A review on the applications of bayesian network in web service

. Web services (WS) are the preferred approach in realizing the service-oriented computing paradigm. However, this comes with challenges such as complexity and uncertainty that hinder their practical application. Bayesian networks (BNs) are one of the techniques used to address these challenges. The objective of this mapping study was to determine what is known about the use of Bayesian networks in web services research. To do this, we identified and selected rigorously 69 articles (out of the 532 identified) published on the subject in 2001-2021. We then classified and analyzed these articles by Web service themes (Service Composition, Service Management, Service Engineering), Objectives (Prediction, Description, Prescription), Types of BN (Basic, Combined, Extended), and Evaluation methods (Proof of concept, Experiment, No evaluation). In doing so, we hope to provide a clear understanding of the subject. We also identify and suggest avenues for future research. Thus, the review results can help researchers and practitioners interested by the application of BNs in WS research.


Introduction
Web services (WS) have revolutionized software development practices.Defined as "software components that were self-described, loosely coupled, and easily integrated with one another" (Driss et al., 2022), WS are present in practically all fields (Bouguettaya et al., 2017;Zhao et al., 2022).This success is fueled, among other things, by the possibilities offered by WS in terms of cost reduction, ease of reuse and operational efficiency (Papazoglou et al., 2008;Zhao et al., 2022).However, the dynamic and unpredictable nature of WS (Papazoglou et al., 2008) leads to various problems, including complexity and uncertainty (Alférez and Pelechano, 2013;Gabarró and Stewart, 2021).All things that make their implementation difficult in practice.One of the ways to address these issues is the use of Machine Learning (ML) techniques (e.g., Purohit and Kumar, 2021;Razian et al, 2022;Song, 2021).Bayesian Networks (BNs), specially adapted to complex and uncertain situations (Rohmer, 2020), are one of these techniques.
Since their introduction in the 1980s (Pearl, 1986), BNs have become very popular as evidenced by the large number of fields where they are used (e.g., Bielza and Larrañaga, 2014;Chen et al., 2021;Hosseini and Ivanov, 2020;Kyrimi et al., 2021;Rosário et al., 2022;Xu et al., 2022Xu et al., , 2023)).This success is maintained despite the decades and the emergence of equally popular techniques such as the Artificial Neural Network (ANN), Support Vector Machine (SVM) or Deep Learning (DL).Indeed, compared to these techniques, BNs have certain advantages that make them unique (Correa et al., 2009;Hosseini and Ivanov, 2020;Kazem et al., 2015;Malekmohamadi et al., 2011;Müller et al., 2020;Weber et al., 2012).For example, with BNs, incomplete data or data of various kinds from WS, especially during their composition, can be integrated into the same model (Kaya et al., 2023;Larrañaga and Moral, 2011;Weber et al., 2012;Rohmer, 2020;Xu et al., 2023).In addition, the structure of this model allows to clearly distinguish the links between its elements (Larrañaga and Moral, 2011).Finally, the use of the model gives explainable results (Lacave and Diez, 2002;Müller et al., 2020), responding at the same time to the requirement of a predictable and responsible Artificial Intelligence (Kitson et al., 2023;Mauro et al., 2022).These properties make BNs a natural choice in fields such as WS (Hwang et al., 2007).
Furthermore, several literature reviews on the use of ML techniques in WS research are published (e.g., Batra She et al., 2019).But, to our knowledge, none of these reviews is devoted to BNs.
In others words, there is little information on how BNs are effectively used in the WS research.
Therefore, and in order to help interested researchers and practitioners, it is necessary and timely to understand how BNs are used to deal with the problems of WS.
The purpose of this research is to contribute to this understanding.Accordingly, we carried out a mapping study of 69 articles published during the period 2001-2021 in order to: • Describe the general profile of this literature (years, types, and countries of publication); • Determine the conditions of application of BNs based on a classification framework with four dimensions: (i) WS themes, (ii) objectives, (iii) types of BN, and (iv) evaluation methods; • Identify and propose avenues for future research based on this analysis.
The rest of this paper is organized as follows.The basic concepts are briefly defined in section 2. Section 3 describes the methods of the review, the results of which are presented discussed in section 4-8.Section 9 concludes the paper.

