Higher-dimension market segments method and device based on canonical correlationTechnical field
The present invention relates to a kind of market segments method, in particular to a kind of higher-dimension market segments method based on canonical correlationAnd device.
Background technique
With the continuous propulsion of Marketing Concept, in order to reduce the cost of enterprise, precision marketing is become more and more important.However, the differentiation of consumer demand, so that marketing becomes more and more difficult.The premise of realization precision marketing is will be to consumer groupBody is finely divided, that is, the market segments.The market segments (market segmentation) refer to that enterprise will according to certain indexCustomer in the market is divided into several customer bases, and it is the base of market targeting that each customer base, which constitutes a sub- market,Plinth.The effective market segments keep marketing investment more targeted, so that benefit is higher, carries out differentiated marketing, customize battalionPin, it is necessary to since the market segments.However it can be used for there are many indexs of the market segments: geographical location (Classification Index: areaDomain, development degree, city rank etc.), population characteristic's (Classification Index: age, gender, income, household size, education degree, residenceFirmly condition etc.), use and buying behavior (Classification Index: usage amount, frequency of use, purchasing channel, decision process, expense expenditureDeng), profit potential (Classification Index: income, procurement cost, cost of serving etc.), values and life style (Classification Index: valenceBe worth orientation, life attitudes, living habit etc.), demand and motivation (Classification Index: buying selection factor, unmet demand), attitude(Classification Index: to category attitude, to brand attitude) and use occasion (Classification Index: place, time, usage mode) etc.,If without reasonable subdivision scheme, it will be difficult to effectively be segmented to market.
Currently, method most popular for the market segments mainly has subdivision based on single index, based on multi objectiveSubdivision and subdivision based on linear index set of variables.Subdivision based on single index is intuitively divided into consumer severalGroup, such as be divided into according to consumption level medium, low and high；After being divided into 70 according to the age, it is after 80s, after 90s and 00 afterDeng；It is medium to be divided into the four corners of the world according to region.This subdivision based on single index is most simple and easy extensive subdivision, thisKind subdivision mode operability is relatively weak, poor with practical application stickiness.Subdivision based on multi objective is two or twoAbove index is cross-up, forms multidimensional and intersects, each crosspoint forms one small segment market, it may be necessary to according to eachThe otherness to segment market in demand side is combined to increase and can grasp with practical stickiness, this divided method to adjacent setsThe property made is stronger, but is easy to be influenced by subjective factor.Subdivision based on linear index set of variables be one dependent variable group of research withRelationship between one independent variable group, what it was analyzed is the whole correlation between set of variables, is suitable for single order dimension, and for mistakeComprehensive complicated, various dimensions variable market status subdivision adaptability is poor.
Summary of the invention
In view of the above deficiencies, the present invention provides a kind of applied widely, high reliablity higher-dimension city based on canonical correlationField fine method and device.
A kind of higher-dimension market segments method based on canonical correlation of the invention, described method includes following steps:
S1, according to sample, establish the data file of original variable；
S2, according to the original variable in S1, select multiple target variable groups, carry out non-linear canonical correlation analysis, determineThe canonical correlation coefficient of the dimension of sample and each dimension, when determining canonical correlation coefficient is greater than given threshold, then according to allusion quotationType related coefficient obtains the object score in each dimension of sample；
S3, the object score according to sample, classify to customer base with clustering；
S4, according to classification results, market segments result is described.
Preferably, in the S2, multiple target variable groups are selected in original variable, include in each target variable groupMultiple observational variables seek the dimension m under the optimal scale of the observational variable of selection1、m2..., make dimension miTo each observational variableThe average degree of correlation of group linear combination reaches highest, then according to every one-dimensional average loss, per one-dimensional characteristic value, degree of fittingDetermine the dimension m in output spaceiValue, i=1,2 ....
Preferably, in the S2, each target variable group indicates to influence a factor of the market segments, variable it is optimalScale type includes ordinal scale, single classification, polynary classification and discrete values.
Preferably, in the S2, adjustment target variable group number, the variable in each set of variables of increase and decrease and/or change are optimalScale type carries out non-linear canonical correlation analysis, determines the dimension of sample and the canonical correlation coefficient of each dimension, Zhi DaodianType related coefficient is greater than given threshold.
