CN106095887A - Context aware Web service recommendation method based on weighted space-time effect - Google Patents

Abstract

The invention discloses a kind of context aware web service recommendation method based on weighted space-time effect, including, set up weight temporal attenuation model, thus find the user set similar to active user's preference in the case of considering time attenuation；Use location aware similarity mining algorithm, it is thus achieved that with user's set that user is presently in context aware；Set up the relational model of temporal correlation and user preference, thus obtain the web services call record that preference current with user is consistent；Based on user's set that preference is similar in the case of considering time attenuation to active user obtained above, user is presently in the web services call record that user gathers and preference current with user is consistent of context aware and utilizes Bayes theorem to carry out servicing qos value prediction, obtains the service that satisfaction is the highest；The service obtained is carried out outcome evaluation.Realize improving the advantage recommending accuracy rate.

Description

Context aware web service recommendation method based on weighted space-time effect

Technical field

The present invention relates to internet arena, in particular it relates to a kind of context aware Web service based on weighted space-time effect Recommendation method.

Background technology

Web service is a kind of service-oriented application program, and it provides service by the Web agreement of standard, it is therefore an objective to protect The application service of card different platform is capable of interoperability.Common Web service be provided with weather forecast query Web service, The Web service etc. that application program is downloaded is provided.Along with developing rapidly of network technology, a large amount of functionally similar Web services exist Internet is upper to be issued.Therefore, how to screen for user and to recommend the Web service meeting its individual demand to be one and compel to be essential Problem to be solved.Being directed to this, industrial quarters and academia have carried out numerous studies in fact from multiple perspective to Web service recommendation Trample, it is proposed that the various technology of Web service recommendation and system.These technology are roughly divided into three classes: recommendation based on collaborative filtering Method, the recommendation method of QoS perception, and recommendation method based on context aware.

C.Wu, J.Wu, Z.Zheng et al. propose a kind of recommendation method of collaborative filtering, i.e. remember according to the history of user Record or " user-service " rating matrix calculate user's similarity or service similarity is recommended interested for user or marks High Web service.It addition, Xiaohui Cui etc. are considering the factor of time decay during collaborative filtering, adjust the distance and work as The QoS scoring that the front time is different, gives different weights, i.e. distance current time is the nearest, solves in user interest similarity Weight shared by during is the biggest.The shortcoming of this technical scheme is: user exists huge difference to the history scoring of different Web services Different, but this method ignores the scoring of the differentiation Different Effects to Web service recommendation, therefore during Similarity Measure Cause the inefficient defect of proposed algorithm.

Zibin Zheng et al. proposes a kind of web service recommendation method based on QoS perception.The core of the method is Propose a kind of user collaborative mechanism utilizing history Web service QoS information, and be current based on the history QoS data gathered The qos value of user in predicting Web service.Shortcoming: the qos value (such as response time) of Web service is largely dependent upon network bar The situational factoies such as part.Situation residing for two users is the most close, then both are the most similar to the qos value of identical Web service.So And, the method does not accounts for above-mentioned context aware.Therefore, the defect that QoS forecasting accuracy is low is caused.

The proposed algorithm of context aware mainly considers the situational factoies such as community network, space, time, carries out more for user Network service is recommended accurately.Kuang Li, M.Tang, Y.Hu et al. consideration user and the locus situation of network service, Propose a kind of collaborative filtering recommending method based on location aware, it is possible to according to user's location information and Web service Service quality, make comprehensive service recommendation for user.The shortcoming of said method is: the locus residing for user can be along with The change of time and change, but prior art seldom considers the temporal correlation impact on user preference.

In prior art, three representative and relevant with the present invention technical schemes of set are:

1) Yan Hu et al. and Xiaohui Cui et al. carry out different modelings to time decay present in recommending, for User recommends.

2) Zheng et al. proposes according to information such as user's request, interest preference and historical records, and utilizes " user-clothes Business " rating matrix recommends the interested or method of the highest Web service of scoring for user.

3) M.Tang, Y.Hu et al. propose the web service recommendation method of context aware, it is intended to active user recommend with The similar service call record of residing contextual information in service.

The shortcoming of above-mentioned three set solutions is: first, considers time attenuation effect during collaborative filtering, but Do not account for different QoS numerical value and collaborative filtering is also had the impact of different weight, i.e. lack weight temporal attenuation effect Consider；Second, have ignored the contextual information that user is presently in, the qos value (such as response time) of Web service largely depends on Rely in situational factoies such as network conditions.Situation residing for two users is the most close, then both are to the qos value of identical Web service just The most similar.But, the method lacks the impact considering context aware effect.Therefore, low the lacking of QoS forecasting accuracy is caused Fall into；3rd, only using contextual information as " filter ", i.e. utilize the history relevant to active user's preference obtained by screening Call data to recommend, but owing to not accounting for temporal correlation to the impact between user preference, thus have impact on recommendation The accuracy rate of method.

