Improving the Performance of Web Service Recommender Uing Semantic Similarity Juan Manuel Adán-Coello, Carlo Miguel Tobar, Yang Yuming Faculdade de Engenharia de Computação, Pontifícia Univeridade Católica de Campina (PUC-Campina) Campina, SP, Brail {juan,tobar}@puc-campina.edu.br; lyonbrcn@gmail.com Abtract: Thi paper addree iue related to recommending Semantic Web Service (SWS) uing collaborative filtering (CF). The focu i on reducing the problem ariing from data parity, one of the main difficultie for CF algorithm. Two CF algorithm are preented and dicued: a memory-baed algorithm, uing the k-nn method, and a model-baed algorithm, uing the k-mean method. In both algorithm, imilarity between uer i computed uing the Pearon Correlation Coefficient (PCC). One of the limitation of uing the PCC in thi context i that in thoe intance where uer have not rated item in common it i not poible to compute their imilarity. In addition, when the number of common item that were rated i low, the reliability of the computed imilarity degree may alo be low. To overcome thee limitation, the preented algorithm compute the imilarity between two uer taking into account ervice that both uer acceed and alo emantically imilar ervice. Likewie, to predict the rating for a not yet acceed target ervice, the algorithm conider the rating that neighbor uer aigned to the target ervice, a i normally the cae, while alo conidering the rating aigned to ervice that are emantically imilar to the target ervice. The experiment decribed in the paper how that thi approach ha a ignificantly poitive impact on prediction accuracy, particularly when the uer-item matrix i pare. Keyword: Collaborative filtering, Recommender ytem, Semantic imilarity, Semantic Web Service, Spare data. 1. INTRODUCTION Service-Oriented Computing (SOC) i a new computing paradigm that ue ervice a building block to accelerate the development of ditributed application in heterogeneou computer environment. SOC promie a world of cooperating ervice where application component are combined with little effort into a network of looely coupled ervice for creating flexible and dynamic buine procee that can pread over many organization and computing platform [1] Among the key challenge for the effective ue of Web ervice i the dicovery of ervice that meet the functional and non-functional requirement of it uer and that take into account their preference [2]. In Web ervice dicovery ytem, three entitie can typically be ditinguihed: the ervice requeter (a uer or a program), the ervice provider and the ervice regitry. Entitie eeking ervice make ervice requet to the regitry. In the regitry, the decription of the ervice requeted i compared with the decription of ervice advertied by ervice provider, uing a matching algorithm, to identify whether there are ervice that meet the requet. If the matching i ucceful, the regitry provide the decription of identified ervice intance to the requeter, including the neceary detail for their invocation. Architecture for ervice dicovery, uually baed on the WSDL pecification [3], have eriou limitation ariing from the ervice decription technology and matching algorithm ued. Thee limitation are due, in part, to the ue of informal decription of ervice functionality and capability, written in natural language, uually lacking a common vocabulary for the ervice requeter and provider. Semantic Web Service (SWS) and Linked Service are recent approache that try to overcome thee limitation by combining Web ervice technology with element of the Semantic Web [4][5]. In SWS dicovery architecture, advertied ervice are decribed uing ervice annotation ontologie in addition to WSDL parameter and operation name. Thee ontologie define a emantic model for the decription of a Web ervice from everal perpective, including functionality, execution flow and invocation detail. They define a et of attribute for ervice capability decription, the mot common being the o-called IOPE (Input, Output, Precondition and Effect). Service annotation, in accordance with a ervice annotation ontology, ue concept contained in domain ontologie intead of nontandardized word, which are more commonly ued in conventional non-emantic approache. Domain ontologie decribe the terminology and the relationhip between term of a pecific domain uing an ontology language uch a OWL or RDFS [6][7]. Each ontology language ha it own unique expreive power, but all can model, at the minimum, hierarchie of concept and role of concept, uch a propertie, attribute and relationhip. When performing a earch, the characteritic of the deired ervice, uch a input and output, are pecified by term that repreent ontology concept. Matchmaking algorithm baed on logical inference can then eek matche for the requet parameter, taking into account the parameter of the available ervice. For each match found, a value that characterize the matching degree (imilarity) i computed. Finally, the identified ervice are returned to the requeter in decending order of matching degree. Search algorithm for emantic Web ervice preent good reult when the uer i able to adequately decribe the deired ervice. However, thi i not alway the cae, and a requet for a ervice cannot correpond fully to the intention of the requeter. For example, there may be a publihed ervice that partially matche the requet and accomplihe the intention of the requeter, or the oppoite cenario could alo conceivably occur [8]. A the number of available ervice on the Web increae, thi problem woren. Currently, a pointed out in [9], one of the mot challenging iue in Web ervice proviion i not the matchmaking proce but the election of good ervice for a target uer. 80