Background 2.1. Web service
Web services (WS) are applications developed and deployed on the Web according to the principles of Service-oriented computing (SOC).At the heart of the SOC is the notion of services, i.e., "... self-describing, open components that support rapid, low-cost composition of distributed applications."(Papazoglou and Georgakopoulos, 2003).To concretely translate its principles, SOC is based on the Service Oriented Architecture (SOA) which is "a means of structuring and reorganizing distributed software applications into a set of composed and interactive pre-existing services."(Driss et al., 2022).In addition to these high-level technologies, there are several essential concepts around WS.These notions are clearly summarized as follows: "... three key features of services are crucial: functionality, behavior, and quality.Functionality is specified by the operations offered by a service; Behavior reflects how the service operations can be invoked and is decided by the dependency constraints between service operations; Quality determines the non-functional properties of a service."(Bouguettaya et al., 2017, p. 70).In particular, non-functional properties are defined in the form of parameters (Hwang et al., 2007;She et al., 2019) which can be grouped into three categories (Driss et al., 2022): Quality of Service (QoS), Quality of Experience (QoE), and Quality of Business (QoBiz).QoS is "a set of parameters describing the behavior of Web services in terms of performance parameters."(Driss et al., 2022).Among these parameters, we can cite accessibility, availability, reliability, response time, robustness, scalability (Driss et al., 2022;She et al., 2019;Yu et al., 2008).QoE is "a measure of the end-to-end performance of a whole system as both resulting and taken from the user's point of view."(Driss et al., 2022).These parameters can be friendliness, success rate, and reputation.Finally, QoBiz aims to "describe the financial aspects of service provisioning, such as the price of service, the costs of service provisioning, the service provisioning revenue, and the revenue per transaction (comprised of cost per transaction) parameters."(Driss et al., 2022).(ii) constitute the qualitative part (structure) of the BN, while element (iii) is the quantitative part (parameter).Building a BN therefore comes down to determining these three elements (Kaya and Yet, 2019).In general, the values of the nodes (discrete or continuous) come either from data (real or synthetic) or from expert knowledge or from the combination of both.

Bayesian Network
The uncertainty of the causal relationship is represented by the conditional probability distributions P(  |Pa(  )) associated with each node   , where Pa(  ) is the parent set of   .

Methods
Due to the exploratory nature of this research, we chose to use the mapping study method (Petersen et al., 2015).A scoping review allows to identify, classify thematically, and describe the articles devoted to a given subject.Its purpose is to provide an overview of this subject by determining its nature, evolution, and limits (Petersen et al., 2015).The remainder of this section is organized into the following phases based on (Petersen et al., 2015): (i) Questions definition, (ii) Article identification, (iii) Article selection, and (iv) Article classification.

Questions definition
Our objective is to provide a snapshot of the published research work on the application of BNs in WS.To do this, we organized the review around the following research questions: • RQ1.What web service themes are addressed by the application of bayesian network?
• RQ2.What objectives are pursued when applying bayesian network in web services?
• RQ3.What types of bayesian network are frequently applied in web services?
• RQ4.What methods are used to evaluate the proposed bayesian network models?
The intent of question RQ1 is to determine which aspects of WS are explored through the use of BNs.With question RQ2, the goal is to identify the reasons that motivate the use of BNs to respond to the problems identified in RQ1.Question RQ3 aims to specify which types of BNs are actually implemented to answer question RQ2.Finally, question RQ4 characterizes how the performance of the BN models used is measured.

Article identification
For searches, we used ACM Digital Library (http://dl.acm.org),Google Scholar (https://scholar.google.com/),and IEEE Xplore (http://ieeexplore.ieee.org).These tools index the main publications on the subject at hand and have been used in similar reviews (e.g., Di Francesco et al., 2019; Rodríguez et al., 2016;She et al., 2019).In particular, we combined two groups of search expressions.The first group to refers terms related to the "web service" and the second to the "bayesian network" (see Table 1).The searches (done on September 27, 2022) produced 523 articles which were saved in EndNote X.9 (Thomson Reuters, Philadelphia, USA).After removing duplicates, the remaining 384 articles were submitted for selection.