Summary of the invention

It is an object of the invention to, for the problems referred to above, propose a kind of context aware Web based on weighted space-time effect clothes Business recommendation method, to realize improving the advantage recommending accuracy rate.

For achieving the above object, the technical solution used in the present invention is:

A kind of context aware web service recommendation method based on weighted space-time effect, including,

Step 1, set up weight temporal attenuation model, thus find with active user in the situation considering time attenuation User's set that lower preference is similar；

Step 2, employing location aware similarity mining algorithm, it is thus achieved that with user's collection that user is presently in context aware Close；

Step 3, set up the relational model of temporal correlation and user preference, thus obtain what preference current with user was consistent Web services call record；

Step 4, based on the use that preference is similar in the case of considering time attenuation to active user obtained above Family is gathered, and user is presently in the web services call record that user gathers and preference current with user is consistent of context aware Utilize Bayes theorem to carry out servicing qos value prediction, obtain the service that satisfaction is the highest；

Step 5, the service obtaining above-mentioned steps 4 carry out outcome evaluation.

Preferably, described step 1 sets up weight temporal attenuation model particularly as follows:

The function setting up time decay is:

$f(\mathrm{\α},t)=e^{-\mathrm{\α}\left|t_{current}-\mathrm{\Δ}t\right|},$

For different QoS, can obtain it affects factor alpha:

$\mathrm{\α}=\frac{1}{\sqrt{{(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})}^2+{(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})}^2}}$

Wherein,WithRepresent user u respectively_iWith user u_jTo service s_kUser satisfaction,WithRespectively Represent user u_iWith user u_jSatisfaction meansigma methods to all used services of common tune；

In conjunction with Pearson's equation, calculate the similarity size of consideration time attenuation effect, as follows,

$sim(u_i,u_j,t)=\frac{{\Σ}_{s_k\∈W_{u_i,u_j}}(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})f(\mathrm{\α},t)}{\sqrt{{\Σ}_{s_k\∈W_{u_i,u_j}}{(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})}^2}\sqrt{{\Σ}_{s_k\∈W_{u_i,u_j}}{(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})}^2}}$

Wherein,It is user u_iWith user u_jThe set of the network service jointly called, s_kIt isIn any One service,It is user u_iAverage to all-network service scoring,It is user u_jThat marks all-network service is equal Value, t_currentRepresent current time,.

Preferably, in the location aware similarity mining algorithm of described step 2, Euclidean distance is used to calculate two use The similarity of different time present position between family, its formula is as follows:

$Sim(l_{u_i,t},l_{u_j,t})=\sqrt{{\Σ}_{k=1}^N{(l_{i,t,k}-l_{j,t,k})}^2}.$

Wherein,Represent location similarity between two users, l_i,t,kAnd l_j,t,kRepresent user respectively u_iWith user u_jKth dimension in the position of t.

Preferably, described step 3 is set up in the relational model of temporal correlation and user preference, specially set up network away from From the model that affects on user preference, model formation is as follows:

$P_{NDS\left(t\right)}=P_{0}Dis{(l_{u_i,t},l_{s_k})}_{nor}$

Wherein, P₀It is a constant,Represent the network distance between user and service,Being the normalized carried out in an order of magnitude for the quantization affecting preference, result is returned One changes between 0-1；

P_{RC S(t)}And P_{ND S(t)}It is temporal correlation impact that user preference is caused, is assigned to the weight that they are different respectively, Available overall preference impact is as follows,

$P_{S\left(t\right)}=w_1{P}_{RCS\left(t\right)}+w_2{P}_{0}Dis{(l_{u_i,t},l_{s_k})}_{nor}$

Finally, P is utilized_S(t)Obtain the web services call record that preference current with user is consistent, w₁And w₂Represent respectively The weight of QoS attribute.

Preferably, Bayes theorem, it is formulated as:

$P(OS=1|s_i)=\frac{P\left(s_i\right|OS=1)*P(OS=1)}{P\left(s_i\right)}$

Wherein, P (OS=1 | s_i) represent service s_iSatisfaction, P (OS=1) represents the satisfaction rate to all services, P (s_i | OS=1) service be satisfied with is s_iProbability, P (s_i) represent, all of calling, record calls service s_iProbability.

Preferably, the service to obtaining of the described step 5 carry out outcome evaluation particularly as follows:

MAE/RMSE value is used to be predicted the assessment of result,

The concrete formula using mean absolute error:

$MAE=\frac{{\Σ}_{u,s}\left|Q_{u,s}-{\hat{Q}}_{u,s}\right|}{N}$

The formula of root-mean-square error:

$RMSE=\sqrt{\frac{{\Σ}_{u,s}{(Q_{u,s}-{\hat{Q}}_{u,s})}^2}{N}}$

Wherein, Q_u,sRepresent the user u actual value to the service overall qos value of s,Represent user u to the service overall QoS of s The predictive value of value, N represents the total degree of prediction.The value of MAE/RMSE is the least, illustrates that the error of prediction is the least, i.e. commending system is pre- Record the most accurate.