In addition, a the number of available Web ervice grow, there may be a lot of intereting available ervice that uer are not aware of, and that they therefore will not take the initiative to requet. Additionally, in the context of mobile and ubiquitou computing, it i unreaonable to aume that a uer i contantly earching for intereting ervice available at the uer locale. In thi context, it i deirable to have a recommender ytem capable of identifying and of proactively recommending potentially intereting ervice to the uer in the right ituation. A web ervice recommender can alo be very valuable to proactively deal with failure and to recover to ervice workflow that have partially failed and in dynamic compoition cenario, provided the ervice and the recommender can deal with emantic markup [10] [11]. The recommendation problem can be reduced to an iue of etimating rating for item that have not been ued before by a uer; item with higher etimated rating are a a conequence recommended to the uer. Recommender are uually implemented uing filtering algorithm claified into three main categorie, depending on how the recommendation are performed: (1) Content-Baed algorithm (CB) filter and recommend item that are imilar to other the uer ha acceed in the pat; (2) Collaborative Filtering (CF) algorithm filter and recommend item baed on the preference of other uer with imilar tate and preference; knowledge-baed (KB) recommender ue knowledge about uer and item to generate a recommendation. It i alo frequent to find hybrid ytem that combine method taken from two or more of the previou categorie of recommender [12] [13]. Content-baed recommender have their root in the information retrieval field and were uccefully implemented in domain where the item to be recommended are decribed through textual information. Thee ytem are, however, limited by the feature that are explicitly aociated with the item. They are alo limited to recommended item that are imilar to thoe already rated by the uer (over pecialization). A particularly difficult tak for thi type of algorithm i to deal with new uer, becaue new uer have to rate a ufficient number of item before the ytem can undertand their preference and tart making ueful recommendation. CF algorithm do not have ome of the abovementioned hortcoming of content baed algorithm. Since they employ the uer' rating, they can deal with any kind of content and recommend any type of item, even item that are diimilar to thoe acceed in the pat. However CF ytem have their own challenge, including coping with pare data and caling with increaing number of uer and item. Several tructural difficultie related to pare data may be encountered, including the cold tart problem, the reduced coverage problem and the neighbor tranitivity problem. The cold tart problem occur when new uer or item are inerted into the ytem. New item cannot be recommended until they are rated by ome uer, and, in turn, new uer are unlikely to receive good recommendation becaue they lack a rating hitory. The reduced coverage problem occur when the number of rating i very mall compared with the number of item in the rating databae. In thi ituation, the ytem may be unable to generate recommendation for uch uer. The neighbor tranitivity problem occur when uer with imilar tate do not have rated item in common and thu cannot be identified a imilar. Knowledge-baed recommender ytem avoid ome of the drawback of content and CF ytem ince their recommendation do not depend on a bae of uer rating. Their main drawback i the well known knowledge acquiition bottleneck. Algorithm for CF, the primary focu of thi paper, can be further claified into two main categorie: memory-baed and model-baed. Memory-baed algorithm contruct a neighborhood of uer who have imilar rating to the target uer uing directly the available data. In thi circumtance, the rating of neighbor are ued to predict how a target uer will rate an item he ha not yet acceed. Model-baed technique employ available rating data to learn a model to make prediction, uually uing data mining or a machine learning algorithm. Then the model i ued to make prediction for target item, intead of uing raw rating data, a i done with memory-baed algorithm. When comparing memory-baed and model-baed CF algorithm it i uually accepted that memory-baed algorithm are eay to implement and have higher prediction accuracy, particularly for dene dataet. Model-baed algorithm are, in turn, more calable and le vulnerable to profile injection attack [12]. In the recent pat, recommender ytem have been built for recommending different type of item in divere domain, including CD, Web page, book, new, movie and coure. However, reearch on Web ervice recommendation i in it preliminary tage and uually focue on predicting ervice QoS (Quality of Service) parameter [14], which i a very limited way of capturing uer interet [15]. In thi paper, we preent algorithm for contructing Web ervice recommender ytem aimed at reducing the problem ariing from pare data. The propoed approach combine CF algorithm with logical inference to determine the emantic imilarity between ervice, and between uer. The rationale behind thi approach i that if two uer have not rated a common et of ervice but have rated imilar ervice, thee rating can till be an indication of uer imilarity and therefore contribute to reduce the effect of data parene. The remainder of the paper i organized a follow: ection 2 preent memory and model-baed CF algorithm for Semantic Web ervice recommendation; ection 3 dicue the experimental et up ued to evaluate the algorithm and the reult that were obtained; ection 4 preent related work; and, finally, ection 5 conclude the paper by pointing out our main reult and direction for future work. 2. CF ALGORITHMS FOR SEMANTIC WEB SERVICE RECOMMENDATION In thi ection, variation of two recommender algorithm that exploit emantic imilaritie among web ervice are preented. Their performance will be compared in Section 3. Intance of uer feedback 1 are tored in a uer-item matrix, repreented a a et T U S F, where U 1 In thi paper we ue the term feedback, core and rating a ynonym. 81