Preferably, whole related between the multiple target variable group.
Preferably, the type of the variable includes numeric type variable, ordinal variable, class variable and these typess of variablesAny combination.
Preferably, the S1 includes:
The all categories variable of non-linear canonical correlation analysis will be carried out as the original variable of data file, every an objectCorresponding classification number takes as variate-value, the classification number according to the maximum attribute of load in the direct selection factor on variableValue.
It preferably, include nominal level variable, ordinal data, spacing variable, the fixed feature than variable per one-dimensional characteristic valueValue and the value of statistical indicant of significance test.
The present invention also provides a kind of higher-dimension market segments device based on canonical correlation, described device include:
Original variable memory module to establish the data file of original variable according to sample, and stores；
Object score obtains module, to the original variable according to storage, selects multiple target variable groups, carries out non-linearCanonical correlation analysis determines the dimension of sample and the canonical correlation coefficient of each dimension, when determining canonical correlation coefficient is greater thanGiven threshold obtains the object score in each dimension of sample then according to canonical correlation coefficient；
Cluster module classifies to customer base with clustering to the object score according to sample；
As a result output module, to describe market segments result according to classification results.
Preferably, the object score obtains module, and multiple target variable groups, each index are selected in original variableInclude multiple observational variables in set of variables, seeks the dimension m under the optimal scale of the observational variable of selection1、m2..., make dimension miHighest is reached to the average degree of correlation of each observational variable group linear combination, then according to every one-dimensional average loss, per one-dimensionalCharacteristic value, degree of fitting determine the dimension m in output spaceiValue, i=1,2 ....
Preferably, each target variable group indicates to influence a factor of the market segments, the optimal scale type of variableIncluding ordinal scale, single classification, polynary classification and discrete values.
Preferably, the object score obtains module, to adjust target variable group number, increase and decrease the change in each set of variablesAmount and/or change optimal scale type, carry out non-linear canonical correlation analysis, determine the dimension of sample and the typical case of each dimensionRelated coefficient, until canonical correlation coefficient is greater than given threshold.
Preferably, whole related between the multiple target variable group.
Preferably, the type of the variable includes numeric type variable, ordinal variable, class variable and these typess of variablesAny combination.
Preferably, the original variable memory module, all categories of non-linear canonical correlation analysis will be carried outOriginal variable of the variable as data file, every an object classification number corresponding on variable is as variate-value, the classificationNumber according to directly selection the factor in the maximum attribute value of load.
It preferably, include nominal level variable, ordinal data, spacing variable, the fixed feature than variable per one-dimensional characteristic valueValue and the value of statistical indicant of significance test.
Above-mentioned technical characteristic may be combined in various suitable ways or be substituted by equivalent technical characteristic, as long as can reachTo the purpose of the present invention.
The beneficial effects of the present invention are the present invention is a kind of subdivision based on the mutual relationship of multiple groups target variable groupMethod is realized suitable for the market segments of various dimensions variable by class variable number by non-linear canonical correlation analysis meansQuantization and data dimensionality reduction, it is stronger with current intricate, various dimensions variable market status stickiness, it can consider consumerNeeds, also can be in view of the influence of the factors such as values, life style, while can also consider Demographic and consumptionPerson's behavior so that final subdivision group be comprehensive various factors, it is inherent, have logic.Applied widely, subdivision knotFruit high reliablity.
Detailed description of the invention
Fig. 1 is the standard drawing of the market segments；
Fig. 2 is the schematic diagram of the canonical correlation analysis of present embodiment, and wherein connecting key indicates each set of variables population characteristic valuve；
Fig. 3 is the statistical chart that how four class crowds reach hospital after clustering in embodiment；
Fig. 4 is the statistical chart that four class crowds select which type of hospital after clustering in embodiment；
Fig. 5 is the statistical chart that four class crowds are used for drug expenditure after clustering in embodiment.
Specific embodiment
Following will be combined with the drawings in the embodiments of the present invention, and technical solution in the embodiment of the present invention carries out clear, completeSite preparation description, it is clear that described embodiments are only a part of the embodiments of the present invention, instead of all the embodiments.It is based onEmbodiment in the present invention, those of ordinary skill in the art without creative labor it is obtained it is all itsHis embodiment, shall fall within the protection scope of the present invention.