Preferably, described step 1 also includes the step of data set pretreatment before setting up weight temporal attenuation model,

Described data set pretreatment, including data set is divided training set and test set, the pretreatment of data set, each Call calculating and the setting of threshold value q of record qos value.

Preferably, the division proportion of described training set and test set is: 14:1,13；2,12:3,11:4,10:5,9:6 or 8:7。

Preferably, described data set pretreatment uses Gauss standard method, the data of Gauss method standardization QoS characteristic Formula is as follows:

$q_k^{i,j^{\′}}=0.5+(q_k^{i,j}-\stackrel{\‾}{q_k^j})/(2*3{\mathrm{\σ}}_j)$

Wherein,Represent service s_iBy user u_jThe numerical values recited of kth QoS characteristic when calling,It is rightNormalizing Result after change,It is user u_jThe meansigma methods of kth characteristic, σ_jIt is user u_jThe standard deviation of kth characteristic.

Preferably, each the computing formula calling record qos value is:

QoS=w₁*v_RTT+w₂*v_{Data Size}+w₃*v_{R HTTP Message}

Wherein, v_RTT、v_{Data Size}、v_{R HTTP Message}It is in response to the qos value of time, size of data, response message respectively, w₁、w₂、w₃It is the weight of three QoS attributes respectively.

Preferably, w₁、w₂And w₃It is entered as 0.35,0.05 and 0.60 respectively.

Technical scheme has the advantages that

Present invention primarily contemplates the impact of time effect in network service, first, during collaborative filtering, by original Time attenuation effect combines with the impact of different QoS value, carries out new modeling；Secondly, it is considered to context aware residing for user Impact, the situation residing for i.e. two users is the most close, then both are the most similar to the qos value of identical Web service, accordingly for using Family is made and being recommended more accurately；Finally, it is considered to space situation time dependent temporal and spatial correlations sexual factor, it is modeled inclined to user Good impact, makes real-time personalized recommendation for user.

Below by drawings and Examples, technical scheme is described in further detail.

Accompanying drawing explanation

Fig. 1 is that the Weather Forecast Web service of the consideration time effect described in the embodiment of the present invention recommends the scene of method to show It is intended to；

Fig. 2 is the stream of the context aware web service recommendation method based on weighted space-time effect described in the embodiment of the present invention Cheng Tu；

Fig. 3 is the time attenuation effect model schematic described in the embodiment of the present invention；

Fig. 4 is the context aware web service recommendation method module based on weighted space-time effect described in the embodiment of the present invention Schematic diagram；

Fig. 5 is the tool of the context aware web service recommendation method based on weighted space-time effect described in the embodiment of the present invention Body embodiment theory diagram；

Fig. 6 a and MAE and the RMSE result schematic diagram that Fig. 6 b is algorithms of different；

Fig. 7 a and Fig. 7 b is MAE and the RMSE result schematic diagram of various algorithms under different test set and training set ratio；

Fig. 8 a and Fig. 8 b is MAE and the RMSE result schematic diagram under different neighboring user numbers.

Detailed description of the invention

Below in conjunction with accompanying drawing, the preferred embodiments of the present invention are illustrated, it will be appreciated that preferred reality described herein Execute example be merely to illustrate and explain the present invention, be not intended to limit the present invention.

The defect that prior art exists has: first, although prior art take into account the time during collaborative filtering Attenuation effect, but never in view of different QoS numerical value, collaborative filtering is also had the impact of different weight；Second, existing skill Art is seldom in view of the impact of context aware, and the situation residing for i.e. two users is the most close, then both are to identical Web service Qos value is the most similar；3rd, prior art seldom has the consideration temporal correlation impact on user preference, the i.e. position of user Change over time and change, owing to Web service itself and network are closely bound up, this can cause the preference of user to become Change.

The defect existed for prior art, the present invention considers different QoS numerical value during having initially set up collaborative filtering The weight temporal attenuation effect model of impact；Secondly, context aware effects model is established；Again, to temporal correlation User preference impact is modeled；Finally combine Bayesian forecasting method, it is achieved that a kind of based on weighted space-time effect Property network service recommendation mechanisms, solve prior art recommend the problems such as accuracy rate is the highest, recommendation effect is poor.

First, it is considered to the temporal correlation impact on user preference.Accompanying drawing 1 illustrates Kongxiang when considering disclosed by the invention The weather forecast network service of the impact of user preference is recommended the scene graph of method by closing property.Figure comprises a service layer (inner Bread is containing many services, and these services are widely distributed in all over the world), a space layer and a time horizon.

Assume inside this service memory storehouse, to include many online weather forecast service (s₁And s₂Represent American National The weather forecast service of weather bureau, s₃Represent the weather forecast service of China National Meteorological Center, s₅Also the sky of meteorological China is represented Gas reporting services, s₄Represent the weather forecast service of Britain BBC).

