CN103473291A - Personalized service recommendation system and method based on latent semantic probability models - Google Patents

Abstract

The invention relates to a personalized service recommendation system and method based on a latent semantic probability models and belongs to the technical field of service computing. The method includes: determining a QoS (quality of service) index system for evaluating performance of a series of services with similar functions; building latent semantic probability models among users, user index preferences and service situations; collecting index preference information provided by user-system interaction when different users use services with different functions under different service situations in a system, and saving the information in a historical experience database; using collected data to train parameters of the latent semantic probability models; when a user is unfamiliar with a certain service situation and needs to call a service with a special function under the service situation, using the trained latent semantic probability models to predict index preference of the user; comprehensively screening candidate services according to the predicted personalized QoS index preference, selecting the service most suitable for the user's requirements, and personalized service recommendation is achieved.

Description

A kind of personalized service recommendation system and method based on enigmatic language justice probability model

Technical field

The invention belongs to the service compute technical field, be specifically related to a kind of personalized service recommendation system and method based on enigmatic language justice probability model.

Background technology

Develop rapidly along with Internet technology, the service compute technology is widely applied, Web service is exactly the loose coupling software systems that a kind of like this distribution runs on Internet, supports interoperability between different platform, it mainly allows between service user and supplier to form loose binding relationship by the pattern of " issue-search-binding ", and this use for service is laid a good foundation.But the characteristic that the user of web services and supplier are separated, increased the difficulty that the service user understands service, simultaneously along with the web services quantity of the upper operation of Internet is on the increase, the service user need to comform and select or the one group of service that meets self-demand most in service like multi-functional photograph, this is undoubtedly a heavy task concerning the service user of great majority shortage professional knowledge, so develop the inevitable demand that effective service recommendation technology is services selection.

Through the retrieval to prior art, find, Chinese Patent Application No. 200710162463.8, put down in writing a kind of self-adapting service recommendation equipment, and this device mainly comprises: semantic analysis device, service selection device, service recommendation device; The semantic analysis device is for carrying out analysis semantically to user's inquiry; Service selection device is used for finding out the service of the selection corresponding with the inquiry after semantic analysis, and according to the service update service Relational database of selecting; Last service recommendation device is searched the service related data storehouse to the user, to recommend related service for the service that utilizes the selection of obtaining.

Further retrieval is found, Chinese Patent Application No. 200910236492.3, a kind of personalized service recommendation system and method have been put down in writing, the method mainly comprises: the various operation informations that the monitoring of user information collection device is carried out in terminal, and carry out depositing User Information Database in after pre-service, if the User Information Database method is upgraded, start the user behavior analysis device and analyzed; User behavior analysis device scanning User Information Database, extract new user profile and count resource information database, calculates new recommendation strategy and charge to the recommendation policy database; Context-aware processor perception user's current context, export current context-descriptive information, starts the personalized recommendation processor; After the personalized recommendation processor receives the message from the context-aware processor, obtain current context information, recommend policy database by retrieval, strategy is recommended in the left and right that obtains coupling, and recommend strategy by the retrieve resources information database according to left and right, mate suitable resource information, generate in real time the personalized recommendation service.

Further retrieval is found, Chinese Patent Application No. 20121014234.2 has recorded the spatial Information Service matching process of a kind of based on the context perception and user preference, on the basis of the functional coupling of user's services request and candidate's spatial Information Service and non-functional coupling, consider the user preference of context-sensitive, calculate the matching degree of each candidate's spatial Information Service in users service needs and intelligent space, then according to matching degree, candidate's spatial Information Service is recommended to the user; Although the user may have preference fixing or that repeat, but these preferences are not to be correlated with at any time, the spatial Information Service matching process of this invention based on the context perception and user preference is simplified user preference from contextual angle, rejects the user preference irrelevant with the user; This invention is carried out matching degree calculating from function match and two aspects of NOT-function coupling, improves the accuracy rate of spatial Information Service coupling.

Further retrieval is found, Chinese Patent Application No. 201210253884.2 has recorded a kind of individuation search method of recommending for Web service, comprise the following steps: step 1, pre-service WSDL document: by removing stop words and extracting two pre-treatment step of stem, form the word bag; Step 2, extract user interest: use improved TF-IDF formula to calculate the weight of each word in the word bag, and be multiplied by the time decay factor of this word, obtain new weight; Before selecting weight from large to small, k word be as user's interest word, and the respective weights of each word, forms the user interest vector that k ties up; Step 3, calculate Interest Similarity: set similarity threshold, the user who surpasses threshold value enters to elect as targeted customer's neighbor user; Step 4, sequence service retrieval result, according to the similarity of neighbor user and select the recommendation predicted value of the number of times calculation services of service, and by result for retrieval according to recommending the predicted value descending sort, thereby obtain personalized search results.