It should be noted that in the absence of conflict, the feature in embodiment and embodiment in the present invention can phaseMutually combination.
The present invention will be further explained below with reference to the attached drawings and specific examples, but not as the limitation of the invention.
The standards of the market segments as shown in Figure 1, present embodiment a kind of higher-dimension market segments side based on canonical correlationMethod includes the following steps:
S1, according to sample, establish the data file of original variable；
Sample is the questionnaire treating the object of subdivision and doing, including attitude measurement, behavior measure, demand measurement, lifeForm acquisition etc., according to the content for including in questionnaire, selects original variable, and assign characteristic value；
S2, according to the original variable in S1, select multiple target variable groups, for example, attitude variables group as shown in Figure 2,Behavior set of variables, demand variable group, lifestyle set of variables and Demographic's set of variables carry out non-linear canonical correlation pointAnalysis, determine the dimension of sample and the canonical correlation coefficient of each dimension, when determining canonical correlation coefficient be greater than given threshold,That is: it achieves the desired results, then according to canonical correlation coefficient, obtains the object score in each dimension of sample；
Canonical correlation coefficient is equivalent to the regression coefficient in multiple linear regression；On the one hand, it shows that each observation becomesMeasure the importance to each canonical variable；On the other hand, it is constituted together with the optimal number of each level of each observational variableThe foundation of computing object score.
Characteristic value is converted into canonical correlation coefficient:
Wherein RiFor the canonical correlation coefficient of i-th dimension, λiFor the characteristic value in i-th dimension, K is set of variables number.
Object score is according to the optimal number of canonical correlation coefficient and each level of each observational variable, to every an objectThe average value of calculated typical case's score.
Target variable group in this step is very flexibly, can to need be defined according to oneself, mostly can also lack,Qualitative variable can receive with quantitative variable.
S3, the object score according to sample, classify to customer base with clustering；And it is verified and is classified with discriminant analysisAs a result；
S4, according to classification results, carry out profile analysis, describe market segments result.
Present embodiment is suitable for the market segments of various dimensions variable, realizes class variable quantification and data dimensionality reduction,Stronger with current intricate, various dimensions variable market status stickiness, subdivision result reliability is high.
In preferred embodiment, in the S2 of present embodiment, multiple target variable groups, each index are selected in original variableInclude multiple observational variables in set of variables, seeks the dimension m under the optimal scale of the observational variable of selection1、m2..., make dimension miHighest is reached to the average degree of correlation of each observational variable group linear combination, then according to every one-dimensional average loss, per one-dimensionalCharacteristic value, degree of fitting determine the dimension m in output spaceiValue, i=1,2 ....
For each collection and per one-dimensional, the multiple correlation coefficient R, 1-R of the collection canonical variable and object score are asked2Referred to as shouldCollect the loss in the dimension.It is the average value of loss of each collection in the dimension per one-dimensional average loss.And per on one-dimensionalCharacteristic value is equal to 1- average loss.
The sum of characteristic value in each dimension claims degree of fitting.The maximum value of degree of fitting is equal to dimension.Certain variable i-th dimension canonical correlationCoefficient square be known as the variable i-th dimension degree of fitting.The degree of fitting of the sum of the degree of fitting of variable in each dimension title variable.Degree of fitting and dimension and every one-dimensional average loss have following relationship: the average loss of all dimensions adds up to+degree of fitting=dimension；
In preferred embodiment, in S2, as shown in Fig. 2, each target variable group indicates to influence a factor of the market segments,The optimal scale type of variable includes ordinal scale, single classification, polynary classification and discrete values.Present embodiment both consideredTo the needs of consumer, it is also considered that the influence of the factors such as values, life style, allow also for Demographic andConsumer behaviour so that final subdivision group be comprehensive various factors, it is inherent, have logic, solve existing cityThe one-sidedness problem of field fine method.
