CN102591872A - Client feature library generating method and device - Google Patents

Abstract

The invention provides a client feature library generating method, which comprises the steps of: calculating the sum d of intra-class distances under different cluster numbers k, and drawing a k-d value curve; finding a turning point according to the k-d value curve, finding a k value corresponding to the turning point, looking for an optimal cluster number kopt within a range near the k value, and utilizing a cluster result corresponding to the optimal cluster number as a final cluster result. The invention also provides a client feature library generating device. The scheme of the invention can improve the cluster searching efficiency and precision, and the method of the invention is used for performing client analysis according to the obtained client feature library to improve the recommendation success rate and client satisfaction degree of a marketing campaign.

Description

A kind of client characteristics library generating method and device

Technical field

The present invention relates to technical field of information processing, particularly a kind of client characteristics library generating method and device.

Background technology

As its name suggests, classify customers exactly, and the client who is divided carried out the database of classification and storage in the client characteristics storehouse according to client characteristics information.The client characteristics storehouse plays crucial effects in the marketing implementation:

For managerial personnel,, can realize the client traffic coupling through client characteristics coupling and its business that is fit to, channel etc.; Accurate localizing objects customer group to the marketing rule of differentiation client poor designs alienation, experiences the user to become more meticulous fast and accurately the marketing design, realizes that marketing management is intelligent;

For line marketing personnel, can help it to find the business that the client is fit to fast accurately, clear recommendation list, the unified bore of recommending;

For the customer, the rate coupling through client characteristics storehouse accurately provides can find the rate that conform to own consumption, and the perception of lifting client rate lets the client feel more economical; Professional coupling can help the client to find the business that conforms to its service feature, promotes the client traffic perception, lets the client feel more satisfied; Through implementing client's channel hobby coupling, can promote the perception of client's channel, let the client feel more convenient; Through the sales promotion coupling, implement the management of client's contact point, promote client's contact point sales promotion perception, let the client feel more intimate.

In the prior art, often adopt the K_MEANS clustering algorithm to realize the division of customers.The cost function that the K_MEANS clustering algorithm adopts is:

$J=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\mathrm{Ji}=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\left(\underset{x_j\∈G_k}{\mathrm{\Σ}}{\left|\right|x_j-c_i\left|\right|}^2\right)---\left(1\right)$

Wherein k is the initial clustering number, c _iBe cluster centre.Fig. 1 provides in the prior art, and the mobile subscriber is carried out the flow process that customers divide, and comprises the steps:

Step 101: from containing n user's customer group x _i(i ∈ (and 1,2 ...., n)) in select k user at random, with a selected k user as initialized user clustering center c _j(j ∈ (1,2 ...., k)).

Step 102: calculate each sample data object and user clustering center c _jDistance

Wherein,

The characteristic information of expression user's a certain quantification, for example age, income situation etc.If d (x _i, c _k)=min{d (x _i, c _j), j=1,2....k}, wherein j ∈ (1,2 ...., k), i ∈ (1,2 ...., n), then with user x _iBe divided into customer group G _k

Step 103: the value of given price value function J.

Step 104: if cost function J convergence, then algorithm finishes, and carries out otherwise return step 101.

The cost function practical meaning is exactly to pass judgment on the rationality that this central point is divided, and is more reasonable if the client of center of a sample's point chooses, promptly the customer representative of this center of a sample's point other client's general character, then cost function is a convergent.Yet the algorithm of this model almost can carry out cluster to each user, and the complexity of algorithm is very high, means that also system resources consumption is very high.

Because it is a kind of basic and use the widest division methods in the clustering method that the K-MEANS clustering algorithm belongs to; Be a kind of method of in not having type label data, finding type with type center, can't satisfy in marketing activity requirement the characteristic user becomes more meticulous, differentiation is analyzed.If directly traditional K-MEANS algorithm is applied to customer segmentation, consider userbase, the cost of its resource consumption will be quite big.