Above-mentioned method has all related to personalized service recommendation, but seldom for the NOT-function attribute selection of Web service, serve to meet user's individual demand, the NOT-function attribute of service is the major embodiment of service performance often, so performance how to utilize the NOT-function attribute of service to estimate objectively service can meet again the individual demand of different user simultaneously, is the problem that the service recommendation needs solve.

Summary of the invention

The technology of the present invention is dealt with problems: the deficiency existed for prior art, a kind of personalized service recommendation system and method based on enigmatic language justice probability model is provided, be intended to provide under unfamiliar service situation for the user prediction of index preference, thereby complete the service colligate sequence based on the multi-QoS index, for the user provides personalized service recommendation result.

Technical solution of the present invention: the user's index preference Forecasting Methodology based on enigmatic language justice probability model in a kind of personalized service recommendation, concrete steps are as follows:

Step 1, definite service QoS index system of estimating the service performance quality

Described service QoS index system refer to that whole Web service systematic unity adopts for estimating the set of the good and bad QoS index used of a series of functional similarity service performances, different systems can be chosen QoS index system that suitable QoS index forms oneself as required for estimating the performance quality of service;

Step 2, set up the enigmatic language justice probability model between user, user's index preference and service situation three

Described user's index preference refers to the preference degree of user to each QoS index in the index system of service QoS described in step 1, to the preference degree value of each QoS index, between 0 to 1, and is 1 to the preference value summation of each QoS index; Described service situation refers to the service of using which kind of function under which kind of scene, each service situation tlv triple (w for e ₁, w ₂, w ₃) expression, wherein w ₁mean service that complete with basic function independent of service, as video capability, navigation feature etc., w ₂mean the concrete business activity that the user uses the service of this function to complete, as academic conference, military navigation etc., w ₃mean that the user calls the terminal device of service, as mobile phone, PC etc.; Enigmatic language justice probability model between described user, user's index preference, service situation three refers to that unique user relies on the hidden class of different users with different probability and exists, single serve situation and rely on the different hidden classes of service situation with different probability and exist, the index preference relies on the hidden class of different users simultaneously and serves the probability model that the hidden class of situation exists with different probability simultaneously, and the image conversion of model means as shown in 7.

The hidden class of described user refers to that non-artificial pre-determined but user that obtain from historical experience data learning clusters; The hidden class of described service situation refers to that non-artificial pre-determined but service situation that obtain from historical experience data learning clusters;

The index preference information that step 3, collection different user independently provide while using the difference in functionality service under different service situations, deposit database in as the historical experience data, for the parameter of the enigmatic language justice probability model set up in training step 2 is prepared, storage format is (user, service situation, index preference) tlv triple;

Step 4, by the parameter of EM algorithm and the historical experience data of having collected training enigmatic language justice probability model

The parameter of described enigmatic language justice probability model refers to the hidden class { U of all users ₁, U ₂..., U _iprior probability P (U _i) (1≤i≤I), the hidden class { E of service situation ₁, E ₂..., E _jprior probability P (E _j) (1≤j≤J), a given hidden class U of user _isituation under the conditional probability P (u|U that occurs of unique user u _i) (1≤i≤I), a given hidden class E of service situation _jsituation place an order and serve the conditional probability P (e|E that situation e occurs _j) (1≤j≤J), the hidden class U of given user _iwith the hidden class E of service situation _jsituation under the probability P (r|U that occurs of user's index preference vector r _i, E _j) (1≤i≤I, 1≤j≤J); I wherein, J means respectively total number of the hidden class of user and the hidden class of service situation, i, j means respectively the numbering of the hidden class of user and the hidden class of service situation;

Step 5, acquisition user need the request of service recommendation, comprise the service situation that personal information, service recommendation rely on;

The unknown index preference of enigmatic language justice probability model prediction designated user under the specific service situation that step 6, use have been trained;

The QoS index preference of the user individual that step 7, basis dope, carry out Integrated Selection to candidate service, thereby select the service of pressing close to this user's request most, and recommendation results is returned to the user.

The process of collecting the historical experience data in described step 3 is as follows:

(1) can be directly while calling familiar web services under user's service situation familiar at it and system interaction, use analytical hierarchy process to provide the comparative result in twos oneself provided each index preference, set up comparator matrix;

$C=\left[\begin{array}{cccc}{c}_{11},& c_{12},& ...,& c_{1n}\\ c_{21},& c_{22},& ...,& c_{3n}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ c_{n1},& c_{n2},& ...,& c_{\mathrm{nn}}\end{array}\right],$

(2) whether within the acceptable range the consistance of comparison matrix, if can accept, enters (3), and no person returns to (1);

(3) calculate the eigenvalue of maximum characteristic of correspondence vector of comparator matrix, by resulting after this proper vector normalization be exactly the preference vector W of user to each QoS index _qoS, deposit (user, service situation, index preference) tlv triple in database.