Article selection
In addition to the constraints introduced in the search, each article had to meet the following criteria to be selected: i. Published in English; ii. Published in refereed journals, international conferences (congresses), or workshops; iii.Published after 1999; iv.Focused on BNs application in WS.
This implies that are excluded: i.No peer reviewed publications (e.g., Research reports, dissertations or theses, books and book chapters, preprints, working papers), editorials, opinion pieces, commentaries, reviews, etc.; ii. Articles not available in full text; iii.Conference papers subsequently published in a journal.
These criteria were used to select articles in two stages.First, the examination of the title and the abstract of the articles allows to select 156 articles.Then, the full text of these 156 articles was read to determine their relevance.At the end of this process, 69 articles were selected and classified.

Article classification
To structure and facilitate the classification of the papers, we developed a scheme (Fig. 1) based on the questions of the review.The details of this scheme are presented in the following subsections.

Evaluation methods
This aspect concerns methods used to evaluate the performance of the BN models.In this review, we consider the following groups of methods: i. Proof of concept.Concerns papers that evaluate the performance of their BN model by example/demonstration/illustration/proof of concept; ii. Empirical.Groups papers that use empirical methods (e.g., experiments, simulation, survey) to evaluate their BN model; and iii.No evaluation.Refers to papers that do not evaluate their BN model.
These categories were used to manually classify the articles.Even if the categories are not necessarily disjoint, as a general rule, an article is classified in one and only one category.In case an article covers more than one category, we choose the one that best suits the authors' objective.

Limitations
Like any research, our review has limitations.These relate in particular to the identification, selection, and classification of articles.With regard to the identification of articles, it may be that, despite all the rigor we have put into developing them, our search expressions are limited.
And, as a result, some relevant articles were not found.However, the number and diversity of the articles finally included in the review ensure that the subject is well represented.As for the selection of articles, we probably eliminated a few by mistake.To reduce this problem, we defined, a priori, eligibility criteria that we applied as rigorously as possible.Finally, when it comes to the classification of articles, another researcher may not obtain exactly the same results as us.Nevertheless, we have defined and used a classification framework whose categories come from both the literature on the subject and the selected articles.We hope that these categories are sufficiently clear and high level to facilitate the reproduction of the classification.

Overview of the selected papers
The profile and evolution of the selected literature was determined by examining its (i) years of publication, (ii) types of publication, and (iii) geographical distribution (See Table A1 in appendix for details).Fig. 2 presents the results of this analysis, the details of which are described in the following subsections.Publication types and venues.In terms of publication types, Fig. 2 shows that 55% of the papers come from conferences, 36% from journals, and 9% from workshops.According to  Geographical distribution.Geographically (i.e., the country of affiliation of the main author of each paper), Fig. 2 shows that Asia, with 62% of papers, largely dominates the list of continents that publish on the subject.The Americas and Europe follow by far with 17% and 13% of papers respectively.Finally, Africa and Oceania bring up the rear with 4% of papers each.When examining the papers in detail, we identify that the 69 papers originate from 19 countries.China (28 papers), followed by India (8 papers) and the USA (8 papers) are the three countries that contribute the most to the subject.

Summary
From a bibliographic point of view, the results reveal that the 69 included papers (i) are published mainly in conferences; (ii) increased sharply in number between 2001-2007 and 2008-2014; and (iii) have authors predominantly from Asia (particularly China).

Web service themes (RQ1)
The thematic classification of the 69 papers (Table 3) shows that Service composition comes first (45% of papers), followed by Service management (33%), and Service engineering (22%).

Results
In this section, the different sub-themes are defined as well as a description of the papers they group together.

