1 Introduction
The concept of personalization in its many forms has gained traction driven by the demands of computer-mediated interactions generally implemented in large-scale distributed systems (LSDS), which may include ad hoc wireless networks. Personalization requires the identification and selection of entities based on a defined profile (a context). While research into personalization has generally addressed individuals, investigations are increasingly considering the application of context to a diverse range of entities; an entity has been defined as: ‘a person, place, or physical or computational object’ (Dey & Abowd, Reference Dey and Abowd1999). Recent developments have shown that the underlying motivation for the use of context is personalized service provision (PSP) to a range of entities in an increasingly broad and diverse range of domains and systems.
Historically, investigations into context have predominantly focused on two domains: (1) office applications and (2) tourist guides; the primary focus being on the use of location, identity, and proximate data (Moore et al., Reference Moore, Hu and Wan2010b). Investigations are, however, seeking to expand the domains addressed along with the range of data (contextual information) utilized. These advances are exemplified in the growing body of research that has investigated the application of context in domains including pedagogic systems, health monitoring, memory aids, field research tools, motion capture, group interaction support, pervasive computing, adaptation, ad hoc wireless networks, and group-based decision-support systems (Moore et al., Reference Moore, Hu and Wan2010b; Moore & Pham, Reference Moore and Pham2012).
Documented research has addressed context as it applies to information systems (IS). Space restricts a detailed discussion on IS, however, a review of related research with a synopsis of the contextual information used and the approaches to implementation can be found in Moore et al. (Reference Moore, Hu and Wan2010b). There is no commonly agreed definition of the term IS or what constitutes an IS (Alter, Reference Alter2008); in fact any system which operates on captured data which can be codified, digitized, and is machine processable to produce useful information can be viewed as an IS.
The traditional view of information resulting from data processing has been extended by Checkland and Holwell (Reference Checkland and Holwell1997) to include an intermediate stage termed Capta which is defined as ‘data selected, created, or to which attention is paid’ and ‘involves the selection of data for processing into information useful to users’. This is analogous to the concept of relevance (Hildreth, Reference Hildreth1998; Coppola & Della, Reference Coppola and Della2004). The observations of Checkland and Holwell (Reference Checkland and Holwell1997), Hildreth (Reference Hildreth1998), and Coppola and Della (Reference Coppola and Della2004) clearly identify the synergy between relevance and PSP. The objectives for PSP are: the provision of relevant and appropriate services to users concomitant with the mitigation of the potential to deliver irrelevant services. PSP implemented in intelligent context-aware decision-support systems has the potential to increase the relevance of PSP, mitigate the issues surrounding information overload (Moore et al., Reference Moore, Hu and Jackson2010a), and accommodate constraint satisfaction (CS) and preference compliance (Moore et al., Reference Moore, Jackson and Hu2010d).
The approach proposed in this article to enable the realization of PSP is an event-driven fuzzy rule-based decision-support system with ontology-based context modeling (OBCM) (Moore et al., Reference Moore, Hu, Zhu, Campbell and Ratcliffe2007) implemented using Semantic Web technologies (W3C, 2012) with the Jena API (Jena, 2012); a discussion on rule-based systems (RBS) and the Semantic Web can be found in Moore et al. (Reference Moore, Hu and Jackson2011). Research into RBS has investigated rules and rule strategies in diverse domains including (1) database systems, (2) data mining and warehousing, (3) fuzzy logic and fuzzy rule-based systems (FRBS), (4) decision trees, (5) artificial neural systems, and (6) logical reasoning and inference (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b). A discussion on RBS and rule strategies can be found in Hayes-Roth (Reference Hayes-Roth1985), Gonzales and Dankel (Reference Gonzales and Dankel1993), Mitchell (Reference Mitchell1997), Moore et al. (Reference Moore, Hu and Jackson2011); a comprehensive exposition on FRBS is presented in Berkan and Trubatch (Reference Berkan and Trubatch1997). In summary, the literature demonstrates the efficacy of RBS and FRBS in a broad range of applications in diverse domains, systems, and technologies (Moore et al., Reference Moore, Hu and Jackson2011).