Owing to the dependency of accuracy and the region of weather forecast service is very big, naturally enough, user prefer to select away from The weather forecast network service closer from oneself current location.In the scene of accompanying drawing 1, use the when of on January 12nd, 2015 Family u₁In New York, then it can tend to select the weather forecast network service s from New York₁Or s₂.And on January 13rd, 2015 arrives On January 16th, 2015, user u₁Stay in Beijing owing to going on business, then it tends to select from Pekinese's weather forecast network always Service s₃And s₅.Additionally, for service incoherent with region, due to response time, accordingly result, size of data etc. all with net Network position is relevant, and the network between user and service is the nearest, and its invocation performance is the best, even so servicing for user Itself is uncorrelated with region, is also intended to select away from oneself nearer service.In a word, when change over time, user's When position changes, the preference of user also can occur to change accordingly therewith, sets up user's temporal correlation and user accordingly Model is affected between preference.

Secondly, it is considered to the impact on collaborative filtering process of the different qos values.Prior art is generally in the process of collaborative filtering Middle consideration time attenuation effect, change the most over time, the service call record before user is played when recommending Effect also can change, the most remote the calling record and user's similarity may be solved role more of distance current time Little.And different qos values also can produce impact to this process, such as, user A calls service 1, and response time is 0.60 second, And during the similar service 2 of calling function, response time is 20 seconds, for the angle of Consumer's Experience, user's preference to service 1 Larger, if carrying out when solving of Interest Similarity with another one user B, we should amplify, and " user A more likes Service 1 " this is true.

Additionally, it is contemplated that the impact of context aware effect.If so-called context aware effect is that is two current institutes of user The situation (such as network speed, bandwidth etc.) at place equally, owing to the QoS characteristic of Web service is closely related with those situations, then call together During one Web service, the user of context aware can obtain similar QoS characteristic.Accordingly, we can be that user finds context aware User, make for it and more reasonably recommending.

As in figure 2 it is shown, context aware web service recommendation method based on weighted space-time effect, comprise the following steps:

Step 1, set up weight temporal attenuation model:

Traditional collaborative filtering usually utilizes user-commodity rating matrix to find similar user or similar business Product, and then recommend for user.But the user that in practical situations both, As time goes on, distance current time is distant- Commodity scoring record role during Similarity Measure can slowly reduce, and the use that distance current time is closer Family-commodity scoring record during Similarity Measure role comparatively speaking can ratio larger, i.e. user-commodity As time goes on scoring record can role can decay.The present invention sets up time attenuation model, below in conjunction with Fig. 3 illustrates.

T in Fig. 3_startRepresent time started, t_ikRepresent user i and call network service s_kTime, t_jkRepresent user j to adjust Use network service s_kTime, t_currentRepresent current time, Δ t_iIt is t_ikWith current time t_currentTime difference, Δ t_jIt is t_jkWith current time t_currentTime difference.

For different QoS, can obtain it affects factor alpha:

$\mathrm{\α}=\frac{1}{\sqrt{{(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})}^2+{(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})}^2}},$

The function that the present invention sets up time decay is as follows,

$f(\mathrm{\α},t)=e^{-\mathrm{\α}\left|t_{current}-\mathrm{\Δ}t\right|},$

When carrying out the calculating of similarity, user i and user j calls network service s jointly_k, they are about network service s_k Similarity overall similar action should be had the decay of a time.To this end, we take interlude Δ t=(Δ t_i+Δ t_j)/2。

In conjunction with traditional similarity calculating method (Pearson's equation), the similarity of consideration time attenuation effect can be calculated Size, as follows,

$sim(u_i,u_j,t)=\frac{{\Σ}_{s_k\∈W_{u_i,u_j}}(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})f(\mathrm{\α},t)}{\sqrt{{\Σ}_{s_k\∈W_{u_i,u_j}}{(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})}^2}\sqrt{{\Σ}_{s_k\∈W_{u_i,u_j}}{(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})}^2}},$

Wherein,It is user u_iWith user u_jThe set of the network service jointly called, s_kIt isIn any One service,It is user u_iAverage to all-network service scoring.Thus can find with active user in the consideration time User's set that in the case of attenuation, preference is similar.

Step 2, location aware similarity mining algorithm:

Two users, the situation residing when them is the most similar, and they more likely select similar service.By situation phase Excavation like degree, it is thus achieved that with user's set that user is presently in context aware.But, change over time, residing for user Situation can change, therefore the present invention consider over time change situation when changing context aware excavate, calculate not With the similarity of the many tuples of situation between user and user under the time.Present invention employs Euclidean distance and calculate two use The similarity of different time present position between family, its formula is as follows:

$Sim(l_{u_i,t},l_{u_j,t})=\sqrt{{\Σ}_{k=1}^N{(l_{i,t,k}-l_{j,t,k})}^2},$

For a new user, it is possible to use Euclidean distance formula calculates its contextual information and each bunch of heart Distance, that bunch of closest heart place bunch, the user inside it is most like with current user context.