Service composition
Table 3 shows the distribution of the 31 papers of this theme according to the three sub-themes.
The majority of papers (55%) concerns the Service selection sub-theme.Next comes the Service discovery sub-theme (29%), which represents just over half of the papers included in the Service selection.Finally, the Service recommendation sub-theme (16%), accounts for less than a third of the papers included in the Service selection.Detailed information on these subthemes is available in the appendix (Table A2).
Service selection consists of "Choosing the most adequate service among discovered candidates, according to functional or non-functional properties" (Huf and Siqueira, 2019). According

Service management
According to the Table 3, the 23 papers of this theme are subdivided almost equally in two subthemes: Service control (52%), and Service monitoring (48%).Detailed information on these sub-themes is available in the appendix (Table A3).
Service control aims to "… improve the service quality through a set of control mechanisms

Service engineering
Table 3 shows that of the 15 papers of this theme, 60% are devoted to Service development, and 40% to Service application.Detailed information on these sub-themes is available in the appendix (Table A4).
Service development concerns the design and development of WS.The papers of this subtheme relate to a prototype of WS to support the daily life (P25) or tourism services (P11).
Others focus on web-based online data (dependency) analysis tool (P12), diagnosis service (P18), an ontology-based WS (P19) or WS API which computes learners' competence and capability assessment (P24).Finally, others explore the field of the supply chain (P14, P16) and an intelligent WS (P27).
Service application refers to the use of BNs to examine existing WS integrated in frameworks, prototypes, tools, etc.In a prototype called "whereabouts diary", a white-pages WS are used to extract information about places visited by users and BNs to classify places (P13).In (P15), an intelligent system based on spatial WS (GIS functions) to provide personalized recommendations for tourist attractions is proposed.In the (P20), a geospatial WS is integrated into Enterprise Business System.Furthermore, a diagnostic functionality is exposed through a web API in (P21), and in (P22), an interactive recommender system based on a WS, is used to manage patient information.Finally, in (P26), the authors used BNs to analyze the sensitivity of a prototype of WS.

Summary
In response to question RQ1 ("What web service themes are addressed by the application of bayesian network?"), the review reveals the predominance of the Service composition (Fig. 3).This is hardly surprising considering that service composition is the "raison d'être" of the SOC paradigm (Papazoglou et al., 2008)   Service engineering.Table 4 shows that in this theme, 47% (7/15) of papers have a Descriptive objective (P11, P13, P14, P18, P21, P22, P25), 40% a Predictive objective (P12, P15, P16, P24, P26, P27), and 13% a Prescriptive objective (P19, P20).In these last papers, the prescription takes the form of Framework (P20) and Method (P19).Table 6 also shows the almost equal distribution of the different types of objective between the two sub-themes.

Summary
Concerning the question "What objectives are pursued when applying bayesian network in web services?"(RQ2), Fig. 4 suggests that the main reason for using BN is for prediction with a focus on the composition and management of services.These results are quite logical since these two themes contain activities aimed at predicting or evaluating WS according to predefined criteria.Furthermore, the results highlight the lack of popularity of the Prescriptive objective, particularly in Services Engineering (only 2 papers are concerned -see Table 4).
This aligns perfectly with (Bouguettaya et al., 2017, p. 68) who remarked that "Service systems have so far been built without an adequate rigorous basis from which to reason about them".
However, the activities of the Service engineering must be carried out according to precise prescriptions in order to design and develop applications based on the WS.Thus, we suggest that future research should pay more attention to how BNs are used for prescriptive purposes in service engineering.For this, we can rely on models such as those proposed in (Kurniawan   Service management.According to Table 5, 70% of the 23 papers in this theme used Basic BN (P23, P28, P29, P30, P31, P32, P33, P34, P37, P38, P40, P42, P43, P44, P46, P49), 21% used the Extended BN (P39, P41, P45, P47, P48), and 9% used the Combined BN (P35, P36).
In the papers that used the combined type, the BN is associated with an Agent (P36) and Ontology (P35).For the Extended type, all the BN types were Dynamic BNs.We also note that the Basic BN is generally used in the Service control sub-theme, while the Combined BN mainly concerns Service monitoring.Finally, the Extended BN is used equally in both sub-themes.
Service engineering.In term of the types of BN, Table 5 reveals that, among the 15 papers of this theme, the Basic BN was used in 80% of papers (P11, P12, P13, P14, P16, P18, P20, P21, P22, P24, P25, P26), and the Combined BN in 20% of papers (P15, P19, P27).Regarding the combined type, the BN was combined with Neural network and Ontology (P19), Multi-entity (P27) and Analytic hierarchy process (P15).It is important to note the total absence of use of the Extend BN type in the concerned sub-themes.Moreover, the Basic BN and Combined BN were mainly used in the Service development sub-theme.