Summary of the invention

The invention provides a kind of client characteristics library generating method and device, can reduce the resource consumption in the cluster calculating process.

A kind of client characteristics library generating method that the embodiment of the invention proposes comprises the steps:

A, input comprise m the wide table of basic index attribute of n mobile client, and each mobile subscriber's m basic index constitutes a m dimensional vector;

B, for all possible cluster numbers k; Vector in the wide table of said basic index attribute carries out the cluster computing, and corresponding type of record is interior apart from the value of sum

and the output result after the cluster;

a _i∈ A _i, A _iBe any cluster under the cluster numbers k, a _iBe cluster A _iCentral point, x is cluster A _iThe sample that is comprised;

C, draw the k-d change curve, find the turning point of k-d change curve and be in the k value of this point, note is k ₀

D, to k ∈ [k ₀-θ, k ₀+ θ] all interior k values of scope, carry out the computing of K-MEANS clustering algorithm,

And calculate under the corresponding k value cluster effect function

wherein θ be integer constant;

Obtain among E, the comparison step D

Value, get K value hour is best cluster numbers k _Opt

F, with k=k among the step B _OptThe time cluster result output in the client characteristics storehouse as net result.

Preferably; The said all possible cluster numbers of step B is all the k values from k=1 to

, and symbol

expression rounds up.

Preferably, before the said steps A, further comprise:

Select out n targeted customer according to marketing activity;

According to the purpose that corresponding marketing activity is analyzed, extracting objects user's a m basic index.

The embodiment of the invention also proposes a kind of client characteristics storehouse generating apparatus, comprising:

Load module is used to receive m the wide table of basic index attribute that comprises n mobile client, and each mobile subscriber's m basic index constitutes a m dimensional vector;

The first cluster computing module is used for for all possible cluster numbers k, and the vector in the wide table of said basic index attribute is carried out the cluster computing, and corresponding type of record is interior apart from sum

Value and the output result after the cluster; a _i∈ A _i, A _iBe any cluster under the cluster numbers k, a _iBe cluster A _iCentral point, x is cluster A _iThe sample that is comprised;

The k-d relationship module is used for the cluster computing structure that obtains according to said cluster computing module, draws the k-d change curve, finds the turning point of k-d change curve and is in the k value of this point, and note is k ₀

The second cluster computing module is used for the [k to k ∈ ₀-θ, k ₀+ θ] all interior k values of scope, carry out the computing of K-MEANS clustering algorithm, and calculate the cluster effect function under the corresponding k value

Wherein θ is an integer constant;

Comparison module, it is resulting to be used for the comparison second cluster computing module

Value, get K value hour is best cluster numbers k _Opt

Output module is with the resulting k=k of the first cluster computing module _OptThe time cluster result output in the client characteristics storehouse as net result.

Preferably; Said all possible cluster numbers is all the k values from k=1 to , and symbol

expression rounds up.

Can find out from above technical scheme, interior through the class of calculating under the different cluster numbers k apart from sum d, depict k-d value curve; Find its turning point according to k-d value curve, find the corresponding k value of this turning point then, seek optimum cluster numbers k near the scope this k value _Opt, the cluster result that optimum cluster numbers is corresponding is as final cluster result.Through cluster effect function

Can find optimum cluster numbers

Pass through k _OptThe cluster of carrying out can improve the precision of K-MEANS algorithm, and using can be effectively k apart from the change curve of sum d in the class _OptDelineation is k ∈ (k in a small range _Opt-θ, k _Opt+ θ), with respect to traditional

, improved the efficient of algorithm, make time complexity by O (n ²Logn) be kept to O (2n θ logn), wherein θ is a constant, and 2 θ＜＜n.Therefore; The present invention program can reduce the resource that the cluster calculation process is consumed with respect to traditional scheme; And cluster result is more accurate, carries out customer analysis according to the client characteristics storehouse that the inventive method obtains, and can improve the recommendation success ratio and the CSAT of marketing activity.