= {u 1, u 2,, u m } i the et off all uer, S = { 1, 2,, n } i the et of all rated ervice, F = {f 1, f 2,, f m } i the et of intance of feedback related to ervice in S and collected from the uer in U. Each f u F i an n- dimenional vector over the pace of all intance of uer feedback, i.e, f u =( f 1, f 2,, f n ) where f j [0..1] i the feedback given by uer u to ervice j. If a ervice wa not rated it feedback i repreented a φ (null). Although the collaborative filtering algorithm decribed in thi ection are independent of the notation ued to decribe ervice emantic, when they allow for the meaurement of the level of emantic imilarity among two ervice, a prototype for ervice decribed uing OWL-S wa implemented for the validation of the algorithm. OWL-S i an upper ontology that pecifie that a ervice can be decribed by at mot one ervice model, and a grounding mut be aociated with exactly one ervice [16]. OWL-S i a W3C recommendation baed on the W3C tandard OWL, an ontology language for the Semantic Web with formally defined meaning [6]. Computing Service Similarity In our prototype implementation, the degree of imilarity between OWL-S ervice i computed uing a hybrid emantic ervice matching algorithm decribed in [17] that take advantage of both logic-baed reaoning and IR technique. If R repreent a requet for a ervice and S a ervice regitered in the ervice databae, the emantic matching algorithm compute the following matching degree: Exact match (S exactly matche R) - The I/O (Input/Output) ignature of S perfectly matche requet R with repect to the logic-baed equivalence of their formal emantic. Plug-in match (S plug into R) - All input parameter concept of S match more pecific one in R. In addition, S i expected to return more pecific output data. Subumed match (R ubume S) - Thi matching degree i weaker than plug-in matching. The output of S i more pecific than requeted by R a before, but the contraint of immediate output concept ubumption i relaxed to arbitrary output concept ubumption. Subumed-by match (R i ubumed by S) - The output of S i lightly more general than requeted (direct parent output concept). Nearet-neighbor match (S i the nearet neighbor of R) - It i checked if the degree of text imilarity, SynSim(S,R), between the input and output concept of S and R i greater than or equal to a defined yntactic imilarity threhold α. Thi degree i computed a the averaged yntactic imilarity of the erialized input and output concept of S and R, according to a given imilarity metric. A et of concept i erialized by mean of their expanion through the ontology implemented and by the conjunctive concatenation of the reult into one untructured text document, including only logical operator and primitive component of the baic vocabulary that i preent in the ontological terminology. In the cae of vectorpace-baed text imilarity meaurement, thee document are repreented a weighted keyword vector baed on a term-weighting cheme. Fail (S doe not match with R) - None of the above matching degree wa obtained. Memory-baed Feedback Prediction with K-NN Thi recommendation algorithm i baed on the contruction of neighborhood of imilar uer. The neighbor rating can then be ued to make prediction for unrated item. A neighborhood i contructed comparing the imilarity of each pair of exiting uer uing the Pearon Correlation Coefficient (PCC). Two variant of the algorithm were implemented. In the firt, named PCC, the imilarity between two uer u and v, im(v), i computed a hown in Eq. (1), where S uv = { f φ and f v, φ } i the et of ervice that both uer, u and v, have rated, f [0..1] i the feedback given by uer u to ervice and f u and f v are the average of the intance of feedback given by uer u and v, repectively. im(v)= S uv S uv ( f ( f f u)( fv, f v) (1) f u) 2 Su v ( f v, f v) In Eq. (1), if uer u and v have not rated item in common it i not poible to compute their imilarity. Alo, if the number of common item that were rated i very low, the computed imilarity may be unreliable. In the econd variant of the algorithm, named PCC- SS (PCC with imilar ervice), it i not required that uer u and v rate the ame ervice to compute their imilarity a it take into conideration the rating of imilar ervice. The imilarity between ervice i computed uing the emantic matching algorithm preented in the previou ubection. PCC-SS compute the imilarity between two uer, u and v, uing Eq. (2). In that equation, t i the ervice rated by v that i mot imilar to (rated by u), repecting a minimum threhold of imilarity δ. When both uer have rated the ame ervice, and t repreent the ame ervice (the imilarity between and t i 1). im(v)= u S ( f f u)( fv, t f v), (2) Su t Sv ( f f ) 2 u t Sv ( f v, t 2 f v) The imilarity between two uer, im(v), computed uing Eq. (1) or Eq. (2), range from 1 to 1. A value of 1 implie a line that decribe the relationhip between feedback f and f v, given from uer u and v, repectively, for ervice (or a imilar ervice), with all data point (intance of feedback) lying on the line where f v, increae a f increae. A value of 1 implie that all data point lie on the line where f v, decreae a f increae. A value of 0 implie that there i no linear correlation between the variou intance of feedback. In our implementation only im(v) value higher than 0 were conidered relevant. The feedback a uer u would give to a ervice that he ha not yet rated can be etimated uing the rating that neighbor uer aigned to that ervice. Having a neighborhood V, the feedback uer u would give to ervice, f, can be predicted uing two variant of the weighted average of all neighbor rating, a hown in Eq. (3) and Eq. (4). 2 82