Ordinal scale: retain the size order between each level of variable in quantization；Single classification: variable it is each it is horizontal onlyAs different classification logotypes, but optimal number of each value in all dimensions is identical；Polynary classification: each level of variableDifferent classification logotypes is served only as, while allowing each horizontal optimal number in each dimension different；Discrete values: variableAs spacing variable, retain the distance relation between each level.
Variable in preferred embodiment, in the S2 of present embodiment, in adjustment target variable group number, each set of variables of increase and decreaseAnd/or change optimal scale type, non-linear canonical correlation analysis is carried out, determines the dimension of sample and the typical phase of each dimensionRelationship number, until canonical correlation coefficient is greater than given threshold.
Present embodiment adjusts target variable group number, the variable in each set of variables of increase and decrease and/or changes optimal scale type,It is repeatedly attempted, exports the dimension of sample and the canonical correlation coefficient of each dimension, until canonical correlation coefficient is greater than settingThreshold value.
It is whole related between the multiple target variable groups of present embodiment, rather than about between variable in multiple set of variablesIt is related.As the connecting key in Fig. 2 indicates population characteristic valuve；
The type of variable is generalized to various types of variables from numeric type by present embodiment, and the type of variable includes numerical valueType variable, ordinal variable, class variable and these typess of variables any combination.
In present embodiment S1, all categories variable of non-linear canonical correlation analysis will be carried out as the original of data fileBeginning variable, every an object classification number corresponding on variable is as variate-value, and the classification number is according in the directly selection factorThe maximum attribute value of load.
It include nominal level variable, ordinal data, spacing variable, the fixed spy than variable per one-dimensional characteristic value in present embodimentValue indicative and the value of statistical indicant of significance test have been well solved using numerous different dimensions, different measurement level variablesThe problem of being finely divided；
Specific embodiment:
Most of pharmaceutical manufacturers use all in accordance with this classification standard corresponding marketing strategy and resource distribution, for example,Tertiary hospitals arrange special Pharmaceutical Sales Representative.Now with certain manufacturer of drugs by recognizing in previous repeatedly research: hospital is not onlyIt is an important channel of drug product, while is also the channel that one effective and consumer is linked up, it is therefore desirable to is rightDifferent hospitals, which uses, the strategy being directed to.So the medical treatment group characteristic of Different hospital and its for drug demand whether andIs existing hospital's classification corresponding? on the basis of summarizing previous experiences, manufacturer discovery is by the classification conduct of hospital's administrative gradeFor communicative channel, on corresponding with hospitalier not very precisely, although such as hospital's rank for having it is not high, due toA certain training has special reputation, thus hospitalier may be from different regions, although and some hospital's ranks compareHeight, but due to being located in residential block, hospitalier is but some residents on hospital periphery instead, is had instead more convenience-for-peopleDemand.Consumer is orientated the medical treatment of Different hospital so that the medical treatment crowd of hospital be provided with it is not consistent with hospital gradeFeature.For this purpose, the characteristics of client is determined to according to medical treatment crowd, the factors such as attitude re-start subdivision to hospital.Research usesQuantitative questionnaire, is investigated for the medical treatment crowd of certain class disease areas belonging to client, has collected 2400 samples.Following the description (part) is contained in questionnaire:
City where consumer: including 10 cities such as Beijing, Shanghai, Guangzhou.
The gender of medical staff: including male and female.
The age of medical treatment crowd: 18 to 60 years old.
The Income situation of medical treatment crowd: mainly using monthly family income as index.
The experience seen a doctor recently: oneself is seen a doctor, and accompanies other people.
Disease is on the severity influenced caused by life: according to the scale of 5 points of scales, carrying out self by medical staffEvaluation.
Medical staff is directed to this hospital: giving respectively from hospital grade, public private, training or comprehensive angleDefinition.
The mode of transportation of medical treatment: self-driving, taxi, public transport, walking or other.
The time of entire medical treatment process spends.
The expense and the means of payment of entire medical treatment process: it is divided into and pure curing at one's own expense, using social medical insurance, using businessTreat insurance.
The type of medical treatment consulting room: general out patient service, medical expert's consultation, acute disease, fever clinic, convenience-for-people/simple outpatient service etc..
The rank of selected doctor: intern, resident doctor, attending physician, deputy director's grade doctor, director grade doctorIt is raw.