Description of drawings

Fig. 1 is in the prior art, and the mobile subscriber is carried out the schematic flow sheet that customers divide;

Fig. 2 is the calculating synoptic diagram of mean distance between the class under the cluster numbers k=2 situation;

Fig. 3 is the calculating synoptic diagram of mean distance between the class under the cluster numbers k=3 situation;

Fig. 4 for type in apart from sum d with k value change curve;

Fig. 5 is the realization flow figure of the client characteristics library generating method of embodiment of the invention proposition;

The customer group classification synoptic diagram that Fig. 6 obtains for the client characteristics library generating method that adopts the present invention to propose.

Embodiment

Usually, a good clustering method should make between each cluster and comparatively disperse, and each cluster inside is comparatively compact.Just like jljl with cluster, the people is with gregarious, the otherness in same customer group inside is smaller, and that the otherness between the different customer group is the rate of exchange is big.For satisfying the demand, the present invention proposes a kind of new client characteristics library generating method.

At first the principle to the present invention program describes.

If cluster data collection G contains the vectorial sample { x of n m dimension ₁, x ₂..x _n, being designated as G{x, n} supposes G{x, n} is k type of { A by cluster ₁, A ₂... A _k, 1＜k＜n; For any two m dimensional vector (x _iAnd x _j) between distance, its Euclidean distance is:

$d(x_i,x_j)=\sqrt{[(x_{i1}-x_{j1})2+(x_{i2}-x_{j2})2+...+(x_{\mathrm{im}}-x_{\mathrm{jm}})2]}$

Definition one: if given cluster numbers k; Then at cluster set G{x, (can prove, be under the situation of k in the hypothesis cluster numbers to choose k maximum apart from sum between any two sample point of vector among the n}; K the point that vector is maximum apart from sum between any two must belong to k cluster in this case.Here do not prove, directly use this conclusion), and with this k the point carry out cluster, finally form k cluster; Then type mean distance be defined as this k cluster centre put all cluster center of a sample's points apart from sum divided by cluster numbers k.

$\stackrel{\‾}{D}=\frac{1}{k}\underset{i=1}{\overset{k}{\mathrm{\Σ}}}|a_i-p|$ a _i∈A _i

Wherein

Be mean distance between class, k is a cluster numbers, a _iBe class A _iCentral point, p is the central point of all samples.Fig. 2 and Fig. 3 are respectively the calculating synoptic diagram of mean distance between the class under cluster numbers k=2 and the k=3 situation.Wherein, each circle or cluster of ellipse representation; Each Diamond spot is represented the vector that cluster data is concentrated.X1, x2 and x3 represent respectively in each cluster in twos apart from maximum vector.

As shown in Figure 2, p is the central point of all samples in the sample space, maximum apart from the sum between any two some x of sample among Fig. 2 when k=2 ₁, x ₂, selected these 2 are carried out the cluster computing as the cluster initial point, and then sample space is divided into 2 class A ₁, A ₂Central point is respectively a in its type ₁, a ₂Then among Fig. 2

$\underset{k=2}{\overset{\‾}{D}}=\frac{1}{2}\left(\right|a_1-p|+|a_2-p\left|\right)$

As shown in Figure 3, during k=3, then according to maximum apart from sum between any two principle, existing x ₁, x ₂Situation under, then the 3rd point should be x ₃, but not other the point.Among the figure two

Definition two: apart from sum, be defined as corresponding with it the respectively distances of clustering centers sum of the contained sample of this k cluster in type.

$d=\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{x\∈\mathrm{Ai}}{\mathrm{\Σ}}|x-a_i|$ a _i∈A _i

Wherein d type of being is interior apart from sum, and k is a cluster numbers, a _iBe class A _iCentral point, x type of being A _iThe sample that is comprised.