im( v)( fv, f v) V f = f u + v (3) im( v) v V For the reult f in Eq. (3), hereafter named WAAR (Weighted Average of All Rating), the neighborhood V i formed by the k mot imilar uer to u that rated ervice. im( v)( fv, t f v) v V fu = f u + (4), im( v) v V For the reult f in Eq. (4), hereafter named WAAR- SS (Weighted Average of All Rating with Service Similarity), the neighborhood V i formed by the k mot imilar uer to u that rated ervice or a ervice t that i emantically imilar to. If V or V i empty, repectively in Eq. (3) or (4), f i made equal to f u. Model-baed Feedback Prediction with K-mean Memory-baed filtering algorithm tend to be more accurate than model-baed algorithm, but the latter are more calable and le vulnerable to profile injection attack [18]. Conidering that the number of available ervice in the Web i continuouly increaing, and that in the context of Web-baed open collaborative recommender the likelihood of attack i not negligible, model-baed recommender algorithm can be good alternative to memory-baed algorithm, provided that their accuracy i acceptable We decribe in thi ection a model-baed CF algorithm for emantic Web ervice that ue the k-mean clutering method and the concept of emantic ervice imilarity. The k-mean method i ued to partition a et of point or obervation into cluter. If we conider that f u F define the profile of uer where f u i the vector of intance of feedback given by uer u for the available ervice, the k-mean algorithm can be ued to cluter uer with imilar profile. Once the cluter are defined, their centroid can be interpreted a aggregated profile of the uer in the cluter a done in [19]. The clutering algorithm work a follow. Initially k point (f vector) are randomly choen a the initial cluter centroid, after which an aignment tep and an update tep are repeated until the algorithm converge. In the aignment tep, each point i aigned to the cluter with the cloet centroid. In the update tep, cluter centroid are updated to the mean of the point aigned to the cluter. The algorithm converge when the centroid no longer change. In the aignment tep, the ditance between a point and a cluter centroid i computed uing the PCC or the PCC-SS (Eq. 1 and Eq. 2, repectively). Following the aignment tep, the update tep compute a new centroid f c =( f c,1, f c,2,, f c,n ) for each cluter c. The new centroid vector i the mean of the uer profile aigned to cluter c. That i, f c,i, for i = 1 to n, i computed by Eq. (5). 1 = fu i (5) c f c, i, u C When applying Eq. (5), if ome f i i equal to φ (meaning that uer u ha not rated ervice i,), the average core of the item rated by u i intead ued. When the algorithm converge, each cluter centroid i een a an aggregation of the uer profile in their repective cluter. Uer intance of feedback for unrated ervice are then etimated uing Eq. (3) or Eq. (4), taking into conideration the neighborhood formed by the k cluter (repreented by their centroid) mot imilar to the target uer profile (repreented by hi feedback vector). 3. EXPERIMENTAL EVALUATION The purpoe of thi ection i to compare the performance of the algorithm preented on ection 2. The lack of public rating dataet i a major difficulty when validating recommender ytem for Web ervice. To circumvent thi difficulty, reearcher uually adapt popular dataet contructed to recommend other type of item. For example, [20] ue the Movielen 2 dataet and conider that a movie in the dataet repreent a Web ervice. The evaluation of the algorithm we propoe, add an additional level of difficulty becaue we need a dataet of uer rating for emantic Web ervice. In thi context, an alternative i to yntheize a dataet that matche the propertie of the target domain and tak [21]. Following thi approach we created a ynthetic ueritem matrix that can be ued to provide ome inight into the behavior of the implemented algorithm and erve a a proof of concept. We ued ervice from the OWL-S Service Retrieval Tet Collection - OWLS-TC 3, verion 2.2, a collection of 1004 Web ervice from everal domain, pecified according to the OWL-S ontology. In the experiment, two group with 50 uer each were defined. Each uer rated 56 ervice from the following four categorie: car, camera, hotel and urf. Service rating were et according to a bae feedback defined for each pair (uer_group, ervice_category). Each feedback wa added to a value that varie from -1 to 1 according to the normal ditribution. The main objective of the experiment wa to analyze the behavior of the propoed algorithm conidering dene and pare data cenario. Thee cenario were imulated by progreively hiding a number of ervice rating from the algorithm: the 56 ervice rating for each uer were progreively reduced in tep of 10 until only 6 rating were available for each uer. After each removal tep, the value of the removed core were etimated uing the algorithm previouly dicued, with and without taking into conideration imilar ervice, following which the average error of the prediction wa computed. The experiment for each removal cenario were repeated 10 time and the reult averaged. The time needed to compute the imilaritie between ervice wa not taken into conideration becaue the computation were performed before running the experiment. The prediction performance of the algorithm wa meaured uing the Mean Abolute Error (MAE) and the Normalized Mean Abolute Error (NMAE), defined by Eq. (6) and Eq. (7), repectively. 2 http://www.grouplen.org/node/73 3 http://project.emwebcentral.org/project/owl-tc 83