Step 3, temporal correlation and the relational model of user preference:

It is known that As time goes on the position of user can change, this temporal correlation meeting of user Affect the preference of user.Such as 1) scene introduce described in, for the network service relevant to region, user can prefer to away from From the service that oneself is closer, set up the impact on user preference of the region dependency accordingly, as follows

$P_{RCS\left(t\right)}=\{\begin{array}{ccc}1& if& webserviceisrelatedtoregion\\ 0& if& webserviceisnotrelatedtoregion\end{array},$

For service incoherent with region, all have with network site due to response time, accordingly result, size of data etc. Close, user and service between network the nearest, its invocation performance is the best, even so for user service itself with Territory is uncorrelated, is also intended to select away from oneself nearer service.Network distance can be set up user preference affected model, As follows

$P_{NDS\left(t\right)}=P_{0}Dis{(l_{u_i,t},l_{s_k})}_{nor},$

Wherein, P₀Being a constant, in order to the quantization affecting preference is in an order of magnitude, the present invention is by P₀It is set to 1.Represent the network distance between user and service,Also for what preference was affected Quantify the normalized carried out in an order of magnitude, result is normalized between 0-1.

P_{RC S(t)}And P_{ND S(t)}It is all temporal correlation impact that user preference is caused, is assigned to the power that they are different respectively Weight, available overall preference impact is as follows,

$P_{S\left(t\right)}=w_1{P}_{RCS\left(t\right)}+w_2{P}_{0}Dis{(l_{u_i,t},l_{s_k})}_{nor},$

Finally, P is utilized_S(t)The web services call record that preference current with user is consistent can be obtained.

Step 4, predict based on Bayesian QoS:

The service call record bigger with active user's dependency is obtained through above step.On this basis, utilize Bayes theorem carries out servicing the prediction of qos value (for active user).

First provide Bayes theorem, be formulated as:

$P(OS=1|s_i)=\frac{P\left(s_i\right|OS=1)*P(OS=1)}{P\left(s_i\right)},$

Wherein, P (OS=1 | s_i) represent service s_iSatisfaction, P (OS=1) represents the satisfaction rate to all services, P (s_i | OS=1) the inside be satisfied with is s_iProbability.

Illustrate the prediction how utilizing Bayes theorem to carry out overall QoS below according to table 1.

	QoS	OS
			<s<sub>1</sub>,u<sub>1</sub>>	0.85	1
<s<sub>1</sub>,u<sub>2</sub>>	0.75	1
			<s<sub>1</sub>,u<sub>3</sub>>	0.45	0
<s<sub>2</sub>,u<sub>1</sub>>	0.8	1
			<s<sub>2</sub>,u<sub>2</sub>>	0.5	0
<s<sub>2</sub>,u<sub>3</sub>>	0.6	0
			<s<sub>3</sub>,u<sub>1</sub>>	0.75	1
<s<sub>3</sub>,u<sub>2</sub>>	0.55	0

Table 1, Bayes theorem application table,

In Table 1, service s₁With service s₂Called respectively by similar users 3 times, service s₃Be have invoked 2 times by similar users. QoS is the QoS size after every time calling, and OS is to determine whether satisfied according to the threshold value that is satisfied with set, and 0 representative is dissatisfied, 1 generation It is satisfied with.Utilize Bayesian formula, each service size to the qos value of active user measurable is as follows:

$\begin{array}{c}QoS\left(s_1\right)=P(OS=1|s_1)=\frac{P\left(\right(s_1|OS=1\left)\right)*P(OS=1)}{P\left(s_1\right)}\\ =\frac{\frac{1}{2}*1/2}{3/8}=2/3\end{array}$

$\begin{array}{c}QoS\left(s_2\right)=P(OS=1|s_2)=\frac{P\left(\right(s_2|OS=1\left)\right)*P(OS=1)}{P\left(s_2\right)}\\ =\frac{1/4*1/2}{3/8}=1/3\end{array}$

$\begin{array}{c}QoS\left(s_3\right)=P(OS=1|s_3)=\frac{P\left(\right(s_3|OS=1\left)\right)*P(OS=1)}{P\left(s_3\right)}\\ =\frac{\frac{1}{4}*1/2}{2/5}=1/2\end{array}$

Can be seen that by predicting the outcome above, it was predicted that the service that satisfaction is the highest is s₁。

Step 5, predict the outcome assessment:

The appraisal procedure predicting the outcome service recommendation mainly has two kinds.One be by precision (accuracy rate)/ The method of recall (recall rate) is assessed, and this is mainly used in the algorithm once recommending multiple service to user；Another kind is Being assessed by usually said MAE (mean absolute error)/RMSE (root-mean-square error), the value of MAE/RMSE is the least, explanation The error of prediction is the least, i.e. commending system records the most accurate in advance, and this method is mainly used in the algorithm providing QoS predictive value.By It is to be given with the form of QoS in predicting the outcome of the present invention, so using MAE/RMSE value to be predicted the assessment of result.