Select hospital main standard and its on these attributes (table 1), to medical treatment hospital satisfaction, adoptWith 5 points of scale scales.
Table 1
| There are many different drugs is selective |
| Drug quality is secure |
| Drug specific aim is relatively high |
| The reasonable price of drug |
| There is the drug that I wants |
| Doctor's attitude is good |
| Physician specialty is horizontal |
| Nurse's attitude is good |
| Nurse's service level is good |
| The prescription information that doctor outputs is clearly understandable |
| This hospital is the authority of this disease |
| Hospital's hardware facility is good |
| Medical environment is good |
| Medical link instruction is clear |
| Have a good transport service/close from my inhabitation/work place |
| Hospital's popularity is high |
| I is familiar with this hospital of family |
| There is my known doctor |
| The rank of hospital is high |
| There is the medical expert's consultation of this kind of disease |
| Queuing time is few |
| Weekend can go to a doctor |
| There is convenience-for-people outpatient service/simple outpatient service |
| All kinds of medical insurances can be used |
| The reasonable price of medical/inspection |
| It will not arrange unnecessary inspection |
| Respect the requirement of patient |
It is the selection of variable first, the present embodiment has selected three target variable groups:
Group one: population and disease feature
Age (age-based size is divided into 4 sections, and optimal scale type is sequence)
Income (is divided into 5 sections by income height, optimal scale type is sequence)
Influence degree (be divided into 5 section, optimal scale type be sequence) of the disease to life
Group two: medial demand
The rank of selected hospital (optimal scale type is single classification)
(optimal scale type is discrete values, since part hospitalier may not have in drug for cost on drugExpense expenditure, therefore can be each value plus 1)
Group two: Hospital choice standard
Nurse's attitude is good (5 points of scales, optimal scale type are sequence)
The rank of hospital is high (5 points of scales, optimal scale type are sequence)
It is medical, the reasonable price of inspection (5 points of scales, optimal scale type are sequence)
It is selective (5 points of scales, optimal scale type are sequence) that there are many different drugs
Medical environment is good (5 points of scales, optimal scale type are sequence)
Queuing time is few (5 points of scales, optimal scale type are sequence)
There is my known doctor (5 points of scales, optimal scale type are sequence)
It has a good transport service, lives from me, the place of work closely (5 points of scales, optimal scale type are sequence)
All kinds of medical insurances can be used
What needs to be explained here is that usually having 2 kinds of ways in practice, one is straight for the variable-value of such as third groupSelecting takes the factor, converts data according to factor score and (1-5 points is typically translated into, because atypia correlation analysis requiresIt is assigned a value of the integer of 1 beginning)；Another way is directly selected in the factor, and the maximum attribute of factor loading takes its originalValue (divides scaled score in this research for 1-5).The present embodiment is using second of value mode, in general, the latter is moreIt is easy to get preferable effect.For the present embodiment, due to no polynary class variable, 14 dimensions can be at most defined(3+2+9) is being adjusted the variable increased and decreased in each set of variables and/or is changing optimal scale type, carrying out non-linear allusion quotationType correlation analysis, it is final to determine that 9 dimensions carry out dimension for further analysis after attempting several times.
Table 2 is the degree of fitting of the present embodiment, and the sum of the characteristic value in each dimension claims degree of fitting.The maximum value of degree of fitting is equal to dimensionNumber.Degree of fitting and dimension and every one-dimensional average loss have following relationship: the average loss of all dimensions adds up to+degree of fitting=dimensionNumber.It is noted above, when canonical correlation analysis is used for the market segments, it is important that the meaning of each dimension, withoutIt is it explains original variable how much information.Thus, in the sense that emphasis is every dimension；
Table 2
Analysis and summary
Table 3 is multiple fitting Multiple Fit, for checking distinguishing ability of each variable on each different dimensions.ExampleSuch as from the data of table 3 as it can be seen that there is better distinctiveness at the age than income, and then opposite for dimension 2 for dimension 1.Table 4 isSingle fitting Single Fit, for assessing contribution of each variable for the dimension, such as from the data of table 4 as it can be seen that the ageIt is contributed largely in the first dimension, and bigger variable is contributed to the first dimension there are also hospital category, the factors such as hospital grade.