MAE = p N f MAE (7) NMAE = f N In Eq. (6), p denote the predicted feedback that uer u will give to ervice, f denote the actual (hidden) feedback that uer u gave to ervice, and N i the number of predicted intance of feedback. Lower value for MAE and NMAE indicate better prediction quality. A MAE or NMAE equal to zero correpond to an ideal cenario, where all prediction are equal to the actual intance of feedback. Evaluating the K-NN Memory-baed Feedback Prediction Algorithm In the experiment decribed in thi ection two ervice are conidered imilar if their matching degree i Exact, Plug-in, Subume, Subumed-by or Nearet-neighbor with a threhold α of 0.8. Two imple etimation cheme, the item-mean and the uer-mean algorithm, were alo implemented to be ued a baeline. The item-mean (IMEAN) algorithm etimate the core for an item (a ervice) a the mean of the core the target item received from all uer that rated it. The uer-mean (UMEAN) algorithm etimate the core for an item a the mean of the core the target uer gave to the item he rated. When applying Eq. (3) or Eq. (4) (WAAR and WAAR-SS), the neighborhood ued to etimate a core i defined by uer with a degree of imilarity to the target uer that i greater than or equal to 0.8, a computed by Eq. (1) or Eq. (2) (PCC and PCC-SS). When etting thi imilarity threhold, we have to conider that if it i too low uer with low imilarity can be conidered neighbor, negatively affecting the accuracy of the algorithm. On the other hand, if the threhold i very high it i poible that no neighbor will be found, making it impoible to predict feedback from the target uer-ervice pair. A can be oberved in Figure 1, the prediction error when uing the PCC and WAAR (without uing ervice imilarity) i ignificantly lower than when the IMEAN and UMEAN algorithm are ued. In other word, NMAE 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 (6) IMEAN UMEAN K-NN without ervice imilarity 0.00 Number of rating removed Figure 1. Prediction accuracy of IMEAN, UMEAN and k-nn without ervice imilarity NMAE 0.45 0.40 0.35 0.30 0.25 0.20 0.15 0.10 0.05 Without ervice imilarity WAAR w ith ervice imilarity 0.00 Number of rating removed PCC with ervice imilarity PCC and WAAR w ith ervice imilarity Figure 2. Impact of ervice imilarity on the accuracy of the KNN-baed prediction algorithm conidering a neighborhood of imilar uer to predict uer feedback i better than uing raw uer or item average. Figure 2 how that conidering ervice imilarity increae the prediction performance to an even greater extent. Thi happen when ervice imilarity i ued only to compute the PCC-SS (Eq. (2)) for the purpoe of finding a neighborhood, or to etimate core with WAAR-SS (Eq. (4)). Uing ervice imilarity both to compute the PCC-SS and the WAAR-SS produce even more accurate prediction. Thee reult can be explained a follow. When the PCC i computed without taking into conideration ervice imilarity, everal imilar uer are not identified becaue the PCC equation correlate only uer that rated a common et of ervice. When ervice imilarity i taken into account, uer who rated imilar ervice are alo taken into conideration, increaing the neighborhood and, a a conequence, the accuracy of the algorithm. In addition, uing ervice imilarity to predict a rating (WAAR-SS) contribute to increae the accuracy becaue it allow more core to be conidered when calculating the prediction. Thi happen becaue intead of only conidering ervice core that the target uer and their imilar uer rated, core for imilar ervice are alo included. Figure 2 alo how that the effect of conidering ervice imilarity are not ignificant when a mall amount of core i removed, but are more dramatic when the amount of removed core increae, that i, when the uer-item matrix become parer. A hown in figure 2, when 50 out of 56 core are removed, the NMAE i equal to 0.23 if ervice imilarity i conidered in both the PCC and WAAR, while when it i not conidered in any of the method it rie to 0.41, an increae of 78%. Evaluating The K-mean Model-baed Feedback Prediction Algorithm Uing the ame cenario from the previou ection, experiment were conducted to evaluate the performance of the prediction approach baed on k-mean. One of the important parameter for thi algorithm i the number of cluter, k. If k i too mall uer profile with little imilarity are clutered together, reducing the accuracy of the algorithm; on the other hand, if k i too high the calability of the algorithm (one of it main expected advantage over the k-nn baed algorithm) can be negatively affected. In the experiment preented in thi ection k wa et to 8, a value choen after ome 84