An analysis of the domains addressed and the methodologies used to capture and process contextual data clearly demonstrates (1) the reliance on location, identity, and proximate data in the case of the early context-aware systems and (2) the growing innovation in the usage of the available data in the provision of PSP. Important conclusions drawn are the diverse nature of context, its inherent complexity, and the difficulty in defining dynamically a user’s context (Moore, Reference Moore2009; Moore & Pham, 2011, Reference Moore and Pham2012). Addressing these limitations demands PSP and the approach proposed in this article, in applying intelligent context processing (CP), has been shown to provide an effective basis upon which computational intelligence can be implemented and the inherent complexity of context leveraged.
The remainder of this article is structured as follows: personalization in LSDS is introduced with consideration of the relationship between such systems, context, and CP. This is followed by a discussion around rule strategies and conditional relationships in decision support. Logic systems are addressed with an overview of the open world assumption (OWA) versus the closed world assumption (CWA) and their relationship with the Semantic Web. The event-driven rule-based approach (which forms the basis upon which intelligent CP can be realized) is presented with an evaluation and proof-of-concept. There is a discussion which considers the issues and challenges identified in the research with potential solutions and research directions; alternative approaches to CP are discussed. The article closes with conclusions and open research questions.
2 The computing landscape
The developments in the computing are reflected in the PSP in LSDS which are characterized by the ubiquity of mobile and wireless systems.
There is a large body of documented research addressing personalization; the focus is (generally) on accommodating user preferences to enable service mediation and content adaptation based on context. Recently, however, there has been interest in expanding the use of context in systems where PSP with decision support forms a central function. Viewed from this perspective, the concept of personalization has been expanded to address an increasingly diverse range of potential contextual information in IS; this is exemplified by the growing interest in the use of kansei words in applications using kansei engineering (Salem et al., Reference Salem, Nakatsu and Rauterberg2009).
The traditional paradigm of centralized Internet networks has been largely replaced by a new distributed Internet-based communications paradigm characterized by networks created using centralized systems and distributed ad hoc wireless networks that are increasingly accessing geospatial, temporal, and cloud-based systems (Kermarrec et al., Reference Kermarrec, Massoulie and Ganesh2003; Nurmi et al., Reference Nurmi, Wolsk, Grzegorczyk, Obertelli, Soman, Youseff and Zagorodnov2009). These developments have resulted in systems in which a user’s current state (termed context) (which describes and defines a user’s prevailing personal, environmental, proximate, and social situation) may be static or, in mobile systems, highly dynamic. Mobile systems incorporate a diverse range of geographically diverse infrastructures with a potentially large and highly dynamic user base and a broad range of fixed and wearable mobile devices as discussed in Moore et al. (Reference Moore, Hu and Wan2010b).
The Internet systems paradigm is frequently characterized by LSDS (Kermarrec et al., Reference Kermarrec, Massoulie and Ganesh2003; Nurmi et al., Reference Nurmi, Wolsk, Grzegorczyk, Obertelli, Soman, Youseff and Zagorodnov2009; Lan et al., Reference Lan, Zheng and Yawei2010) in which interactions between infrastructure components and individual users (more accurately the users devices) are inherently complex, the complexity increasing exponentially as the nodes in the system increases. This complexity places great strain on communication systems with issues in the management of the interactions and the ability to enable PSP that requires both user and network infrastructure contextual information to effectively target service provision. LSDS are inherently context aware, context performing an increasingly important role. The rule-based approach presented in this article is posited as an effective approach to enable (1) targeted service provision, (2) the processing of data from a diverse range of geographically and technologically diverse sources, (3) the capability to handle the inherent complexity of context, (4) the ability to manage CS, and (5) the ability to realize predictable decision support under uncertainty.
3 Context and context-aware systems
The issues of personalization, intelligent context and its inherent complexity, CP, and context matching (CM) with partial matching (PM), and CS have been explored in Moore (Reference Moore2009), Moore et al. (Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Wan2010c, Reference Moore, Jackson and Hu2010d), and Moore and Pham (2011, Reference Moore and Pham2012). This section considers the system design requirements as they relate to intelligent context-aware decision-support systems, context, and the nature of potential contextual information.