Through defining one, can type of using between mean distance

To sample space G{x, discrete case all kinds of among the n} quantize, in theory,

When value is big, then all kinds of { A ₁, A ₂... A _kSpacing bigger, the cluster effect is also better,

Value is described the dispersion degree between each cluster; Through defining two, we with type in apart from sum d to all kinds of { A ₁, A ₂... A _kInner compact degree quantizes, and d value is more little, and the distance between the then all kinds of interior samples is also less, and then each cluster also seems more compact, and the d value is described the compact degree of each sample of all kinds of inside.

Close definition two based on above-mentioned definition one, proposed a kind of new cluster validity effect function:

Definition three: cluster effect function

$F(\stackrel{\‾}{D},d)=\frac{\stackrel{\‾}{D}}{d}=\frac{1}{k}\frac{\underset{i=1}{\overset{k}{\mathrm{\Σ}}}|a_i-p|}{\underset{i=1}{\overset{k}{\mathrm{\Σ}}}\underset{x\∈A_i}{\mathrm{\Σ}}|x-a_i|}$ a _i ∈A _i 1＜k＜n

Since at value big with the d value than hour; Cluster can obtain effect preferably, thereby we get

According to the digital simulation analysis, the inventor finds: when k＜＜k _OptThe cluster of Shi Jinhang, the inner similarity of each cluster is lower, thereby bigger apart from sum d in the class; Along with the increase of k value, demonstrate downward trend apart from sum d in type, this is because because the cluster division is more and more thinner, make that all kinds of inner similarities are higher, thereby reduces apart from sum d in the class; But, when k reaches k _OptAfter; Obviously slow down apart from sum d downtrending in type; This is because because k has surpassed desirable cluster numbers; Then newly-increased type must be the division of from desirable cluster, carrying out once more, and distance does not have greatly changed because of the increase of cluster numbers in its type, thus type in just do not have too big variation apart from sum yet.As shown in Figure 4 with k value change curve in type apart from sum d.Can find through observing Fig. 4, in the turning point of curve, be exactly k in fact _OptMost probable value.

Based on above analysis, the scope that the present invention proposes a kind of new cluster numbers k is confirmed method, through calculating the d value under the different value of K, depicts k-d value curve; Find its turning point according to k-d value curve, find the corresponding k value of this turning point then.Regulation k ± θ is k value scope, and wherein θ is an integer constant, and the value of θ should be done suitable adjustment according to the size of sample space, if the bigger then θ of sample space should get bigger constant, vice versa.

The realization flow of the client characteristics library generating method that the embodiment of the invention proposes is as shown in Figure 5, comprises the steps:

Step 501: select out targeted customer's scope according to marketing activity.Suppose selected n targeted customer.

Step 502: according to the purpose that corresponding marketing activity is analyzed, extracting objects user's index dimension; Suppose to have extracted m basic index.

Step 503: input comprises m the wide table of basic index attribute of n mobile client.Each mobile subscriber's m basic index constitutes a m dimensional vector.

Step 504: from all possible k value; Vector in the wide table of said basic index attribute carries out the cluster computing; Select cluster centre at random; And corresponding type of record is interior apart from the value of sum d and the output result after the cluster, and the output result is included in which user is each cluster comprised under this K value, i.e. the user of each cluster set.In the embodiment of the invention; Possible k value scope from k=1 wherein to

, symbol

expression rounds up.

Step 505: draw the k-d change curve, find the turning point of k-d change curve and be in the k value of this point, note is k ₀

Step 506: to k ∈ [k ₀-θ, k ₀+ θ] (θ is a constant, confirms according to the sample space size) all interior k values of scope, carry out the computing of K-MEANS clustering algorithm, and calculate under the corresponding k value

Cluster effect function.

Step 507: in the comparison step 506

Value, get K value hour is k _OptBe best cluster numbers, on behalf of n user of input, best cluster numbers promptly be divided into k _OptIndividual type can obtain optimum polymerization effect down.