preliminary tet demontrated that it i a good choice for the data et ued. The neighborhood ued to predict a feedback to a target uer i formed by the cluter centroid that have a degree of imilarity to that uer (computed uing the PCC and PCC-SS) greater than or equal to 0.8. A can be oberved in figure 3, the k-mean prediction algorithm without ervice imilarity ha a prediction error ignificantly lower than that which i obtained when applying the IMEAN and UMEAN algorithm, except when the number of available core i very low (when 50 out of 56 are removed). Under uch circumtance, the mall number of available uer profile prevent the contruction of repreentative uer group, everely affecting the prediction accuracy of the algorithm. Under uch pare data condition, the ue of ervice imilarity account for an appreciable increae in accuracy. A already verified for the k-nn algorithm, the bet reult are oberved when ervice imilarity information i ued for computing both the PCC and the WAAR. Thee reult can be explained in the ame manner a done for the k-nn algorithm: when running the algorithm without ervice imilarity information, everal imilar uer are not identified a uch and are not clutered together, becaue only uer that rated the ame et of ervice can be conidered imilar; when ervice imilarity i taken into account, it i alo poible to identify imilar uer among thoe uer that rated imilar ervice. In addition, when computing the WAAR, the ue of ervice imilarity information contribute to increae the accuracy becaue it allow for the conideration of more core to calculate a prediction. Figure 3 how that when 50 out of 56 core are removed, characterizing a ituation of carcity of evaluation, uing ervice imilarity for computing the PCC and WAAR account for a NMAE of 0.32, while when thi information i not ued the NMAE rie to 0.89, an increae of 178%. Comparing the K-NN and the K-mean Prediction Algorithm The literature ay that memory-baed prediction algorithm, like thoe baed on the k-nn, often have greater accuracy than model-baed algorithm, uch a thoe baed on the k-mean, but model-baed algorithm are more calable becaue they require le memory and are fater. Figure 4 confirm the firt claue of the NMAE 1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 UMEAN k-nn with ervice imilarity k-mean w ith ervice imilarity 0.00 Number of rating removed k-nn w ithout ervice imilarity k-mean without ervice imilarity Figure 4. Comparing the prediction performance of k-nn and k-mean algorithm previou entence. However, it i worth noting that the k- mean algorithm with ervice imilarity i more accurate than the k-nn one without ervice imilarity. The lower accuracy of the k-mean algorithm with repect to k-nn can be explained by the fact that the k- mean method ue cluter centroid and not the profile of imilar uer to predict the core. Profile are grouped into cluter baed on the imilarity of each profile to a cluter centroid; thu a poorly choen centroid directly influence the quality of the cluter. In the implementation decribed, the initial eight centroid were choen randomly among the available profile. The particularly bad reult for the k-mean algorithm when many core are removed and imilar ervice are not conidered can be explained by the difficulty in finding imilar uer to group together when data i pare. Figure 5 how the time required by the algorithm to predict the removed core when uing a notebook with an Intel Core Duo 1.66 GHz proceor and 2 GB of RAM. Regarding the k-mean algorithm, the required time for core prediction with already created cluter i hown. Under thee condition, the run time i lower for the k-mean algorithm, particularly when the uer-item matrix i dene. Thi reult wa expected becaue a high number of profile are conidered in the computation of the PCC and the WAAR when uing the k-nn method, while only a mall number of cluter centroid are ued when applying the k-mean method. NMAE IMEAN UMEAN Without ervice imilarity PCC w ith ervice imilarity WAAR w ith ervice imilarity PCC and WAAR w ith ervice imilarity 1.00 0.90 0.80 0.70 0.60 0.50 0.40 0.30 0.20 0.10 0.00 Number of rating removed Time (m) K-NN without ervice imilarity K-mean without ervice imilarity K-NN with ervice imilarity K-mean with ervice imilarity 40 35 30 25 20 15 10 5 0 Number of rating removed Figure 3. Impact of ervice imilarity on the prediction performance of the k-mean algorithm Figure 5. Run-time of the k-nn and k-mean algorithm for core prediction 85
Time (m) 1400 1200 1000 800 600 400 200 k-mean w ithout ervice imilarity 0 Number of rating removed The run-time for the k-nn algorithm tend to increae harply a the amount of uer and ervice increae. Under uch condition, the lower accuracy of the k-mean algorithm can be acceptable provided that it run time i atifactory. Figure 5 and 6 can be ued to demontrate thi point. Figure 5 how that the run time needed to predict the core i lower for the k-mean algorithm when the uer profile are already grouped. However, a can be een in Figure 6, the time needed by the k-mean algorithm to cluter the uer i very high when compared to the time needed by the k-nn algorithm to predict core. That mean that the k-mean prediction algorithm would be uitable when the uer-item matrix i not updated very often, in which cae the clutering procedure can be run off-line or in background. Beide accuracy and run-time, other criteria could alo be conidered when chooing between both algorithm. In particular, in open environment it i uually not difficult to perform profile injection attack that inert fake uer on the uer-item matrix in order to manipulate the recommendation. When that i the cae, model-baed algorithm, uch a the k-mean one, may be the bet choice, ince it ha been hown that they are more reitant to thi type of attack [18]. 