaining positive training sample is then as the positive training sample of Equations of The Second Kind.In such as all positive training samples, 2/3rds as the positive training sample of the first kind, remaining three/first as Equations of The Second Kind training sample.

Candidate parameter value chooses module 1332, for getting the candidate parameter value of predetermined number in the span of model parameter.

Model parameter has span, and this span can rule of thumb set, and also can determine by calculating, can provide the step of the span calculating this model parameter below.In the span of model parameter, a candidate parameter value can be chosen every a fixed step size, form the set of candidate parameter value.Candidate parameter value is the model parameter of hypothesis.

Statistic of classification module 1333, for generating candidate user state list disaggregated model respectively according to the positive training sample of the first kind and each candidate parameter value, to carry out classification and statistical classification accuracy rate to the positive training sample of Equations of The Second Kind.

Particularly, substituted in above-mentioned formula (1) by the sampling feature vectors of each candidate parameter value and every part of positive training sample of the first kind and obtain corresponding candidate probability density function model, this candidate probability density function model refers to the probability density estimation under the hypothesis of candidate parameter value.

The positive training sample of Equations of The Second Kind is used for checking the classification accuracy of candidate probability density function model, specifically the sampling feature vectors of positive for each Equations of The Second Kind training sample is inputted candidate probability density function model respectively and obtain corresponding functional value, corresponding functional value is inputted the classification decision model represented by above-mentioned formula (2), by the computing of this classification decision model, the relatively size of candidate functions value and pre-defined function value threshold value, output category result.

Corresponding each candidate parameter value calculates classification accuracy respectively, and this classification accuracy is that the positive training sample of Equations of The Second Kind is classified into the quantity of designated user state class and the ratio of the positive training sample sum of Equations of The Second Kind.

Model parameter determination module 1334, for using candidate parameter value the highest for corresponding classification accuracy as the model parameter estimated.

Particularly, the size of match stop accuracy rate, classification accuracy is higher, and classification performance is better, represent the probability density estimation of corresponding candidate probability density function model close to optimum, using candidate parameter value the highest for the accuracy rate of correspondence as the model parameter estimated.The model parameter this estimated brings probability density estimation function into, thus obtains probability density estimation.

In the present embodiment, by positive training sample being divided into the positive training sample of the first kind and the positive training sample of Equations of The Second Kind, last class is used for training, and a rear class is used for inspection, thus show that algorithm easily realizes close to optimum model parameter.

In one embodiment, sample divides module 1331 specifically for all positive training samples are divided into default number, in turn using a copy of it as the positive training sample of Equations of The Second Kind, and using remaining positive training sample as the positive training sample of the first kind.

As shown in figure 16, in one embodiment, this User Status list disaggregated model trainer 1300 also comprises: Mean Matrix computing module 1380, variance matrix computing module 1390 and model parameter span computing module 1399.

Mean Matrix computing module 1380, for calculating the Mean Matrix of the sampling feature vectors of all positive training samples.

Variance matrix computing module 1390, for calculating variance matrix according to Mean Matrix.

Model parameter span computing module 1399, for substituting into model parameter function to obtain the span of model parameter by variance matrix; Model parameter function is extract square root after the inverse of the dimension of sampling feature vectors is multiplied by the mark of variance matrix, then is multiplied by and adds up to the truth of a matter with positive training sample and get with the business of parametric variable and dimension the exponential function born as index; Parametric variable has default span.

As shown in figure 17, in another embodiment, model parameter estimation module 1330 comprises: Euclidean distance value computing module 1330a, screening module 1330b, first candidate family parameter calculating module 1330c, first auxiliary middle-value calculating module 1330d, second candidate family parameter calculating module 1330e, second auxiliary middle-value calculating module 1330f, current candidate model parameter middle-value calculating module 1330g, current candidate model parameter determination module 1330h, judge module 1330i, model parameter determination module 1330j and iterative computation module 1330k.

Euclidean distance value computing module 1330a, for calculate all positive training samples sampling feature vectors in Euclidean distance value between any two.

Screening module 1330b, for filtering out minimum Eustachian distance value corresponding to each sampling feature vectors.

First candidate family parameter calculating module 1330c, for the square root that calculates the maximal value in all minimum Eustachian distance values using as first candidate family parameter.

First auxiliary middle-value calculating module 1330d, for passing through auxiliary middle-value calculating module 1335 to calculate first auxiliary intermediate value.Auxiliary middle-value calculating module 1335 can belong to User Status list disaggregated model trainer 1300, also can be independently module.

Second candidate family parameter calculating module 1330e, for calculate first candidate family parameter square to be multiplied by positive training sample sum and after the dimension being multiplied by sampling feature vectors adds first candidate family parameter, again divided by after the product of positive training sample sum and dimension, extract square root again, to obtain second candidate family parameter.

Second auxiliary middle-value calculating module 1330f, for calculating second auxiliary intermediate value by auxiliary middle-value calculating module 1335.

Current candidate model parameter middle-value calculating module 1330g, for calculating current candidate family parameter intermediate value by candidate family parameter median operation module 1336.Candidate family parameter median operation module 1336 can belong to User Status list disaggregated model trainer 1300, also can be independently module.