In described step 4, the parameter training process of enigmatic language justice probability model is as follows:

(1) take out all users, service situation, index preference tlv triple as training data from database, the personalized index preference weights that utilize analytical hierarchy process to provide under user's service situation familiar at it, each tlv triple (u, e, r) mean that user u preference weight to each QoS index of service under service situation e is r, r is a K dimensional vector (r ₁, r ₂..., r _k), for the sake of simplicity, suppose the separate and Normal Distribution of every one dimension of weight vectors r, available following two formula of joint probability P (u, e, r) mean,

$P(u,e,r)=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}P\left(U_i\right)P\left(E_j\right)P\left(u\right|U_i)P\left(e\right|E_j)P\left(r\right|U_i,E_j),$

P (r|U wherein _i, E _j) be the normal distribution of K dimension, i, j means respectively the hidden class of user and the hidden class numbering of service situation;

(2) the E step is calculated the associating posterior probability of hidden class,

$P(U_p,E_q|u,e,r)=\frac{{\left[P\left(U_p\right)P\left(E_q\right)P\left(u\right|U_p)P\left(e\right|E_q)P\left(r\right|U_p,E_q)\right]}^b}{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{\left[P\left(U_i\right)P\left(E_j\right)P\left(u\right|U_i)P\left(e\right|E_j)P\left(r\right|U_p,E_q)\right]}^b}.$

Wherein b is a simulated annealing parameter between 0 to 1, P (r|U _p, E _q) be the normal distribution of K dimension, p, q means the hidden class of user and the hidden class numbering of service situation;

(3) posterior probability that the M step is used the E step to calculate is the appraising model parameter again, comprises the prior probability P (U of the hidden class of each user and the hidden class of each service situation _p) and P (E _q), the conditional probability P (u|U that each user occurs when the hidden class of given different user _p), often serve first the conditional probability P (e|E that situation occurs when the hidden class of given different service situation _q), the normal distribution average of K index preference when the given different hidden class of user context and the hidden class of service situation simultaneously

and variance ${{\mathrm{\σ}}^2}_{k,U_p,E_q}:$

$P\left(U_p\right)=\frac{\underset{l}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}P(U_p,E_j|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L},$ $P\left(E_q\right)=\frac{\underset{l}{\mathrm{\Σ}}\underset{i}{\mathrm{\Σ}}P(U_i,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L},$

$P\left(u\right|U_p)=\frac{\underset{l:u_{\left(l\right)}=u}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}P(U_p,E_j|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L*P\left(U_p\right)},$ $P\left(e\right|E_q)=\frac{\underset{l:e_{\left(l\right)}=e}{\mathrm{\Σ}}\underset{i}{\mathrm{\Σ}}P(U_i,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L*P\left(E_q\right)},$

${\mathrm{\μ}}_{k,U_p,E_q}=\frac{\underset{l}{\mathrm{\Σ}}{r}_{k\left(l\right)}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{\underset{l}{\mathrm{\Σ}}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})},$ ${{\mathrm{\σ}}^2}_{k,U_p,E_q}=\frac{\underset{l}{\mathrm{\Σ}}{(r_{k\left(l\right)}-{\mathrm{\μ}}_{k,U_p,E_q})}^{2}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{\underset{l}{\mathrm{\Σ}}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}.$

Wherein l means the numbering of historical training sample, i, and p means the label of the hidden class of user, j, q means to serve the numbering of the hidden class of situation, u, e means unique user and singly serves situation, U, E means the hidden class of user and the hidden class of service situation, k means the numbering of QoS index.

(4) whether the inspection model parameter restrains, if convergence finishes and the preservation model parameter, if do not restrain, returns to (2) and carries out.

In described step 6, utilize the implementation of the enigmatic language justice probability model predictive user index preference of having trained as follows:

(1) user u ^tlog in service recommendation system, the service situation that need to carry out service recommendation e is provided ^t;

(2) the service contextual information that commending system obtains user's personal data and provides, with following formula to this user each QoS index preference under this service situation carry out independent prediction

$R_k(e^t,u^t)=\underset{0}{\overset{1}{\∫}}{r}_k\frac{P(e^t,u^t,r_k)}{\underset{0}{\overset{1}{\∫}}P(e^t,u^t,r_k)d{r}_k}d{r}_k$