Table 3
Table 4
According to both the above table, it is found that preceding 9 dimensions suffer from oneself specific meaning, be difficult to take completely between each otherIn generation, carries out clustering based on this 9 dimensions thus, extracts four class crowds.The characteristics of in order to describe every class crowd.It will clusterObtained crowd and the common factor proposed in the factorial analysis of front carry out crosstab, obtain that the results are shown in Table 5.
Table 5
These four types of crowds are not difficult to find out from table 5, the obvious demand of the first kind is the convenience of hospital, including " whenBetween it is convenient " and " having a good transport service ", this kind of crowds can be referred to as " convenience guiding (convenience type) "；Second class is obviousDemand be " be suitable for medical insurance ", and " service level of hospital ", this kind of crowds can be referred to as to " medical insurance is oriented to(medical insurance type) "；The obvious demand of third analogy is " hospital professional ", this kind of crowds can be referred to as " professionProperty guiding (professional) "；This kind of crowds can be referred to as " price guiding (material benefit by the 4th class crowd then obvious priceType) ".Difference of these four types of crowds in population feature and selected medial demand, therefore base are not difficult to find out by table 6 and Fig. 3-5Originally it may determine that this classification is meaningful.
Table 6
After having basic judgement to the market segments, the content that can be subsequently measured further combined with other in questionnaire,Each subdivision crowd is described in population feature, medial demand, show the attitude of drug, client's brand, client can makeFixed corresponding marketing strategy.
Present embodiment also provides a kind of higher-dimension market segments device based on canonical correlation corresponded to the above method,Described device includes:
Original variable memory module to establish the data file of original variable according to sample, and stores；
Object score obtains module, to the original variable according to storage, selects multiple target variable groups, carries out non-linearCanonical correlation analysis determines the dimension of sample and the canonical correlation coefficient of each dimension, when determining canonical correlation coefficient is greater thanGiven threshold obtains the object score in each dimension of sample then according to canonical correlation coefficient；
Cluster module classifies to customer base with clustering to the object score according to sample；
As a result output module, to describe market segments result according to classification results.
In preferred embodiment, the object score obtains module, multiple target variable groups is selected in original variable, eachInclude multiple observational variables in target variable group, seeks the dimension m under the optimal scale of the observational variable of selection1、m2..., make to tie upNumber miHighest is reached to the average degree of correlation of each observational variable group linear combination, then according to per one-dimensional average loss, eachThe characteristic value of dimension, degree of fitting determine the dimension m in output spaceiValue, i=1,2 ....
In preferred embodiment, each target variable group indicates to influence a factor of the market segments, the optimal scale of variableType includes ordinal scale, single classification, polynary classification and discrete values.
In preferred embodiment, the object score obtains module, to adjust target variable group number, increase and decrease in each set of variablesVariable and/or change optimal scale type, carry out non-linear canonical correlation analysis, determine sample dimension and each dimensionCanonical correlation coefficient, until canonical correlation coefficient is greater than given threshold.
It is whole related between the multiple target variable group in preferred embodiment.
In preferred embodiment, the type of the variable includes numeric type variable, ordinal variable, class variable and these variablesAny combination of type.
In preferred embodiment, the original variable memory module, all of non-linear canonical correlation analysis will be carried outOriginal variable of the class variable as data file, every an object classification number corresponding on variable are described as variate-valueClassification number is according to the maximum attribute value of load in the directly selection factor.
In preferred embodiment, includes nominal level variable per one-dimensional characteristic value, ordinal data, spacing variable, determines than variableCharacteristic value and the value of statistical indicant of significance test.
Although describing the present invention herein with reference to specific embodiment, it should be understood that, these realitiesApply the example that example is only principles and applications.It should therefore be understood that can be carried out to exemplary embodimentMany modifications, and can be designed that other arrangements, without departing from spirit of the invention as defined in the appended claimsAnd range.It should be understood that different appurtenances can be combined by being different from mode described in original claimBenefit requires and feature described herein.It will also be appreciated that the feature in conjunction with described in separate embodiments can be usedIn other described embodiments.