CN105491157A - User collaborative regularization-based personalized Web service composition method - Google Patents

Abstract

The invention discloses a user collaborative regularization-based personalized Web service composition method. According to the user collaborative regularization-based personalized Web service composition method of the invention, based on a matrix decomposition algorithm for fusing historical data of users, the QoS information of services can be efficiently predicted, and requirements of the users for personalized service combination can be satisfied; based on a matrix decomposition acceleration technology, solving efficiency can be greatly enhanced, and therefore, the personalized QoS combination query requests of a plurality of users can be responded in real time.

Description

A kind of personalized web service composition method based on user collaborative regularization

Technical field

The invention belongs to Computer Service technical field, be specifically related to a kind of personalized web service composition method based on user collaborative regularization.

Background technology

In recent years, Web service technology is flourish, and as a kind of Internet resources that at any time can be accessed by the user, Web service to be deployed in widely in Internet enterprises and to meet consumers' demand.Along with the Internet is towards the continuous evolution of target of cloud computing, the development of Web service presents following 4 trend: (1) quantity of service increases fast, the Internet exists ten hundreds of all kinds of services; (2) service describing semantization, semantic information is that the functional description of service provides unified standard; (3) QoS (QualityofService) served receives publicity, and QoS becomes the important indicator that user considers when using service; (4) under cloud computing environment, service execution scenarios is more complicated and changeable, and available service originally may become and cannot use under some specific environment, thus will seem more reliable with the composite services carried into execution a plan more.Under this background, from magnanimity service, how obtain the research emphasis that the composite services can meeting consumers' demand and have optimum QoS as much as possible will be Web service under cloud computing trend rapidly.

At present, the research of user's services selection all comprises a public prerequisite: namely first user must know the QoS information of all services, and then generates strategy and select.But under real conditions, this data premise calls is difficult to be met, reason is as follows: (1) most Service Source is all provided by commercial company, the QoS information obtaining service can only be undertaken by user's method of calling, but this money that user will be spent a large amount of and time cost.(2) contemporary internet topological structure is complicated, and user contains a large amount of noises by the QoS information that method of calling obtains, so that researcher needs to spend a large amount of energy cleaning data.Therefore, under real application scenarios, a large amount of sufficient QoS resource cannot be obtained by the mode of the service of calling.The disappearance of QoS information, directly causes enterprise that Services Subset can only be provided to allow user select, cannot meet consumers' demand well.

Improving the high efficiency method of user's services selection satisfaction, is use Mathematical Modeling to solve QoS scarcity problem.In recent years, how researcher is predicted by the QoS data resource of use machine learning method to disappearance in thinking.In current major part work, the QoS information of the unknown is predicted that main use calculates the similarity between client user or between Web service based on PearsonCorrelationCoefficient (PCC) method.But there are following 2 deficiencies in this computational methods: (1) PCC method needs to take statistics study to the QoS in historical record, depends critically upon accuracy and the completeness of data.But due to the complexity of contemporary internet environment, QoS record might not be all accurately, PCC method Similarity measures accuracy rate under service compute scene is caused to decline.(2) traditional PCC method is widely used in commending system field.But the application scenarios of commending system and service compute also exists the difference of essence, in QoS historical record, each is all determined by the actual use network environment of user, and the feature of this data objectivity directly reduces the precision of PCC Similarity Measure.

In addition, there is the problems such as user satisfaction is low in the Web service Automatic Combined technology of current main flow, and tracing it to its cause is that causing cannot the list of efficient combination candidate service because be subject to noise data interference when the QoS prediction module of core is run.How predicting all QoS information of Web service from extremely sparse historical data learning to rule, is the crucial problem in current service Automatic Combined.

Summary of the invention

For the above-mentioned technical problem existing for prior art, the invention provides a kind of personalized web service composition method based on user collaborative regularization, it can predict the QoS information of service efficiently by the matrix decomposition algorithm merging user's historical data, and supports the personalized service requirements of combination meeting user.

Based on a personalized web service composition method for user collaborative regularization, comprise the steps:

(1) collect the service call data that all users provide, described service call data comprise the QoS data of user to all Web services that it called; And then set up the QoS variable matrix S between user and Web service according to described QoS data;

(2) the neighborhood user set of each user is determined according to described QoS variable matrix S;

(3) gather according to the neighborhood user of described QoS variable matrix S and each user, by the following target function J of SVD method establishment, and this target function J is minimized solve, in the hope of the implicit features matrix U about user and the implicit features matrix V about Web service; And then according to R=U ^tv rebuilds the QoS prediction matrix R between user and Web service;