3.1 System design requirements
The system design requirements which must be accommodated can be summarized as follows:
1. The implementation of predictable decision support. Arriving at a decision requires the matching of the context properties that describe an input with the corresponding context properties describing an entity (the output). This matching process is CM in which the probability of achieving a perfect match is low (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Jackson2011; Moore & Pham, 2011, Reference Moore and Pham2012); therefore, CM must be capable of reaching a decision where PM is the general use-case. Figure 1 graphically models the CM and PM problem.
2. For systems where PSP is the aim, the goal for CP and CM is to arrive at a Boolean decision as to the suitability of an entity (generally an individual user or his/her mobile device) for PSP.
Context matching (CM): shown are the input and the corresponding output contexts with the CM and the decision boundary (the approach to defuzzification as discussed in Sections 4.2, 5, and 7)
The input (context properties) and therefore the corresponding output (context properties) (more accurately stated as the literal values that apply to a property as discussed in subsequent sections) are potentially highly dynamic and can be extremely variable being dependent on the input which is generally unique (every use-case is new and individual). This is emphasized in the approaches considered in this article which include a broad range of applications including recent research addressing health-monitoring systems and a novel recommender system which targets the identification of stock-market trading strategies based on context-aware group-based decision support (Pham et al., Reference Pham, Thang, Nakaoka, Cooper and Kamei2011).
The implementation of CM with PM, CS, and decision support are not independent processes but are closely interconnected. To address these requirements the novel context-processing algorithm (CPA) has been developed as articulated in this article.
3.2 Context and contextual information
Context awareness describes a concept in which the profile of an entity is defined by its context, an entity being defined in Dey and Abowd (Reference Dey and Abowd1999) as ‘a person, place, or a physical or computational object’. Context awareness employs context to identify individuals and implement PSP. Context is highly domain and application specific requiring the identification of domain-specific function(s) and properties (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Wan2010c) and therefore must reflect a user’s current dynamic state (Moore et al., Reference Moore, Hu and Jackson2010a). Location is central to context; context, however, includes more than just location (Moore et al., Reference Moore, Jackson and Hu2010d). As identified in Sheth and Ramakrishnan (Reference Sheth and Ramakrishnan2003) and Moore et al. (Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b), a broad and diverse range of context factors combine to form a context definition, in fact, almost any information available at the time of an interaction can be viewed as contextual information including:
1. the variable tasks demanded by users;
2. the diverse range of mobile devices and the associated service infrastructure(s) in LSDS incorporating Wide Area Networks, Local Area Networks, and Personal Area Networks (PAN) implemented in Internet and Internet systems;
3. resource availability (connectivity, battery condition, display, network, and bandwidth, etc.);
4. nearby resources (accessible devices and hosts including I/O devices;
5. the physical situation (temperature, air quality, light, and noise level, etc.);
6. the social situation (who you are with, people nearby, etc.—proximate information);
7. spatial information (location, orientation, speed and acceleration. etc.);
8. temporal information (time of the day, date, and season of the year);
9. physiological measurements (blood pressure, heart function—electrocardiography (ECG or EKG from the German Elektrokardiogramm), cognitive functions related to brain activity (EEG from Electroencephalography), respiration, galvanic skin response, and motor functions including muscle activity);
10. cognitive and abstract contextual information such as an individual’s emotional responses, intuition, feelings, and sensibilities expressed in terms of semantic terminologies.
Context-aware systems have been developed in recent years to improve user interaction in applications and services which operate in highly dynamic environments which are characterized by adaptable contextual situations (Hong et al., Reference Hong, Suh and Kim2009). Contextual information (captured context data processed into information useful in CP) is fundamental in the development of context-aware applications where improved interactions between computers, devices, and people in LSDS is a system requirement. In the modeling of context the requirements for dynamic environments include scalability, technology, mobility, tolerance, robust systems, security, CS, and decision support. When dealing with context, context-aware systems and applications may run multiple computing devices and sensors that define and describe the evolving states that characterize the dynamic nature of context. CP as proposed in this article has the capability to enhance a context-aware systems performance when sharing data between services and clients in LSDS.