The formula of mean absolute error

$MAE=\frac{{\Σ}_{u,s}\left|Q_{u,s}-{\hat{Q}}_{u,s}\right|}{N},$

The formula of root-mean-square error

$RMSE=\sqrt{\frac{{\Σ}_{u,s}{(Q_{u,s}-{\hat{Q}}_{u,s})}^2}{N}},$

Wherein, Q_u,sRepresent the user u actual value to the service overall qos value of s,Represent user u to the service overall QoS of s The predictive value of value, N represents the total degree of prediction.The value of MAE/RMSE is the least, illustrates that the error of prediction is the least, i.e. commending system is pre- Record the most accurate.

When the time situation of user changes, repeat process above.

1) pretreatment of data set.Mainly include the training set of data set and the division of test set, the pretreatment of data set, Each is called the record calculating of qos value, the setting of threshold value q.

The training set to be divided into of data set and test set, training set is mainly used to train the result drawing prediction, then Compare with the result in test set, draw forecast error.In the present embodiment, data set is divided training set and test set Ratio is 14:1,13；2,12:3,11:4,10:5,9:6,8:7.

As shown in Figure 4, the data concentrated data under MATLAB software are standardized processing.Data normalization belongs to Prior art, presently, there are many methods planted data prediction.In order to avoid the result of data normalization is very big by some Or minimum data are affected, there is employed herein Gauss standard method.The data of Gauss method standardization QoS characteristic (RTT, Data Size), formula is as follows:

$q_k^{i,j}=0.5+(q_k^{i,j}-\stackrel{\‾}{q_k^j})/(2*3{\mathrm{\σ}}_j),$

Wherein,It is user u_jThe meansigma methods of kth characteristic, σ_jIt is user u_jThe standard deviation of kth characteristic.

Again, calculating the total satisfactory grade QoS after user calls service each time, formula is:

QoS=w₁*v_RTT+w₂*v_{Data Size}+w₃*v_{R HTTP Message},

Wherein, v_RTT、v_{Data Size}、v_{R HTTP Message}It is in response to the qos value of time, size of data, response message respectively, w₁、w₂、w₃It is the weight of three QoS attributes respectively.Here according to the importance size of each characteristic, respectively three weights are composed Value is 0.35,0.05 and 0.60.

Finally, expression user is set and this Web service is called the threshold value whether being satisfied with (according to the result of QoS, selection one The individual value having most discrimination).Record is called for each Web service, is then satisfied (being set to 1) more than or equal to this threshold value, little Then dissatisfied (being set to 0) in this threshold value.

As it is shown in figure 5, data set uses the WS-Dream Dataset of Hong Kong Chinese University's service computing laboratory exploitation. This data set is one of the most authoritative data set of current Web service recommendation, calls the User IP ground in record including Web service The information such as location, Web service address, response time, response message, size of data, data throughout, mortality.Data are concentrated with 150 users, 100 services.Every service is called 100 times by each user respectively, adjusts so about having 1,500,000 services With record.

The present embodiment uses MATLAB (2013a version) to realize.The benefit selecting MATLAB is to wrap in this software Containing the function of many mappings, the visualization for experimental result is provided convenience.The environment that above-mentioned software is run is Lenovo ThinkPad E430c, its CPU are 2.5GHz Intel Core I5, and operating system is the Windows 7 of 64.

The reference algorithm compared with result of implementation of the present invention:

All recommend method: this algorithm utilizes all data of data set directly to carry out the prediction of qos value.

UPCC method: it utilizes the user-item rating matrix of user, finds the user's similar to current user interest History is called record and is predicted.Similarity between user utilizes Pearson correlation coefficients equation to calculate.

IPCC method: its similar items utilizing item-user rating matrix to find the item that user once liked recommends To user, the similarity between item is also to be calculated by Pearson correlation coefficients equation.

CASR method: the service call that the method is gathered by finding the user under similar situation is recorded as active user and does Going out to recommend, it is the most similar that it is based on the assumption that two users call situation residing during service, and they more likely select identical Service.

CASR-UP method: the method considers the user preference determined by customer location, and then recommends for user.

ITRP-WS method: the method considers time attenuation effect recommending when.

CASR-TE method: the method considers temporal correlation and time attenuation effect when recommending, but does not consider difference The impact of qos value.

Result of the comparison is specifically shown in Fig. 6 a, Fig. 6 b, Fig. 7 a, Fig. 7 b, Fig. 8 a and Fig. 8 b.

Wherein in Fig. 6 a and Fig. 6 b, the data post in each data set is followed successively by from left to right: UPCC, IPCC, RBA, CASR、ITRP-WS、CASR-UP、CASR-TE、CASR-TSE。

In Fig. 8 a and Fig. 8 b, the data post in each data set is followed successively by from left to right: UPCC, IPCC, ITRP-WS, CASR-TE、CASR-TSE。

By above experimental result it can be seen that the CASR-TSE algorithm of the present invention has less experimental error, than it Its algorithm to be got well.