$J=\frac{1}{2}\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}{I}_{ij}{(S_{ij}-U_i^T{V}_j)}^2+\frac{\mathrm{\α}}{2}\left|\right|U|{|}_2^2+\frac{\mathrm{\α}}{2}|\left|V\right|{|}_2^2+\frac{\mathrm{\β}}{2}\left|\right|U_i-\frac{1}{N}\underset{k\∈K\left(i\right)}{\Σ}{U}_k|{|}_2^2$

Wherein: U _ifor the i-th column vector in implicit features matrix U, V _jfor the jth column vector in implicit features matrix V, K (i) is the neighborhood user set of i-th user, and N is the total number of users that default neighborhood user number and neighborhood user gather in K (i), U _kfor neighborhood user gathers a kth user in K (i) vector corresponding in implicit features matrix U, m is the sum of user, and n is the sum of Web service; S _ijfor the i-th row jth column element value in QoS variable matrix S, I _ijfor S _ijdesignator, if S _ijfor null then I _ij=0, otherwise I _ij=1; || || ₂for 2-norm, ^trepresent transposition, α and β is given control coefrficient;

(4) accept the query composition request of user about Web service, then the candidate combinations list of the Web service of its requesting query is provided to this user according to QoS prediction matrix R.

Set up the QoS variable matrix S between user and Web service according to QoS data in described step (1), be specially: the dimension of described QoS variable matrix S is the i-th row jth column element value S in m × n, QoS variable matrix S _ijadopt i-th user for the QoS data of a jth Web service, if a jth Web service is crossed in never call before i-th user, then element value S _ijfor null.

Adopt Kmeans algorithm to determine the neighborhood user set of each user in described step (2), detailed process is: first, adopts Kmeans algorithm to carry out cluster to total user according to described QoS variable matrix S; Then for arbitrary user, get the neighborhood user that of a sort N number of user with this user forms this user and gather.

K value in described Kmeans algorithm is set as 10.

Adopt in described step (3) dictionary learning algorithm to minimize target function J to solve, and take turns in iterative process at each and record intermediate variable process.

In described step (4), finally the candidate combinations list of the Web service of user institute requesting query is packaged into html page formatting is presented to user.

The present invention can predict the QoS information of service efficiently by the matrix decomposition algorithm merging user's historical data, and supports the personalized service requirements of combination meeting user; In addition, the present invention improves solution efficiency greatly by matrix decomposition speed technology, thus the personalized QoS query composition request of real-time response multi-user.

Accompanying drawing explanation

Fig. 1 is the schematic flow sheet of the web service composition method that the present invention is based on user collaborative regularization.

Fig. 2 is the internal process schematic diagram of core QoS prediction algorithm engine of the present invention.

Embodiment

In order to more specifically describe the present invention, below in conjunction with the drawings and the specific embodiments, technical scheme of the present invention is described in detail.

As shown in Figure 1, the overall procedure that the present invention is based on the personalized web service composition method of user collaborative regularization comprises with lower part:

Flow process 1: the history gathering user calls service QoS data (throughput and response time etc.).Suppose total m user and a n service, so use historical data to produce the QoS matrix S of the user-service of a m × n: wherein each S _ijthat the QoS of user i to service j calls situation.

Flow process 2: the front-end portal page accepts targeted customer's Services Composition inquiry request.

Flow process 3: the inquiry request that flow process 2 gathers is performed an analysis:

If systems axiol-ogy is complete to QoS matrix S, so the QoS inquired about according to user is retrained composite services by system, and generates the front end display engine that candidate service list is sent to flow process 6, generates result back page.If systems axiol-ogy is not complete to QoS matrix S, system is so needed to carry out the fast prediction algorithm engine of flow process 4.

Flow process 4: the QoS prediction algorithm engine based on user collaborative regularization is execution entity of the present invention; As shown in Figure 2, the sub-process of this algorithm engine comprises following a few part:

4.1 calculate according to history QoS data the best neighborhood finding targeted customer.System uses classical Kmeans clustering algorithm.System is arranged to 10 K value, and handle is defined as best neighborhood K (i) with the of a sort customer group of targeted customer i.

4.2 structure optimization equations.In order to implementation algorithm can calculate to a nicety the qos value of targeted customer, first the present invention introduces classical SVD Predicting Technique as algorithm template:

$min\frac{1}{2}{\mathrm{\Σ}}_{i=1}^m{\mathrm{\Σ}}_{j=1}^n{I}_{ij}{(S_{ij}-U_i^T{V}_j)}^2+\frac{\mathrm{\α}}{2}\left|\right|U|{|}_2^2+\frac{\mathrm{\α}}{2}|\left|V\right|{|}_2^2---\left(1\right)$

Wherein: I _ijthe designator (I of original matrix _ij=1 works as S _ijthere is QoS record, on the contrary I _ij=0).U and V is respectively the implicit features matrix of user and the implicit features matrix of service.Latter two is regularization term, avoids U and V over-fitting in primal algorithm model.SVD Predicting Technique produces U and V satisfied condition, all QoS information of last reconstruction matrix S by minimizing formula (1).