Current candidate model parameter determination module 1330h, if be less than 0 for current candidate family parameter intermediate value, then determines current candidate family parameter according to previous candidate family parameter and previous auxiliary intermediate value; If current candidate family parameter intermediate value is greater than 0, then determine current candidate family parameter according to current candidate family parameter intermediate value.

Judge module 1330i, for judging whether current auxiliary intermediate value and current candidate family parameter meet respective stopping criterion for iteration simultaneously.

Model parameter determination module 1330j, if meet respective stopping criterion for iteration for current auxiliary intermediate value and current candidate family parameter simultaneously, then using current candidate family parameter as the model parameter in probability density estimation.

Iterative computation module 1330k, if meet respective stopping criterion for iteration time different with current candidate family parameter for current auxiliary intermediate value, then notify that current candidate model parameter middle-value calculating module 1330g is to continue through candidate family parameter median operation module 1336 to calculate next candidate family parameter intermediate value, and notify that current candidate model parameter determination module 1330h determines next candidate family parameter, until the auxiliary intermediate value of this next one and this next candidate family parameter meet respective stopping criterion for iteration simultaneously.

In one embodiment, candidate family parameter median operation module 1336 is for relative to current candidate family parameter and current auxiliary intermediate value, calculate the first difference of previous auxiliary intermediate value and previous auxiliary intermediate value again, again with the first difference be multiplied by previous candidate family parameter square, be multiplied by again more previous candidate family parameter square, again divided by the second difference, this second difference be previous auxiliary intermediate value and previous candidate family parameter square product, deduct again more previous auxiliary intermediate value and previous candidate family parameter again square product.

As shown in figure 18, in one embodiment, auxiliary middle-value calculating module 1335 comprises: the first intermediary matrix processing module 1335a, element value middle-value calculating module 1335b, the second intermediary matrix constructing module 1335c, the first auxiliary parameter computing module 1335d, the second auxiliary parameter computing module 1335e and auxiliary intermediate value generation module 1335f.

First intermediary matrix processing module 1335a, for calculating each element in the first intermediary matrix, wherein in the first intermediary matrix, the element at the unequal place of ranks sequence number is set to the Euclidean distance value between corresponding sampling feature vectors, and in the first intermediary matrix, the element at the equal place of ranks sequence number is set to the first preset positive value.

Element value middle-value calculating module 1335b, for calculating the difference of the minimum Eustachian distance value corresponding with corresponding sampling feature vectors of each element in the first intermediary matrix, again divided by current candidate family parameter square two times, to generate the element value intermediate value corresponding with each element in the first intermediary matrix.

Second intermediary matrix constructing module 1335c, for construct ranks number consistent with the ranks number of the first intermediary matrix complete zero the second intermediary matrix, if the element value intermediate value corresponding to the element in the first intermediary matrix is less than the second preset positive value, to be then set to element corresponding in the second intermediary matrix with natural constant be the truth of a matter, get with corresponding element value intermediate value the numerical value born as power.

First auxiliary parameter computing module 1335d, the element for often being gone by the second intermediary matrix is added and obtains adding and value of corresponding often row; If this adds and value equals 0, then calculating the first auxiliary parameter corresponding to this row is the 3rd preset positive value; If this adds and value is not equal to 0, then calculating the first auxiliary parameter corresponding to this row is the inverse that this adds and is worth.

Second auxiliary parameter computing module 1335e, for suing for peace after the element multiplication of every row corresponding position in each element of the first intermediary matrix and the second intermediary matrix, to obtain the second auxiliary parameter of corresponding often row.

Auxiliary intermediate value generation module 1335f, sue for peace after respectively the first corresponding for often row auxiliary parameter being multiplied with the second auxiliary parameter, again divided by current candidate family parameter square after, deduct the product of positive training sample sum and the dimension of sampling feature vectors again, using as current auxiliary intermediate value.

One of ordinary skill in the art will appreciate that all or part of flow process realized in above-described embodiment method, that the hardware that can carry out instruction relevant by computer program has come, described program can be stored in a computer read/write memory medium, this program, when performing, can comprise the flow process of the embodiment as above-mentioned each side method.Wherein, described storage medium can be magnetic disc, CD, read-only store-memory body (Read-Only Memory, ROM) or random store-memory body (Random Access Memory, RAM) etc.

The above embodiment only have expressed several embodiment of the present invention, and it describes comparatively concrete and detailed, but therefore can not be interpreted as the restriction to the scope of the claims of the present invention.It should be pointed out that for the person of ordinary skill of the art, without departing from the inventive concept of the premise, can also make some distortion and improvement, these all belong to protection scope of the present invention.Therefore, the protection domain of patent of the present invention should be as the criterion with claims.

Claims (10)

1. a User Status list disaggregated model training method, described method comprises:

Obtain known at least two the positive training samples belonging to designated user state class; Each positive training sample has at least two customer attribute informations;

According to every customer attribute information of each positive training sample, extract the sampling feature vectors of each positive training sample;

Estimation model parameter is carried out according to described sampling feature vectors, and according to the model parameter generating probability density function model estimated;

Generate User Status list disaggregated model, described User Status list disaggregated model comprises the proper vector for receiving input and calculates the probability density estimation of functional value, also comprises the classification decision model whether belonging to the classification results of described designated user state class for calculating expression according to the functional value calculated.