Wherein

$P(e^t,u^t,r_k)=\underset{i,j}{\mathrm{\Σ}}P\left(U_i\right)P\left(E_j\right)P\left(u^t\right|U_i)P\left(e^t\right|E_j)P\left(r_k\right|U_i,E_j)$

$P\left(r_k\right|U_i,E_j)=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_{k,U_i,E_j}}\exp [-\frac{{(r_k-{\mathrm{\μ}}_{k,U_i,E_j})}^2}{2{{\mathrm{\σ}}^2}_{k,U_i,E_j}}].$

Finally prediction obtains user u ^tat this service situation e ^tunder to the preference weight of each QoS index, be $(R_1(e^t,u^t),...,R_K(e^t,u^t)),$ Simply be denoted as $({\hat{r}}_1,{\hat{r}}_2,...,{\hat{r}}_K).$

In described step 7, according to the QoS index preference of the user individual doped, the process that candidate service is screened is as follows: user u ^tat service situation e ^tunder obtained a plurality of intimate candidate service, suppose that the performance on each QoS index of each service can have additive method to obtain, be designated as (q ₁, q ₂..., q _k), the PTS of each service adopts so

mean, finally according to the PTS of each service, provide the sequence of service, as the foundation of service recommendation.

The present invention's advantage compared with prior art is: the probability that technical scheme provided by the invention has been set up between a kind of relevant user, user's index preference, service situation relies on model, with historical data to model training after, can utilize this model to dope easily the index preference vector of specific user under strange service situation, the last NOT-function attribute in conjunction with service and user's index preference of having predicted are carried out comprehensive evaluation to realize personalized service recommendation to candidate service.Existing method is general only to be weighted each dimension qos value of candidate service on average simply, provides accordingly the integrated ordered of candidate service, and weights are generally first determined, obviously can't embody the individual demand of user in service invocation procedure.The personalized service recommendation method based on enigmatic language justice probability model that this patent proposes can dope the personalized index preference of user under particular context in a kind of automatic mode, even when user oneself also can't the preference information of clear expression oneself.Model training of the present invention can carry out by off-line simultaneously, so improved the efficiency that online service is recommended.

The accompanying drawing explanation

Fig. 1 is the functional block diagram of system of the present invention;

The realization flow figure that Fig. 2 is historical information collection module in Fig. 1;

The realization flow figure that Fig. 3 is enigmatic language justice probability model parameter training module in Fig. 1;

The realization flow figure that Fig. 4 is service recommendation request module in Fig. 1;

The realization flow figure that Fig. 5 is personalized index preference prediction module in Fig. 1;

The realization flow figure that Fig. 6 is personalized service recommendation module in Fig. 1;

The image conversion that Fig. 7 is enigmatic language justice probability model means.

Embodiment

As shown in Figure 1, a kind of personalized service recommendation system and method based on enigmatic language justice probability model of the present invention is comprised of historical information collection module, enigmatic language justice probability model parameter training module, service recommendation request module, personalized index preference prediction module, personalized service recommendation module.

Whole implementation procedure is as follows:

Step 1, definite service QoS index system of estimating the service performance quality

Described service QoS index system refer to that whole Web service systematic unity adopts for estimating the set of the good and bad QoS index used of a series of functional similarity service performances, different systems can be chosen QoS index system that suitable QoS index forms oneself as required for estimating the performance quality of service;

Step 2, set up the enigmatic language justice probability model between user, user's index preference and service situation three

Described user's index preference refers to the preference degree of user to each QoS index in the index system of service QoS described in step 1, to the preference degree value of each QoS index, between 0 to 1, and is 1 to the preference value summation of each QoS index; Described service situation refers to the service of using which kind of function under which kind of scene, each service situation tlv triple (w for e ₁, w ₂, w ₃) expression, wherein w ₁mean service that complete with basic function independent of service, as video capability, navigation feature etc., w ₂mean the concrete business activity that the user uses the service of this function to complete, as academic conference, military navigation etc., w ₃mean that the user calls the terminal device of service, as mobile phone, PC etc.; Enigmatic language justice probability model between described user, user's index preference, service situation three refers to that unique user relies on the hidden class of different users with different probability and exists, single serve situation and rely on the different hidden classes of service situation with different probability and exist, the index preference relies on the hidden class of different users simultaneously and serves the probability model that the hidden class of situation exists with different probability simultaneously, and the image conversion of model means as shown in Figure 7.The hidden class of U representative of consumer in figure, the hidden class of E representative service situation, u, e represent respectively unique user and beat a service situation, r represents the index preference vector, Fig. 7 has shown the probability dependence between these key elements, user u, service situation e, index preference vector r weigh with the size of conditional probability P (u|U), P (e|E), P (r|U, E) the size of the dependency degree of the hidden class U of user and the hidden class E of service situation.