For the neighborhood that sub-process 4.1 delimited, the present invention uses for reference the mass-rent thought of current main-stream, is predicted by the qos value of implicit features to targeted customer fully learning best neighborhood.The data characteristics how learning neighborhood is the key component of collaborative filtering thought.The inherent attribute of user is by describing in the present invention, and according to general knowledge, under stable case, the implicit features difference of the user of same group should be very little; Therefore the present invention sets up user collaborative regularization term with this condition:

$min\left|\right|U_i-\frac{1}{\left|K\left(i\right)\right|}{\mathrm{\Σ}}_{j\∈K\left(i\right)}{U}_j|{|}_2^2---\left(2\right)$

Wherein: K (i) defines by sub-process i.Formula (2) fully have learned the feature of neighborhood in average weighted mode, then uses restraint to the inherent attribute of targeted customer; Above-mentioned formula is the mathematical table method mode that the present invention proposes to suppose.

After obtaining above-mentioned formula, targeted customer is expanded inherent attribute in the present invention and traditional SVD Predicting Technique merges, and produces and meets the optimization equation that Web service QoS predicts scene:

$min\frac{1}{2}{\mathrm{\Σ}}_{i=1}^m{\mathrm{\Σ}}_{j=1}^n{I}_{ij}{(Q_{ij}-U_i^T{S}_j)}^2+\frac{\mathrm{\α}}{2}\left|\right|U|{|}_2^2+\frac{\mathrm{\α}}{2}|\left|S\right|{|}_2^2+\frac{\mathrm{\β}}{2}\left|\right|U_i-\frac{1}{\left|K\left(i\right)\right|}{\mathrm{\Σ}}_{j\∈K\left(i\right)}{U}_j|{|}_2^2---\left(3\right)$

4.3 dictionary learning methods solve.

For formula (3), conventional method uses gradient descent method to solve.But for the characteristic of formula (3), the simple gradient descent method that uses will produce a large amount of duplicate keys.These duplicate keys consume a large amount of computational resources, and do not produce effect to the final prediction effect of lifting.In order to solve duplicate keys problem, present invention uses dictionary learning method record intermediate vector table, directly accelerating the solution procedure of formula (3).

Take turns iteration for t, the more new variables of the present invention first in statistical formula (3) iteration also builds the following information of intermediate variable table record that 1 length is m:

$min\left|\right|U_i-\frac{1}{\left|K\left(i\right)\right|}{\mathrm{\Σ}}_{j\∈K\left(i\right)}{U}_j|{|}_2^2---\left(4\right)$

Last the present invention uses dictionary learning method to carry out renewal multivariable process:

$U_i^{(t+1)}=U_i^{\left(t\right)}-\mathrm{\θ}\frac{\∂J^{(t+1)}}{\∂U_i}$

$V_j^{(t+1)}=V_j^{\left(t\right)}-\mathrm{\θ}\frac{\∂J^{(t+1)}}{\∂V_j}$

Wherein:

$\frac{\∂J^{(t+1)}}{\∂U_i}=\underset{j=1}{\overset{n}{\Σ}}{I}_{ij}(S_{ij}-U_i^{\left(t\right)T}{V}_j^{\left(t\right)})(-V_j^{\left(t\right)})+{\mathrm{\αU}}_i^{\left(t\right)}+{\mathrm{\βU}}_i^{{\left(t\right)}_{*}}$

$\frac{\∂J^{(t+1)}}{\∂V_j}=\underset{i=1}{\overset{m}{\Σ}}{I}_{ij}(S_{ij}-U_i^{\left(t\right)T}{V}_j^{\left(t\right)})(-U_i^{\left(t\right)})+{\mathrm{\αV}}_j^{\left(t\right)}$

Whether 4.4 target functions meet convergence.

For matrix variables U and the V of iteration generation each time, we need the result of more new formula (3).The end condition of iteration is:

J-J'≤ε (5) is wherein: ε is iteration threshold, usual ε=0.001.

If loss function meets above-mentioned end condition, then iterative process stops.If do not meet, then return sub-process 4.2 and continue iteration, until satisfy condition.

Flow process 5: for eigenmatrix U and V satisfied condition, the present invention is by all QoS information of matrix combination reconstruct original matrix R:

R＝U ^TV(6)

Flow process 6: front end page accepts the demand of user individual Services Composition, background engine is according to inquiry request automation composite services and be pushed to front-end module, and front-end module shows service prioritizing.