Step 508: with k=k in the step 504 _OptThe time cluster result output in the client characteristics storehouse as net result, the output result is included in which user is each cluster comprised under this K value, promptly the user of each cluster gathers.

So far, the client characteristics storehouse based on the realization of modified K-MEANS algorithm of the present invention's proposition precisely generates the flow process end of method and system.

The embodiment of the invention also proposes a kind of client characteristics storehouse generating apparatus, comprising:

Load module is used to receive m the wide table of basic index attribute that comprises n mobile client, and each mobile subscriber's m basic index constitutes a m dimensional vector;

The first cluster computing module is used for for all possible cluster numbers k, and the vector in the wide table of said basic index attribute is carried out the cluster computing, and corresponding type of record is interior apart from sum

Value and the output result after the cluster; a _i∈ A _i, A _iBe any cluster under the cluster numbers k, a _iBe cluster A _iCentral point, x is cluster A _iThe sample that is comprised;

The k-d relationship module is used for the cluster computing structure that obtains according to said cluster computing module, draws the k-d change curve, finds the turning point of k-d change curve and is in the k value of this point, and note is k ₀

The second cluster computing module is used for the [k to k ∈ ₀-θ, k ₀+ θ] all interior k values of scope, carry out the computing of K-MEANS clustering algorithm, and calculate the cluster effect function under the corresponding k value

Wherein θ is an integer constant;

Comparison module, it is resulting to be used for the comparison second cluster computing module Value, get

K value hour is best cluster numbers k _Opt

Output module is with the resulting k=k of the first cluster computing module _OptThe time cluster result output in the client characteristics storehouse as net result.

Preferably; Said all possible cluster numbers is all the k values from k=1 to

, and symbol

expression rounds up.

The beneficial effect of the client characteristics library generating method that the present invention proposes is following: through cluster effect function

Can find optimum cluster numbers

Pass through k _OptThe cluster of carrying out can improve the precision of K-MEANS algorithm, and in addition through definition two, using can be effectively k apart from the change curve of sum d in the class _OptDelineation is k ∈ (k in a small range _Opt-θ, k _Opt+ θ), with respect to traditional

, improved the efficient of algorithm, make time complexity by O (n ²Logn) be kept to O (2n θ logn), wherein θ is a constant, and 2 θ＜＜n.

The present invention can be widely used in the large-scale consumer segmentation model of moving communicating field; And be applied in the flow process of marketing service; For instance, can extract a part of snapshot user to the 3G subscription crowd, extract a part of common 2G target customer in addition again; 2 certain customers are mixed; The improved k-means algorithm that at last utilizes this paper to propose according to different dimensions carries out cluster to the user to be divided, and we just can therefrom find out the customer group that similar cluster effect is arranged with 3G subscription, so in fact be appreciated that for these certain customers also be potential 3G subscription most probably; When marketing, we market such client and recommend to improve the recommendation success ratio preferably as the potential user with 3G characteristic.In addition, the present invention has improved the K-MEANS algorithm accuracy, makes its data for large data sets can obtain result preferably, and a kind of searching k is provided _OptThe feasibility method.