The hidden class of described user refers to that non-artificial pre-determined but user that obtain from historical experience data learning clusters; The hidden class of described service situation refers to that non-artificial pre-determined but service situation that obtain from historical experience data learning clusters;

The index preference information that step 3, collection different user independently provide while using the difference in functionality service under different service situations, deposit database in as the historical experience data, for the parameter of the enigmatic language justice probability model set up in training step 2 is prepared, storage format is (user, service situation, index preference) tlv triple;

Step 4, by the parameter of EM algorithm and the historical experience data of having collected training enigmatic language justice probability model

The parameter of described enigmatic language justice probability model refers to the hidden class { U of all users ₁, U ₂..., U _iprior probability P (U _i) (1≤i≤I), the hidden class { E of service situation ₁, E ₂..., E _jprior probability P (E _j) (1≤j≤J), a given hidden class U of user _isituation under the conditional probability P (u|U that occurs of unique user u _i) (1≤i≤I), a given hidden class E of service situation _jsituation place an order and serve the conditional probability P (e|E that situation e occurs _j) (1≤j≤J), the hidden class U of given user _iwith the hidden class E of service situation _jsituation under the probability P (r|U that occurs of user's index preference vector r _i, E _j) (1≤i≤I, 1≤j≤J); I, what j represents respectively, I wherein, J means respectively total number of the hidden class of user and the hidden class of service situation, i, j means respectively the numbering of the hidden class of user and the hidden class of service situation;

Step 5, acquisition user need the request of service recommendation, comprise the service situation that personal information, service recommendation rely on;

The unknown index preference of enigmatic language justice probability model prediction designated user under the specific service situation that step 6, use have been trained;

The QoS index preference of the user individual that step 7, basis dope, carry out Integrated Selection to candidate service, thereby select the service of pressing close to this user's request most, and recommendation results is returned to the user.

The specific implementation process of above-mentioned each module is as follows:

1. historical information collection module

Implementation procedure is as shown in Figure 2: collecting user's personal information, comprise name, occupation, annual income, hobby etc., is numbering of each user assignment; Collecting the user provides the index preference residing service contextual information, and the service situation is expressed as tlv triple (service basic function, goal activities, terminal device), such as (navigation feature, go off daily, mobile phone) or (phonetic function, academic conference, PC) etc.Suppose that actually determined each QoS dimension is: reliability, response time, availability, handling capacity, price, the comparator matrix that the user provides under this service situation so is just the Matrix C of 5*5, calculate its eigenvalue of maximum characteristic of correspondence vector normalization, be the index preference, by the user, the service situation, index preference tlv triple deposits database in.

2, enigmatic language justice probability model parameter training module

Implementation procedure is as shown in Figure 3: the historical training data in database is taken out, then the number of the hidden class of user and the hidden class of service manually is set, suppose to be all 5, and the initial parameter of enigmatic language justice probability model: the initial prior probability of 5 hidden classes of user can all be made as 1/5; The probability of 5 hidden classes of service situation, can all be made as 1/5; Each user is at the conditional probability initial value of the hidden class of given different user; The every conditional probability initial value of situation in the hidden class of given different service situation of serving first; The normal distribution average of each index preference when the hidden class of given different user and service situation hidden class and the initial value of variance, then by the parameter of EM algorithm combined training data iteration training pattern until convergence parameter is preserved.

3. service recommendation request module

As shown in Figure 4, the user login services commending system, suppose that this user has been the registered user of system to implementation procedure, and specific login name sign is arranged, and is designated as u ^t, he provides the service situation that need to carry out service recommendation to be designated as e ^t, concrete meaning is (exchange rate conversion, shopping at network, PC), this module can be recorded u ^tand e ^t, submit to personalized index preference prediction module and personalized service recommendation module.

4. personalized index preference prediction module, implementation procedure as shown in Figure 5.

This module need to be obtained the service recommendation solicited message that the model parameter that trained and user provide, and based on index preference Forecasting Methodology presented above, dopes user's index preference, supposes that the index preference of prediction is r=(0.3,0.1,0.2,0.3,0.1).

5. personalized service recommendation module

Implementation procedure as shown in Figure 6, supposes that the service that meets exchange rate conversion has 3 s1, s2, s3, their normalized values on five QoS dimensions can be obtained by other modes, suppose to be designated as respectively: q1=(0.5,0.5,0.6,0.7,0.6), q2=(0.8,0.4,0.7,0.2,0.6), q3=(0.4,0.8,0.2,0.6,0.4), the score of three candidate service is respectively so:

g1=r·q1=(0.3,0.1,0.2,0.3,0.1)·(0.5,0.5,0.6,0.7,0.6)=0.59；

g2=r·q2=(0.3,0.1,0.2,0.3,0.1)·(0.8,0.4,0.7,0.2,0.6)=0.54；

g3=r·q3=(0.3,0.1,0.2,0.3,0.1)·(0.4,0.8,0.2,0.6,0.4)=0.46；

So can obtain g1 > g2 > g3, so provided the service ranking list, be s1, s2, s3, return to the user and select for the user.