Test result:

In order to the QoS Forecasting Methodology representing the present invention's proposition quantized is conducive to improving the satisfaction of user's Services Composition request, we used the accuracy (the larger effect of F1 is better) that traditional F1 index carrys out evaluation prediction.In addition due to system generation is candidate service list, and we are labeled as F1K, and wherein K is that service list generates length.

The data centralization that experiment uses contains the detailed recalls information of 4532 client users to 59492 Web services, and therefore we use the client user of 4532 × 59492 sizes-Web service matrix to store.In an experiment, this matrix has been divided into two parts: 20% training sample and 80% test sample book.We are core-prediction algorithm called after Reg.

Table 1

	F13	F15	F110
				Reg	2.6053	3.4716	4.4007
L2R	1.5566	2.4364	2.3425
				GDistance	1.4790	2.8078	3.2805
NDCG	1.8734	2.2705	3.1093

As shown in table 1, as compared to present method L2R, GDistance with NDCG, the F1 value of the inventive method Reg is under any circumstance all larger, and namely combined effect is better.

Above-mentioned is can understand and apply the invention for ease of those skilled in the art to the description of embodiment.Person skilled in the art obviously easily can make various amendment to above-described embodiment, and General Principle described herein is applied in other embodiments and need not through performing creative labour.Therefore, the invention is not restricted to above-described embodiment, those skilled in the art are according to announcement of the present invention, and the improvement made for the present invention and amendment all should within protection scope of the present invention.

Claims (6)

1., based on a personalized web service composition method for user collaborative regularization, comprise the steps:

(1) collect the service call data that all users provide, described service call data comprise the QoS data of user to all Web services that it called; And then set up the QoS variable matrix S between user and Web service according to described QoS data;

(2) the neighborhood user set of each user is determined according to described QoS variable matrix S;

(3) gather according to the neighborhood user of described QoS variable matrix S and each user, by the following target function J of SVD method establishment, and this target function J is minimized solve, in the hope of the implicit features matrix U about user and the implicit features matrix V about Web service; And then according to R=U ^tv rebuilds the QoS prediction matrix R between user and Web service;

$J=\frac{1}{2}\underset{i=1}{\overset{m}{\Σ}}\underset{j=1}{\overset{n}{\Σ}}{I}_{ij}{(S_{ij}-U_i^T{V}_j)}^2+\frac{\mathrm{\α}}{2}\left|\right|U|{|}_2^2+\frac{\mathrm{\α}}{2}|\left|V\right|{|}_2^2+\frac{\mathrm{\β}}{2}\left|\right|U_i-\frac{1}{N}\underset{k\∈K\left(i\right)}{\Σ}{U}_k|{|}_2^2$

Wherein: U _ifor the i-th column vector in implicit features matrix U, V _jfor the jth column vector in implicit features matrix V, K (i) is the neighborhood user set of i-th user, and N is the total number of users that default neighborhood user number and neighborhood user gather in K (i), U _kfor neighborhood user gathers a kth user in K (i) vector corresponding in implicit features matrix U, m is the sum of user, and n is the sum of Web service; S _ijfor the i-th row jth column element value in QoS variable matrix S, I _ijfor S _ijdesignator, if S _ijfor null then I _ij=0, otherwise I _ij=1; || || ₂for 2-norm, ^trepresent transposition, α and β is given control coefrficient;

(4) accept the query composition request of user about Web service, then the candidate combinations list of the Web service of its requesting query is provided to this user according to QoS prediction matrix R.

2. personalized web service composition method according to claim 1, it is characterized in that: in described step (1), set up the QoS variable matrix S between user and Web service according to QoS data, be specially: the dimension of described QoS variable matrix S is the i-th row jth column element value S in m × n, QoS variable matrix S _ijadopt i-th user for the QoS data of a jth Web service, if a jth Web service is crossed in never call before i-th user, then element value S _ijfor null.

3. personalized web service composition method according to claim 1, it is characterized in that: in described step (2), adopt Kmeans algorithm to determine the neighborhood user set of each user, detailed process is: first, adopts Kmeans algorithm to carry out cluster to total user according to described QoS variable matrix S; Then for arbitrary user, get the neighborhood user that of a sort N number of user with this user forms this user and gather.

4. personalized web service composition method according to claim 3, is characterized in that: the K value in described Kmeans algorithm is set as 10.

5. personalized web service composition method according to claim 1, it is characterized in that: adopt in described step (3) dictionary learning algorithm to minimize target function J and solve, and take turns in iterative process at each and record intermediate variable process.

6. personalized web service composition method according to claim 1, is characterized in that: in described step (4), finally the candidate combinations list of the Web service of user institute requesting query is packaged into html page formatting is presented to user.