Additionally for the replacement of techniques below scheme also among the protection of technical solution of the present invention,

1) there is other method for normalizing and can substitute in the data prediction of the inventive method, as all can carry out normalizing The most alternative data prediction of method changed；

2) the location aware similarity of the inventive method excavates module existence, excavates module such as location aware similarity and exists Some schemes that can replace, the mathematical method of every spacing size solving multi-C vector all can replace this programme to build Mould；

3) the QoS prediction module of the inventive method, can pass through other forecast model such as SVM, decision tree or neutral net Etc. being predicted.Every inventive method technology again completed with upper module by replacement, all be can be considered in the technology of the present invention Among the protection of scheme.

Abbreviation and Key Term explanation:

RS (Recommender System, it is recommended that system): be the information requirement according to user, interest and historical behavior note Information interested for user, product etc. are recommended the Personalized Information Recommendation System of user by records etc..

CARS (Context-aware Recommender System, context aware commending system): be a kind of special Commending system, i.e. on the one hand recommends those product similar to its information requirement or services for user, on the other hand based on situation Information is that user makes personalized recommendation.

Web service (network service): be an application program, it can outwardly be provided one and be called by network Application programming interfaces (API).

QoS (Quality of Service, service quality): when user is by one Web service of network call, can table Reveal the characteristic of some response service quality, as called whether success, response time, transmitted data amount size etc., these characteristics It is referred to as QoS.

In sum, present invention offers following beneficial effect:

1) the inventive method considers the time attenuation effect model of weighting, during improving commending system collaborative filtering Similarity solve accuracy.Every in commending system relate to collaborative filtering method, even for different recommended (as Book, music, film, commodity etc.) may be by this time attenuation effect model weighted, promote phase during collaborative filtering Like the accuracy that solves of degree, and then promote the accuracy rate recommended.

2) the inventive method considers the impact of context aware degree, and the situation residing for i.e. two users is the most close, then both The most similar to the qos value of identical Web service, and establish context aware degree mining model, improve recommendation further accurately Degree.

3) for the recommendation of Web service, the inventive method establishes the relation mould that user preference is affected by temporal correlation Type, this model may apply to Web service and other has in the recommendation of similarity relation model.

4) both the above model and location similarity are excavated and combine by the inventive method, and utilize bayes method to carry out The prediction of QoS, has obtained context aware web service recommendation method based on weighted space-time effect, can promote recommendation further Accuracy.

5) this context aware web service recommendation method based on weighted space-time effect, it is achieved simple, it is not necessary to use other Supplementary module, can be circulated iteration when user occurs migration and variation over time, it is achieved network service prediction and recommendation The decoupling self-correcting of efficiency, improves service recommendation effect.

Finally it is noted that the foregoing is only the preferred embodiments of the present invention, it is not limited to the present invention, Although being described in detail the present invention with reference to previous embodiment, for a person skilled in the art, it still may be used So that the technical scheme described in foregoing embodiments to be modified, or wherein portion of techniques feature is carried out equivalent. All within the spirit and principles in the present invention, any modification, equivalent substitution and improvement etc. made, should be included in the present invention's Within protection domain.

Claims (10)

1. a context aware web service recommendation method based on weighted space-time effect, it is characterised in that include,

Step 1, set up weight temporal attenuation model, thus find with active user in the case of considering time attenuation partially Good similar user's set；

Step 2, employing location aware similarity mining algorithm, it is thus achieved that with user's set that user is presently in context aware；

Step 3, set up the relational model of temporal correlation and user preference, thus obtain the network that preference current with user is consistent Service call record；

Step 4, based on user's collection that preference is similar in the case of considering time attenuation to active user obtained above Closing, user is presently in the web services call record that user gathers and preference current with user is consistent of context aware and utilizes Bayes theorem carries out servicing qos value prediction, obtains the service that satisfaction is the highest；

Step 5, the service obtaining above-mentioned steps 4 carry out outcome evaluation.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 1, its feature exists In, described step 1 is set up weight temporal attenuation model particularly as follows:

The function setting up time decay is:

$f(\mathrm{\α},t)=e^{-\mathrm{\α}|t_{current}-\mathrm{\Δ}t|},$

For different QoS, can obtain it affects factor alpha:

$\mathrm{\α}=\frac{1}{\sqrt{{(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})}^2+{(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})}^2}}$

Wherein,WithRepresent user u respectively_iWith user u_jTo service s_kUser satisfaction,WithGeneration respectively Table user u_iWith user u_jSatisfaction meansigma methods to all used services of common tune；In conjunction with Pearson's equation, when calculating consideration Between the similarity size of attenuation effect, as follows,