4. RELATED WORK k-mean w ith ervice imilarity Figure 6. Computation time needed to contruct the cluter in the k-mean algorithm Recent reearch ha focued on CF for Web ervice recommendation. For example, [14] developed a prediction algorithm of QoS value for Web ervice that combine uer-baed and item-baed CF method. The predicted QoS i ued to recommend ervice to uer. In [22] it i preented a hybrid CF algorithm that cluter uer into region baed on imilaritie of their phyical location and hitorical QoS. The cluter are ued to identify region-enitive ervice, and a nearet neighbor approach predict the QoS of a candidate Web ervice for an active uer. The prediction occur by exploiting hitorical QoS information gathered from uer of highly correlated region. The ervice with the bet predicted QoS i then recommended to the active uer. In [23] it i alo addreed the problem of Qo prediction uing a neighborhood-baed collaborative filtering approach for QoS baed ervice election. While the above cited work ue QoS parameter a indicator of uer interet, [15] point out that Web ervice QoS parameter, uch a availability and repone time, are too limited to capture the experience provided to end uer. In our work, we aume that uer interet in a ervice i repreented by explicit or implicit rate provided by the uer. In [20] it i decribed a framework for Web ervice election inpired by memory-baed CF method that conider the dependencie among Web ervice in compoition procee. The invocation rate of a Web ervice carried out by a uer in different Web ervice compoitional procee i ued a an indicator of the uer preference for that ervice. The experimental evaluation of the framework wa performed uing the Movielen data et to imulate Web ervice compoition. The main focu of the author i ervice election during a compoition proce, when ome or all of the ervice to be compoed are already known. Uing general Web ervice naming tendencie coupled with enhanced yntactical method, the work in [24] aggregate ervice by their meage and proactively ugget candidate ervice to uer. Service imilarity to predict uer feedback i examined in [25], although in a different context from our. The author propoe a method for ervice dicovery that combine multiple matching criteria with uer feedback, baed on the aumption that uer rate how appropriate retrieved ervice are according to the reult of their requet. Conidering a given pair with one requet R and one ervice S, when no rating exit in the databae, the method take into account not only the rating aigned to the current ervice requet R, but alo rating aigned to requet imilar to R. Differently from our work, and from CF method in general, all available uer rating are conidered equally important, independently of uer imilarity. The predicted feedback value (core) i computed a the average of all uer rating for the correponding ervice. In contrat to our work, none of the above urveyed article ue emantic imilarity of ervice a a trategy to increae accuracy under pare data condition. 5. CONCLUSIONS AND FUTURE WORK In thi paper we preented algorithm for the contruction of emantic Web ervice recommender uing CF. The focu of our work wa to ue emantic markup for Web ervice to increae the accuracy of the recommendation baed on CF algorithm when the uer-item matrix i pare. We implemented and evaluated two algorithm for recommendation: a memory-baed algorithm uing the k- NN method, and a model-baed algorithm uing the k- mean method. In both algorithm, the imilarity between uer i computed by the Pearon Correlation Coefficient (PCC). Uually, when the PCC method i employed in ueruer CF algorithm, the imilarity between two uer i computed utilizing the rating given by the uer to item (ervice) rated in common. If the uer have not rated item in common it i not poible to compute their imilarity. In addition, when the number of common rated item i low, the reliability of the computed imilarity degree may alo be low. In our algorithm, intead of only uing the rating of common ervice, the rating of ervice that are emantically imilar to thoe ervice rated by the uer are alo taken into conideration. Likewie, when predicting the rating a target uer will give to a target item he ha not yet acceed, the algorithm conider the rating 86
given to the target item by neighbor uer (or group, in the cae of the k-mean algorithm), a i cutomary, while alo conidering the rating given by neighbor to item that are emantically imilar to the target item. The experimental evaluation decribed how that conidering imilar ervice when computing uer imilarity and predicting uer rating ha a ignificant impact on the accuracy of the implemented algorithm, particularly when the uer-item matrix i pare. A expected, the memory-baed algorithm uing k-nn wa more accurate than the model-baed algorithm baed on k- mean, but the k-mean algorithm i more calable when the dynamic of the application domain permit the clutering proce to be run in background. It i alo intereting to point out that when the k-mean algorithm conider imilar ervice it ha higher prediction accuracy than the k-nn baed algorithm when the latter doe not take ervice imilarity into account. A a final remark it i worth noting that recommender ytem uually preent their recommendation in decreaing order of predicted uer interet, and that uer frequently conider only the top n rated item. In [26] it i oberved that CF algorithm baed on the k-nn method make ome obcure or inaccurate recommendation at the top poition when implemented uing the PCC to find neighborhood. Uually that behavior i not evident becaue the algorithm are commonly rated uing the MAE (a wa done in our work). Thi metric favor algorithm that have a low average error rate over a et of prediction, but that do not necearily place the n bet recommendation at the top of the lit. Thi performance limitation ha two primary ource: (1) target uer with few neighbor who have rated an item and (2) target item rated by neighbor with low correlation to the target uer. The PCC addree the econd problem giving more