Summary
Regarding question RQ3: "What types of bayesian network are frequently applied in web services?", the answer is that the Basic BNs are the most used, and this, mainly in Service composition (Fig. 5).These results can be explained by (i) the relative ease of use of this type of BN and (ii) their ability to visually represent the dependencies between the different elements of a WS.Which is a facilitating element (Zhao et al., 2022) in the particular case of Service composition.But, at the same time, for complex and dynamic phenomena such as WS (Papazoglou, 2008), "description" alone is not enough.We need slightly more adapted techniques like DBN to better understand these phenomena.Our results suggest that, if this form of BN is actually used, it remains marginal (8 papers).This could be explained by the complexity of DBNs (Bielza and Larrañaga, 2014); which can notably increase their computation time (Hosseini and Ivanov, 2020).The same goes for the Combined BN which, like the Extended BN, concerns only 9 papers.However, as "BNs are limited by the modeling aspects that they can deal with" (Weber et al., 2008), it is necessary to combine them with other techniques in order to correctly model the phenomenon under study.Therefore, these constraints must be taken into account when considering using Combined and Extended BNs in a WS context.The predominance of proof of concept suggests the poor quality of the studies proposed.
Indeed, "...these studies are only demonstrations that a technology works..." (Sjøberg et al., 2007).Moreover, the results reveal that the empirical methods used (Experimentation and Simulation) are, for the most part, based on small samples.In other words, the BN models in the reviewed papers are not robustly and convincingly evaluated.This means that these evaluations do not constitute a solid basis for making informed decisions.Therefore, a possible avenue for future research is to improve the quality of BN model evaluations by carefully selecting the methods used.).Thus, their future in the "sphere of Analytics" (Mishra et al, 2023) is more than promising.

A
BN is a directed acyclic graph which represents in the form of arcs or lines causal dependencies between the variables of the phenomenon studied.At the structural level, a BN is composed of three elements (Kaya and Yet, 2019; Xu et al., 2022): (i) Node, which indicates the variables; (ii) Directed arcs/lines with arrows which represent the causation relationship between nodes; (iii) Conditional Probability Table (CPT), which contains the conditional probability of each state of the nodes, to quantify the causation relationship.Elements (i) and

Fig. 2 .
Fig. 2. Distribution of papers by years, publication types, and continents to our results, the 17 papers of this sub-theme propose BN models based on nonfunctional requirements such as QoS(12 papers)  and QoE (5 papers).For QoS, 8 papers explore QoS in general (P04, P09, P17, P55, P58, P63, P64, P68).The other 4 relate to specific QoS parameters such as Service organization (P67), Data quality (P59), Response time (P65) or Performance (P66).Regarding QoE, the papers explore Trust (P57, P60, P61, P62) andTrust and reputation (P56).Service discovery aims to "Locating relevant services that offer some desired data or functionality" (Huf and Siqueira, 2019).In general, three approaches are used for Service discovery: Syntactic-aware, Semantic-aware, and Context-aware (Huang and Zhao, 2022;Rodríguez et al., 2016).According to the results, 7 of the 9 papers of this sub-theme use BN models based on the Semantic-aware approach (P01, P02, P04, P06, P07, P10, P69).The other 2 papers deal with BN models based on the Context-aware approach (P05, P08).In particular, in the paper (P07), the authors used the Semantic-aware mode to explore QoS ("Quality as functionality").No paper used the Syntactic-aware approach.Service recommendation refers to "… the process of automatically identifying the usefulness of services and proactively recommending services to end users."(Yaoet al., 2015, p. 453).In particular, this can facilitate the service composition(Wu et al., 2015).Examination of the papers of this sub-theme suggests that, in general, the Service recommendation is a support task.Thus, the 5 papers included in this sub-theme use the recommendation to support Service discovery (P50, P51) and Service selection (P52, P53, P54).Note that all papers are based on the "QoS-aware recommendation method"(Li et al., 2021).