$sim(u_i,u_j,t)=\frac{{\Σ}_{s_k\∈W_{u_i,u_j}}(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})f(\mathrm{\α},t)}{\sqrt{{\Σ}_{s_k\∈W_{u_i,u_j}}{(r_{u_i,s_k}-{\stackrel{\‾}{r}}_{u_i})}^2}\sqrt{{\Σ}_{s_k\∈W_{u_i,u_j}}{(r_{u_j,s_k}-{\stackrel{\‾}{r}}_{u_j})}^2}}$

Wherein,It is user u_iWith user u_jThe set of the network service jointly called, s_kIt isIn any one Service,It is user u_iAverage to all-network service scoring,It is user u_jAverage to all-network service scoring, t_currentRepresent current time.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 2, its feature exists In, in the location aware similarity mining algorithm of described step 2, when using Euclidean distance to calculate different between two users Between the similarity of present position, its formula is as follows:

$Sim(l_{u_i,t},l_{u_j,t})=\sqrt{{\Σ}_{k=1}^N{(l_{i,t,k}-l_{j,t,k})}^2}$

Wherein,Represent location similarity between two users, l_i,t,kAnd l_j,t,kRepresent user u respectively_iWith User u_jKth dimension in the position of t.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 3, its feature exists In, described step 3 is set up in the relational model of temporal correlation and user preference, specially sets up network distance to user preference Affect model, model formation is as follows:

$P_{NDS\left(t\right)}=P_{0}Dis{(l_{u_i,t},l_{s_k})}_{nor}$

Wherein, P₀It is a constant,Represent the network distance between user and service, Being the normalized carried out in an order of magnitude for the quantization affecting preference, result normalizes between 0-1；

P_RCS(t)And P_NDS(t)It is temporal correlation impact that user preference is caused, is assigned to the weight that they are different respectively, can obtain It is as follows to overall preference impact,

$P_{S\left(t\right)}=w_1{P}_{RCS\left(t\right)}+w_2{P}_{0}Dis{(l_{u_i,t},l_{s_k})}_{nor}$

Finally, P is utilized_s(t)Obtain the web services call record that preference current with user is consistent, w₁And w₂Represent QoS respectively to belong to The weight of property.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 4, its feature exists In,

Bayes theorem, is formulated as:

$P(OS=1|s_i)=\frac{P\left(s_i\right|OS=1)*P(OS=1)}{P\left(s_i\right)}$

Wherein, P (OS=1 | s_i) represent service s_iSatisfaction, P (OS=1) represents the satisfaction rate to all services, P (s_i|OS =1) service being satisfied with is s_iProbability, P (s_i) represent, all of calling, record calls service s_iProbability.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 5, its feature exists Carry out outcome evaluation in, the described step 5 service to obtaining particularly as follows:

MAE/RMSE value is used to be predicted the assessment of result,

The concrete formula using mean absolute error:

$MAE=\frac{{\mathrm{\Σ}}_{u,s}|Q_{u,s}-{\hat{Q}}_{u,s}|}{N}$

The formula of root-mean-square error:

$RMSE=\sqrt{\frac{{\mathrm{\Σ}}_{u,s}{(Q_{u,s}-{\hat{Q}}_{u,s})}^2}{N}}$

Wherein, Q_u,sRepresent the user u actual value to the service overall qos value of s,Represent user u to the service overall qos value of s Predictive value, N represents the total degree of prediction, and the value of MAE/RMSE is the least, illustrates that the error of prediction is the least, i.e. commending system records in advance The most accurate.

7. according to the arbitrary described context aware web service recommendation method based on weighted space-time effect of claim 1 to 6, its Being characterised by, described step 1 also includes the step of data set pretreatment before setting up weight temporal attenuation model,

Described data set pretreatment, including data set is divided training set and test set, the pretreatment of data set, each call The calculating of record qos value and the setting of threshold value q.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 7, its feature exists In, the division proportion of described training set and test set is: 14:1,13；2,12:3,11:4,10:5,9:6 or 8:7.

Context aware web service recommendation method based on weighted space-time effect the most according to claim 8, its feature exists In, described data set pretreatment uses Gauss standard method, and the data formula of Gauss method standardization QoS characteristic is as follows:

$q_k^{i,j^{\′}}=0.5+(q_k^{i,j}-\stackrel{\‾}{q_k^j})/(2*3{\mathrm{\σ}}_j)$

Wherein,Represent service s_iBy user u_jThe numerical values recited of kth QoS characteristic when calling,It is rightAfter normalization Result,It is user u_jThe meansigma methods of kth characteristic, σ_jIt is user u_jThe standard deviation of kth characteristic.

10., according to the context aware web service recommendation method based on weighted space-time effect described in claim 98, its feature exists In, each the computing formula calling record qos value is:

QoS=w₁*v_RTT+w₂*v_{Data Size}+w₃*v_{R HTTP Message}

Wherein, v_RTT、v_{Data Size}、v_{R HTTP Message}It is in response to the qos value of time, size of data, response message, w respectively₁、w₂、 w₃It is the weight of three QoS attributes respectively；

w₁、w₂And w₃It is entered as 0.35,0.05 and 0.60 respectively.