influence to neighbor with higher imilarity. But thi trategy doe not account for cae where all the neighbor have low correlation with the target uer. Although we have not analyzed the quality of the top n recommendation, we can notice that the two mentioned ource of poor performance are related to data parity. And a uch, our algorithm contribute to alleviate both ource of low performance: (1) by enlarging the neighborhood through conidering not only uer who have rated the ame ervice, but alo uer who have rated imilar ervice; and (2) by etting a threhold to the minimum imilarity between two uer that mut be oberved to permit the placement of uer in the ame neighborhood. Reference [1] Papazoglo M. P., and Georgakopoulo, D. (2003). Service-Oriented Computing. Communication of the ACM, 46(10), 25 28. [2] Pan, Y., Tang, Y., & Li, S. (2011). Web Service Dicovery in a Pay- A-You-Go Fahion. Journal of Univeral Computer Science, 17(14), 2029 2047. [3] Chritenen, E., Curbera, F., Meredith, G., and Weerawarana, S. (2001). Web Service Decription Language (WSDL) 1.1, 2001. At http://www. w3. org/tr/2001/note-wdl-20010315. Retrieved from http://www.w3.org/tr/ 2001/NOTE-wdl-20010315 [4] McIlraith, S. A., Son, T. C., and Zeng, H. (2005). Semantic web ervice. Intelligent Sytem, IEEE, 16(2), 46 53. [5] Pedrinaci, C., and Domingue, J. (2010). Toward the Next Wave of Service: Linked Service for the Web of Data. Journal of Univeral Computer Science, 16(13), 1694 1719. [6] W3C OWL Working Group. (2012). OWL 2 Web Ontology Language Document Overview (Second Edition). http://www.w3.org/tr/owl2- overview/ [7] Brickley, D., and Guha, R. V. (2006). RDF Vocabulary Decription Language 1.0: RDF Schema, 2004. Retrieved from http://www. w3. org/tr/rdf-chema [8] Teto, V., Anagnotopoulo, C., and Hadjiefthymiade, S. (2006). On the Evaluation of Semantic Web Service Matchmaking Sytem. Web Service, 2006. ECOWS 06. 4th European Conference on, 255 264. [9] Sreenath, R. M., and Singh, M. P. (2004). Agent-baed ervice election. Web Semantic: Science, Service and Agent on the World Wide Web, 1(3), 261 279. [10] Stein, S., Payne, T. R., & Jenning, N. R. (2009). Flexible proviioning of web ervice workflow. ACM Tranaction on Internet Technology (TOIT), 9(1), 2 [11] Tizzo, N. P., Adán-Coello, J. M., & Cardozo, E. (2011). Automatic compoition of emantic web ervice uing A-Team with genetic agent. In Evolutionary Computation (CEC), 2011 IEEE Congre on (pp. 370 377). [12] S X., and Khohgoftaar, T. M. (2009). A Survey of Collaborative Filtering Technique. Advance in Artificial Intelligence, 2009, 1 19. [13] Burke, R. (2000). Knowledge-baed recommender ytem. Encyclopedia of Library and Information Sytem, 69, 175 186. [14] Zheng, Z., Ma, H., Ly M. R., & King, I. (2011). QoS-aware Web ervice recommendation by collaborative filtering. Service Computing, IEEE Tranaction On, 4(2), 140 152. [15] Van Moorel, A. (2001). Metric for the Internet Age: Quality of Experience and Quality of Buine. Fifth International Workhop on Performability Modeling of Computer and Communication Sytem, Arbeitberichte de Intitut f\ür Informatik, Univerit\ät Erlangen- N\ürnberg, Germany (Vol. 34, pp. 26 31). [16] Martin, D., Burtein, M., Hobb, J., Laila, O., McDermott, D., McIlraith, S., Narayanan, S., et al. (2004). OWL-S: Semantic Markup for Web Service. Retrieved from http://www.w3.org/submiion/owl-s [17] Kluch, M., Frie, B., and Sycara, K. (2009). OWLS-MX: A Hybrid Semantic Web Service Matchmaker for OWL-S Service. Web Semantic: Science, Service and Agent on the World Wide Web, 7(2), 121 133. doi:10.1016/j.webem.2008.10.001 [18] Mobaher, B., Burke, R., and Sandvig, J. J. (2006). Model-Baed Collaborative Filtering a a Defene Againt Profile Injection Attack. Proceeding of the National Conference on Artificial Intelligence (Vol. 21, p. 1388). [19] Mobaher, B., Dai, H., Luo, T., and Nakagawa, M. (2002). Dicovery and Evaluation of Aggregate Uage Profile for Web Peronalization. Data Mining and Knowledge Dicovery, 6(1), 61 82. [20] Rong, W., Li K., and Liang, L. (2009). Peronalized Web Service Ranking via Uer Group Combining Aociation Rule. Proceeding of the 2009 IEEE International Conference on Web Service-Volume 00 (pp. 445 452). [21] Herlocker, J. L., Kontan, J. A., Terveen, L. G., & Riedl, J. T. (2004). Evaluating collaborative filtering recommender ytem. ACM Tranaction on Information Sytem (TOIS), 22(1), 5 53. [22] Chen, X., Li X., Huang, Z., and Sun, H. (2010). RegionKNN: A Scalable Hybrid Collaborative Filtering Algorithm for Peronalized Web Service Recommendation. 2010 IEEE International Conference on Web Service (pp. 9 16). Preented at the 2010 IEEE International Conference on Web Service (ICWS), Miami, FL, USA. doi:10.1109/icws.2010.27 [23] W J., Chen, L., Feng, Y., Zheng, Z., Zho M. C., & W Z. (2013). Predicting quality of ervice for election by neighborhood-baed collaborative filtering. Sytem, Man, and Cybernetic: Sytem, IEEE Tranaction On, 43(2), 428 439. [24] Blake, M. B., & Nowlan, M. F. (2007). A web ervice recommender ytem uing enhanced yntactical matching. In Web Service, 2007. ICWS 2007. IEEE International Conference on (pp. 575 582). [25] Averbakh, A., Kraue, D., & Skouta, D. (2009). Exploiting Uer Feedback to Improve Semantic Web Service Dicovery. Preented at the 8th International Semantic Web Conference (ISWC 2009). [26] McLaughlin, M. R., and Herlocker, J. L. (2004). A collaborative filtering algorithm and evaluation metric that accurately model the uer experience. Proceeding of the 27th annual international ACM SIGIR conference on Reearch and development in information retrieval (pp. 329 336). 87