Fig. 4 .
Fig. 4. Papers on Research objectives by Web service themes

Fig. 5 .
Fig. 5. Papers on Types of BN by Web service themes

Table 1 . Article identification (2022-09-27) Source Description Results
Using the results of the description and the prediction to define or propose a normative approach (framework, method, platform, or procedure) that favorize the use of WS.
Evaluation methodsProof of conceptEmpiricalNo evaluation Research objectives Descriptive Predictive Prescriptive Types of BNs Basic Extended Combined Web service areas Based on (Guerra-Montenegro et al., 2021; Mishra et al, 2023) and the objective of the selected papers, we categorize the reasons for using BNs in WS as descriptive (exploratory), predictive, or prescriptive.i.Descriptive objective.Characterize (classify, explain, model, represent, or understand) WS or its elements by using BNs;ii.Predictive objective.Appraise (assess, calculate, estimate, evaluate, forecast, measure, predict, prognosis, sizing) state of WS or its elements by using BNs;iii.Prescriptive objective.3.4.3.Types of BNIn this review, we distinguished the following types of BNs (e.g.,Larrañagaand Moral, 2011; Marcot and Penman, 2019; Weber et al., 2012): i. Basic.Concerns the standard form of BNs in which data contains only discrete variables; ii.Extended.More elaborated forms of BNs such as dynamic BNs, hierarchical BNs, object-oriented BNs, relational BNs, etc.This type of BN also concerns those that contain (i) both continuous and discrete variables (Hybrid BN), (ii) only continuous variables (Continuous BN); and iii.Combined.Joint use of BNs with other techniques such as AHP, Fuzzy logic, Neural network, simulation, etc.

Table 2 ,
one journal and four conferences published more than one paper.Most publications (78%) are limited to one paper each.Note that the journal Expert Systems with WS, the International Conference on Web Services and the International Conference on Services Computing are among the best in their respective fields.

Table 3 . Papers on Web service themes and sub-themes (N = 69) Theme Papers Service composition (31 papers)
when they plan to study BNs.Finally, regarding the sub-theme Service recommendtion, the results show that no paper mentions the type of recommendation approach used.Therefore, it would be important to explore how recommendation approaches (e.g., collaborative filtering, content-based and hybrid) may be used in concert with BNs in a WS context.
BNs.Among the 69 analyzed papers, 33 (48%) deal with one aspect or another of Quality (22papers on Service composition, and 11 on Service management).Based on these observations, we suggest that researchers pay more attention to the Functionality and Behavior of WS

Table 4 . Papers on Research objectives by web service themes (N = 69) Theme Research objective
Method (P54).More specifically, the Descriptive objective mainly concerns the Service selection and Service discovery sub-themes.The Predictive objective is mainly used in Service selection.Finally, the Prescriptive objective is found mainly in the Service discovery papers.objective (P23, P28).For the papers with prescriptive objective, the prescription takes the form of Approach (P29, P32, P38, P41, P44, P47), Framework (P31, P37) and Method (P30, P34, P36).Note that the Descriptive objective was used only in the papers of the Service control sub-theme.More specifically, the Descriptive and Prescriptive objectives mainly concern the papers of the Service control sub-theme.As for the Predictive objective, it mainly concerns Service monitoring papers.

Table 5 . Papers on Types of bayesian network by web service themes (N = 69) Theme Type of bayesian network
in the form of Dynamic BN (DBN).Finally, in the case of papers that used the combined type, BN is combined with stochastic local search (P09), fuzzy logic (P55), Hidden Markov Model (P65) and cuckoo search algorithm handset (P68).A closer examination shows that Basic BN and Combined BN are mostly used in Service selection.As for the Extended BNs, they are only used in the Service discovery.