Non-elaborated part of the present invention belongs to techniques well known.

The above; be only part embodiment of the present invention, but protection scope of the present invention is not limited to this, in the technical scope that any those skilled in the art disclose in the present invention; the variation that can expect easily or replacement, within all should being encompassed in protection scope of the present invention.

Claims (7)

1. the personalized service recommendation system based on enigmatic language justice probability model is characterized in that comprising:

Historical information collection module: collect the index preference comparator matrix that different user independently provides by analytical hierarchy process under difference service situation, after checking that this matrix meets consistency constraint, calculate this matrix eigenvalue of maximum characteristic of correspondence vector, draw the index preference of different user under difference service situation after normalization, and by the user, service situation, index preference deposit database in triple form, for enigmatic language justice probability model parameter training module provides historical training data;

Enigmatic language justice probability model parameter training module: obtain the user that the historical information collection module is collected from database, service situation, index preference tlv triple are as training sample, enigmatic language justice probability model parameter initial value according to prior setting, utilize EM algorithm iteration training pattern parameter on training data, until model parameter convergence, preserve the gained training parameter, offer personalized index preference prediction module;

User's service recommendation request module: the needs that this module is obtained individual character data information that user login is to provide and the user logins rear typing carry out the service contextual information of service recommendation, finally these two information are offered to personalized index preference prediction module;

Personalized index preference prediction module: obtain userspersonal information and service contextual information that user's service recommendation request module provides, the model parameter trained provided in conjunction with enigmatic language justice probability model parameter training module, predict the index preference of this user under this service situation, and provided the personalized service recommendation module by the index preference of prediction;

Personalized service recommendation module: receive the user's index preference from personalized index preference module, reach the service contextual information from user's service recommendation request module, obtain the similar service of all functions that is applicable to this service situation, and the QoS desired value of each dimension of service, make dot product with the index preference of user individual, provide the integrated ordered of service, as recommendation list, return to the user, the user selects own interested project to be paid close attention to according to the result of recommending.

2. the personalized service recommendation system based on enigmatic language justice probability model according to claim 1 is characterized in that: described enigmatic language justice probability model parameter training module implementation procedure:

(1) obtain user, service situation, the index preference tlv triple that the historical information collection module deposits in from database, as historical training data;

(2) number of the hidden class of user and the hidden class of service situation manually is set, and the parameter initial value of model, the prior probability that comprises the hidden class of user and the hidden class of service situation, the conditional probability of unique user when the hidden class of given different user, single serve the conditional probability of situation when the hidden class of given different service situation, and normal distribution average and the variance of each index preference when the hidden class of given different user and the hidden class of different service situation;

(3) use the EM algorithm, in conjunction with historical training data, model parameter is carried out to the iteration training, until convergence, the preservation model parameter, offer personalized index preference prediction module.

3. the personalized service recommendation method based on enigmatic language justice probability model is characterized in that performing step is as follows:

Step 1, definite service QoS index system of estimating the service performance quality

Described service QoS index system refer to that whole Web service systematic unity adopts for estimating the set of the good and bad QoS index used of a series of functional similarity service performances, different systems can be chosen QoS index system that suitable QoS index forms oneself as required for estimating the performance quality of service;

Step 2, set up the enigmatic language justice probability model between user, user's index preference and service situation three

Described user's index preference refers to the preference degree of user to each QoS index in the index system of service QoS described in step 1, to the preference degree value of each QoS index, between 0 to 1, and is 1 to the preference value summation of each QoS index; Described service situation refers to the service of using which kind of function under which kind of scene, each service situation tlv triple (w for e ₁, w ₂, w ₃) expression, wherein w ₁mean service that complete with basic function independent of service, as video capability, navigation feature etc., w ₂mean the concrete business activity that the user uses the service of this function to complete, w ₃mean that the user calls the terminal device of service;

Enigmatic language justice probability model between described user, user's index preference, service situation three refers to that unique user relies on the hidden class of different users with different probability and exists, single serve situation and rely on the different hidden classes of service situation with different probability and exist, the index preference relies on the hidden class of different users simultaneously and serves the probability model that the hidden class of situation exists with different probability simultaneously

The hidden class of described user refers to that non-artificial pre-determined but user that obtain from historical experience data learning clusters; The hidden class of described service situation refers to that non-artificial pre-determined but service situation that obtain from historical experience data learning clusters;