2010, A economized moving projection the 3G potential user of the whole province is carried out G3 scheming marketing activity, before accomodation of activities; In order to verify the effect of client characteristics storehouse to marketing, we to A province commmunication company randomly drawed 10000 number scales at net 2G subscriber's account, and to this batch user data through index screening, index scoring; 6 expense indexs have been selected; The customer group classification that the client characteristics library generating method that adopts the present invention to propose obtains, the final user has been divided into 7 types, and the result is as shown in Figure 6.Can find out that through Fig. 6 the user has been clustered into 7 types according to the property value of himself, we find that the GPRS expense has accounted for very big ratio in such in class 1; About 90%, further such user's detailed single, bill are analyzed the back and find that such user's major part also belongs to the middle and high end client; The monthly expense that is used to surf the Net, flow are all higher, and rule of thumb, such user possibly like online at ordinary times; And the consumption on surfing Internet with cell phone is also higher; If recommend to move the G3 scheming to such user, can be higher by the possibility that the user accepts, therefore; Such user is defined as the customers with the characteristic of surfing the web, and also is the potential target client of this G3 scheming marketing activity.Further, we carry out finding behind outgoing call, the short-message verification to this batch target customer through the marketing of contact by all kinds of means that the marketing service platform provides; Really, the ratio that such user accepts the G3 scheming is all high than other several types of clients, thus; We can further carry out large-scale G3 scheming marketing activity and promote; And carry out the client characteristics storehouse by means of the K-MEANS algorithm after the improvement of the present invention's proposition and analyze, improve the recommendation success ratio of marketing activity, increase customer satisfaction degree.

The above is merely preferred embodiment of the present invention, and is in order to restriction the present invention, not all within spirit of the present invention and principle, any modification of being made, is equal to replacement, improvement etc., all should be included within the scope that the present invention protects.

Claims (5)

1. a client characteristics library generating method is characterized in that, comprises the steps:

A, input comprise m the wide table of basic index attribute of n mobile client, and each mobile subscriber's m basic index constitutes a m dimensional vector;

B, for all possible cluster numbers k, the vector in the wide table of said basic index attribute is carried out the cluster computing, and in corresponding type of the record apart from sum Value and the output result after the cluster; a _i∈ A _i, A _iBe any cluster under the cluster numbers k, a _iBe cluster A _iCentral point, x is cluster A _iThe sample that is comprised;

C, draw the k-d change curve, find the turning point of k-d change curve and be in the k value of this point, note is k ₀

D, to k ∈ [k ₀-θ, k ₀+ θ] all interior k values of scope, carry out the computing of K-MEANS clustering algorithm, and calculate the cluster effect function under the corresponding k value

Wherein θ is an integer constant;

Obtain among E, the comparison step D

Value, get

K value hour is best cluster numbers k _Opt

F, with k=k among the step B _OptThe time cluster result output in the client characteristics storehouse as net result.

2. method according to claim 1; It is characterized in that; The said all possible cluster numbers of step B is all the k values from k=1 to

, and symbol

expression rounds up.

3. method according to claim 1 is characterized in that, before the said steps A, further comprises:

Select out n targeted customer according to marketing activity;

According to the purpose that corresponding marketing activity is analyzed, extracting objects user's a m basic index.

4. a client characteristics storehouse generating apparatus is characterized in that, comprising:

Load module is used to receive m the wide table of basic index attribute that comprises n mobile client, and each mobile subscriber's m basic index constitutes a m dimensional vector;

The first cluster computing module is used for for all possible cluster numbers k, and the vector in the wide table of said basic index attribute is carried out the cluster computing, and corresponding type of record is interior apart from sum

Value and the output result after the cluster; a _i∈ A _i, A _iBe any cluster under the cluster numbers k, a _iBe cluster A _iCentral point, x is cluster A _iThe sample that is comprised;

The k-d relationship module is used for the cluster computing structure that obtains according to said cluster computing module, draws the k-d change curve, finds the turning point of k-d change curve and is in the k value of this point, and note is k ₀

The second cluster computing module is used for the [k to k ∈ ₀-θ, k ₀+ θ] all interior k values of scope, carry out the computing of K-MEANS clustering algorithm, and calculate the cluster effect function under the corresponding k value

Wherein θ is an integer constant;

Comparison module, it is resulting to be used for the comparison second cluster computing module

Value, get

K value hour is best cluster numbers k _Opt

Output module is with the resulting k=k of the first cluster computing module _OptThe time cluster result output in the client characteristics storehouse as net result.

5. device according to claim 4; It is characterized in that; Said all possible cluster numbers is all the k values from k=1 to

, and symbol

expression rounds up.

Images