2. method according to claim 1, is characterized in that, described method also comprises:

Obtain at least two customer attribute informations corresponding to user ID to be detected;

Proper vector to be detected is extracted according to the customer attribute information got;

Described proper vector to be detected is inputted described User Status list disaggregated model, exports the classification results representing and whether belong to designated user state class, to determine the User Status corresponding to described user ID to be detected.

3. method according to claim 1, is characterized in that, describedly carrys out estimation model parameter according to described sampling feature vectors, comprising:

All positive training samples are divided into the positive training sample of the first kind and the positive training sample of Equations of The Second Kind;

The candidate parameter value of predetermined number is got in the span of model parameter;

Candidate user state list disaggregated model is generated respectively, to carry out classification to the positive training sample of Equations of The Second Kind and statistical classification accuracy rate according to the positive training sample of the first kind and each candidate parameter value;

Using candidate parameter value the highest for corresponding classification accuracy as the model parameter estimated.

4. method as claimed in any of claims 1 to 3, it is characterized in that, described probability density estimation be with natural constant be the truth of a matter, respectively with the exponential function that is index of the function between each sampling feature vectors and the proper vector of input and; The transposition that function between each sampling feature vectors and the proper vector of input is respectively the proper vector of input and the difference of corresponding sampling feature vectors is got negative, be multiplied by the negative second power of model parameter again, then be multiplied by the proper vector of input and the difference of corresponding sampling feature vectors.

5. method according to claim 4, is characterized in that, described method also comprises:

Calculate the Mean Matrix of the sampling feature vectors of all positive training samples;

Variance matrix is calculated according to described Mean Matrix;

Described variance matrix is substituted into model parameter function to obtain the span of model parameter; Described model parameter function is extract square root after the inverse of the dimension of sampling feature vectors is multiplied by the mark of variance matrix, then is multiplied by and adds up to the truth of a matter with positive training sample and get with the business of parametric variable and described dimension the exponential function born as index; Described parametric variable has default span.

6. a User Status list disaggregated model trainer, is characterized in that, described device comprises:

Positive training sample acquisition module, for obtaining known at least two the positive training samples belonging to designated user state class; Each positive training sample has at least two customer attribute informations;

Sampling feature vectors extraction module, for the every customer attribute information according to each positive training sample, extracts the sampling feature vectors of each positive training sample;

Model parameter estimation module, for carrying out estimation model parameter according to described sampling feature vectors, and according to the model parameter generating probability density function model estimated;

Training execution module, for generating User Status list disaggregated model, described User Status list disaggregated model comprises the proper vector for receiving input and calculates the probability density estimation of functional value, also comprises the classification decision model whether belonging to the classification results of described designated user state class for calculating expression according to the functional value calculated.

7. device according to claim 6, is characterized in that, described device also comprises:

Customer attribute information acquisition module, for obtaining at least two customer attribute informations corresponding to user ID to be detected;

Characteristic vector pickup module to be detected, for extracting proper vector to be detected according to the customer attribute information got;

Sort module, for described proper vector to be detected is inputted described User Status list disaggregated model, exports the classification results representing and whether belong to designated user state class, to determine the User Status corresponding to described user ID to be detected.

8. device according to claim 6, is characterized in that, described model parameter estimation module comprises:

Sample divides module, for all positive training samples are divided into the positive training sample of the first kind and the positive training sample of Equations of The Second Kind;

Candidate parameter value chooses module, for getting the candidate parameter value of predetermined number in the span of model parameter;

Statistic of classification module, for generating candidate user state list disaggregated model respectively according to the positive training sample of the first kind and each candidate parameter value, to carry out classification and statistical classification accuracy rate to the positive training sample of Equations of The Second Kind;

Model parameter determination module, for using candidate parameter value the highest for corresponding classification accuracy as the model parameter estimated.

9. according to the device in claim 6 to 8 described in any one, it is characterized in that, described probability density estimation be with natural constant be the truth of a matter, respectively with the exponential function that is index of the function between each sampling feature vectors and the proper vector of input and; The transposition that function between each sampling feature vectors and the proper vector of input is respectively the proper vector of input and the difference of corresponding sampling feature vectors is got negative, be multiplied by the negative second power of model parameter again, then be multiplied by the proper vector of input and the difference of corresponding sampling feature vectors.

10. device according to claim 9, is characterized in that, described device also comprises:

Mean Matrix computing module, for calculating the Mean Matrix of the sampling feature vectors of all positive training samples;

Variance matrix computing module, for calculating variance matrix according to described Mean Matrix;

Model parameter span computing module, for substituting into model parameter function to obtain the span of model parameter by described variance matrix; Described model parameter function is extract square root after the inverse of the dimension of sampling feature vectors is multiplied by the mark of variance matrix, then is multiplied by and adds up to the truth of a matter with positive training sample and get with the business of parametric variable and described dimension the exponential function born as index; Described parametric variable has default span.