The index preference information that step 3, collection different user independently provide while using the difference in functionality service under different service situations, deposit database in as the historical experience data, for the parameter of the enigmatic language justice probability model set up in training step 2 is prepared, storage format is the user, service situation, index preference tlv triple;

Step 4, by the parameter of EM algorithm and the historical experience data of having collected training enigmatic language justice probability model

The parameter of described enigmatic language justice probability model refers to the hidden class { U of all users ₁, U ₂..., U _iprior probability P (U _i) (1≤i≤I), the hidden class { E of service situation ₁, E ₂..., E _jprior probability P (E _j) (1≤j≤J), a given hidden class U of user _isituation under the conditional probability P (u|U that occurs of unique user u _i) (1≤i≤I), a given hidden class E of service situation _jsituation place an order and serve the conditional probability P (e|E that situation e occurs _j) (1≤j≤J), the hidden class U of given user _iwith the hidden class E of service situation _jsituation under the probability P (r|U that occurs of user's index preference vector r _i, E _j) (1≤i≤I, 1≤j≤J); I wherein, J means respectively total number of the hidden class of user and the hidden class of service situation, i, j means respectively the numbering of the hidden class of user and the hidden class of service situation;

Step 5, acquisition user need the request of service recommendation, comprise the service situation that personal information, service recommendation rely on;

The unknown index preference of enigmatic language justice probability model prediction designated user under the specific service situation that step 6, use have been trained;

The QoS index preference of the user individual that step 7, basis dope, carry out Integrated Selection to candidate service, thereby select the service of pressing close to this user's request most, and recommendation results is returned to the user.

4. the personalized service recommendation method based on enigmatic language justice probability model according to claim 3, it is characterized in that: the process of collecting the historical experience data in described step 3 is as follows:

(1) while calling familiar web services under user's service situation familiar at it, directly and system interaction, use analytical hierarchy process to provide the comparative result in twos oneself provided each index preference, set up comparator matrix;

$C=\left[\begin{array}{cccc}{c}_{11},& c_{12},& ...,& c_{1n}\\ c_{21},& c_{22},& ...,& c_{3n}\\ .& .& & .\\ .& .& & .\\ .& .& & .\\ c_{n1},& c_{n2},& ...,& c_{\mathrm{nn}}\end{array}\right],$

(2) whether within the acceptable range the consistance of comparison matrix, if can accept, enters step (3), and no person returns to step (1);

(3) calculate the eigenvalue of maximum characteristic of correspondence vector of comparator matrix, by resulting after this proper vector normalization be exactly the preference vector W of user to each QoS index _qoS, deposit user, service situation, index preference tlv triple in database.

5. the personalized service recommendation method based on enigmatic language justice probability model according to claim 3, it is characterized in that: in described step 4, the parameter training process of enigmatic language justice probability model is as follows,

(1) take out all users, service situation, index preference tlv triple as training data from database, the personalized index preference weights that utilize analytical hierarchy process to provide under user's service situation familiar at it, each tlv triple (u, e, r) mean that user u preference weight to each QoS index of service under service situation e is r, r is a K dimensional vector (r ₁, r ₂..., r _k), for the sake of simplicity, suppose the separate and Normal Distribution of every one dimension of weight vectors r, joint probability P (u, e, r) means with following two formula,

$P(u,e,r)=\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}P\left(U_i\right)P\left(E_j\right)P\left(u\right|U_i)P\left(e\right|E_j)P\left(r\right|U_i,E_j),$

P (r|U wherein _i, E _j) be the normal distribution of K dimension, i, j means respectively the hidden class of user and the hidden class numbering of service situation, K is index number in the QoS index system;

(2) the E step is calculated the associating posterior probability of hidden class,

$P(U_p,E_q|u,e,r)=\frac{{\left[P\left(U_p\right)P\left(E_q\right)P\left(u\right|U_p)P\left(e\right|E_q)P\left(r\right|U_p,E_q)\right]}^b}{\underset{i}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}{\left[P\left(U_i\right)P\left(E_j\right)P\left(u\right|U_i)P\left(e\right|E_j)P\left(r\right|U_p,E_q)\right]}^b}.$

Wherein b is a simulated annealing parameter between 0 to 1, P (r|U _p, E _q) meet the normal distribution of K dimension, i, p means the hidden class numbering of user, j, q means to serve the hidden class numbering of situation;

(3) the hidden class of user that the M step is used the E step to calculate and the associating posterior probability of the hidden class of service situation be the appraising model parameter again, comprises the prior probability P (U of the hidden class of each user and the hidden class of each service situation _p) and P (E _q), the conditional probability P (u|U that each user occurs when the hidden class of given different user _p), often serve first the conditional probability P (e|E that situation occurs when the hidden class of given different service situation _q), the normal distribution average of K index preference when the given different hidden class of user context and the hidden class of service situation simultaneously

and variance

$P\left(U_p\right)=\frac{\underset{l}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}P(U_p,E_j|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L},$ $P\left(E_q\right)=\frac{\underset{l}{\mathrm{\Σ}}\underset{i}{\mathrm{\Σ}}P(U_i,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L},$

$P\left(u\right|U_p)=\frac{\underset{l:u_{\left(l\right)}=u}{\mathrm{\Σ}}\underset{j}{\mathrm{\Σ}}P(U_p,E_j|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L*P\left(U_p\right)},$ $P\left(e\right|E_q)=\frac{\underset{l:e_{\left(l\right)}=e}{\mathrm{\Σ}}\underset{i}{\mathrm{\Σ}}P(U_i,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{L*P\left(E_q\right)},$

${\mathrm{\μ}}_{k,U_p,E_q}=\frac{\underset{l}{\mathrm{\Σ}}{r}_{k\left(l\right)}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{\underset{l}{\mathrm{\Σ}}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})},$ ${{\mathrm{\σ}}^2}_{k,U_p,E_q}=\frac{\underset{l}{\mathrm{\Σ}}{(r_{k\left(l\right)}-{\mathrm{\μ}}_{k,U_p,E_q})}^{2}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}{\underset{l}{\mathrm{\Σ}}P(U_p,E_q|u_{\left(l\right)},e_{\left(l\right)},r_{\left(l\right)})}.$

Wherein l means the numbering of historical training sample, L means the sum of training sample, i, p means the label of the hidden class of user, j, q means to serve the numbering of the hidden class of situation, u, e means unique user and single situation, U of serving, E means respectively the hidden class of user and the hidden class of service situation, and k means the numbering of QoS index;

(4) whether the inspection model parameter restrains, if convergence finishes and the preservation model parameter, if do not restrain, returns to (2) and carries out.

6. the personalized service recommendation method based on enigmatic language justice probability model according to claim 3 is characterized in that: in step 6, utilize the implementation of the enigmatic language justice probability model predictive user index preference of having trained as follows:

(1) user u ^tlog in service recommendation system, the service situation that need to carry out service recommendation e is provided ^t;

(2) the service contextual information that commending system obtains user's personal data and provides, with following formula to this user each QoS index preference under this service situation carry out independent prediction:

$R_k(e^t,u^t)=\underset{0}{\overset{1}{\∫}}{r}_k\frac{P(e^t,u^t,r_k)}{\underset{0}{\overset{1}{\∫}}P(e^t,u^t,r_k)d{r}_k}d{r}_k;$

Wherein:

$P(e^t,u^t,r_k)=\underset{i,j}{\mathrm{\Σ}}P\left(U_i\right)P\left(E_j\right)P\left(u^t\right|U_i)P\left(e^t\right|E_j)P\left(r_k\right|U_i,E_j)$

$P\left(r_k\right|U_i,E_j)=\frac{1}{\sqrt{2\mathrm{\π}}{\mathrm{\σ}}_{k,U_i,E_j}}\exp [-\frac{{(r_k-{\mathrm{\μ}}_{k,U_i,E_j})}^2}{2{{\mathrm{\σ}}^2}_{k,U_i,E_j}}].$

U wherein ^t, e ^tmean respectively user and the residing service situation thereof of index preference to be predicted, k means the numbering of QoS index, R _k(e ^t, u ^t) expression user u ^tat service situation e ^tunder to the predicted value of k index preference, U, E is the hidden class of representative of consumer and the hidden class of service situation respectively, i means the label of the hidden class of user, j means to serve the numbering of the hidden class of situation,

with

mean respectively the hidden class U of given user context _iwith the hidden class E of service situation _jthe time k index preference the normal distribution average and variance

finally prediction obtains user u ^tat this service situation e ^tunder to the preference weight of each QoS index, be $(R_1(e^t,u^t),...,R_K(e^t,u^t)),$ Be denoted as $({\hat{r}}_1,{\hat{r}}_2,...,{\hat{r}}_K).$

7. the personalized service recommendation method based on enigmatic language justice probability model according to claim 3, it is characterized in that: in step 7, according to the QoS index preference of the user individual doped, the process that candidate service is screened is as follows: user u ^tat service situation e ^tunder obtained a plurality of intimate candidate service, suppose that the performance on each QoS index of each service can have additive method to obtain, be designated as (q ₁, q ₂..., q _k), the PTS of each service adopts

mean, finally according to the PTS of each service, provide the sequence of service, as the foundation of service recommendation.

Images