Context modeling is concerned with the modeling of data (on a conceptual level) and includes the metadata and the contextual information (the context properties and their literal values) that describes and defines a context where data are captured from a range of sources. A good context-modeling formalism should (1) be human and machine readable (Moore et al., Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Wan2010c) and (2) reduce the complexity (of a context-aware system) with the aim of improving performance.
Contextual data are (generally) gathered by a set of autonomous sensors. When gathering data the process of context data generation consists of filtering, aggregation, fusion (CM as discussed in this article extends the concept of data fusion), and may additionally include inference and reasoning. Creating a context requires contextual information which describes the resources available by the computing infrastructure and an entities (generally an individual’s or an individual’s mobile device) context. In intelligent context-aware systems context middleware is used to detect and adapt to the changing context(s) and implement PSP. Thus, context-aware systems should facilitate the adaptation of the behavior among services, accommodate dynamic environments, and enable CS. Context modeling should also provide an effective basis upon which inference and reasoning (generally using subsumption and entailment) can be implemented in complex situations.
In semantic intelligent context-aware systems, contexts are dynamically influenced by user intuitions and preferences. An appropriate method called kansei engineering has been developed to deal with human feelings, demands, and impressions of context-aware applications. Kansei is a Japanese term meaning sensibility, impression, and emotion (Salem et al., Reference Salem, Nakatsu and Rauterberg2009). Kansei words are given by adjectives describing human emotion (more accurately emotional response), sensibility, and impression; there is no equivalent term in English, the nearest applicable word is possibly intuition. Kansei evaluation is commonly used for evaluation methods to quantify impressions.
For kansei evaluation, we have determined adjective pairs called kansei words in pairs: (Synonym and Antonym) and (Synonym and Not Synonym). For instance, the pairs of adjectives (good and bad) and (successful and unsuccessful) are kansei words. This approach is discussed in Pham et al. (Reference Pham, Thang, Nakaoka, Cooper and Kamei2011) where kansei words are used in quantifying trader sensibilities about trading decisions, market conditions with uncertain risks in a context-aware-based group decision making. By aggregating user preferences and selecting alternatives, a group of individuals enhance potentially optimal solutions in contextual information. The use of kansei evaluation shows that if data can be captured, measured, codified, and digitized it can be considered to be contextual information.
The range of contextual information identified along with the dynamic nature of the data demonstrates the inherent complexity of context, its domain-specific nature, and the difficulty in defining and measuring it. This difficulty is exemplified in the need to accommodate two general types of context: a static context and a dynamic context (Moore & Hu, Reference Moore and Hu2007; Moore et al., Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Wan2010c). A static context relates to a use-case in which a user has an element of control in context creation; a dynamic context relates to a condition in which the user is seen as being passive, or at least somewhat less in control; in such a use-case the system monitors, analyses, and reacts dynamically to a user’s behavior and ‘state’. The two types of context are reflected in the two principal ways context is used, these are: (1) as a retrieval indicator (a static context) and (2) to tailor system behavior to match users system usage patterns (a dynamic context). The static and dynamic context also uses to provide relevant information, matching with users in uncertain environments.
4 Rules and rule strategies
Rules are generally conditional specifications that instruct, permit, trigger, and inhibit processes, functions, and actions. Many approaches have emerged to create RBS’s; the approaches, however, share certain key properties (Hayes-Roth, Reference Hayes-Roth1985) including (1) the capability to incorporate practical human knowledge in conditional IF-THEN rule strategies, (2) the degree of accuracy improves at a rate proportional to the growth in the size of the knowledge base, (3) the capacity to solve diverse and complex problems, and (4) the ability to determine the best execution sequence.
Rules fall into two general types: (1) reasoning and inference rules which are generally used where inference is drawn to identify conditional relationships such as familial or peer-to-peer and business-to-business relationships and (2) trigger rules which generally implement actions either in the form of a Boolean decision or the firing of another rule. Rule strategies fall into four general classifications (Moore et al., Reference Moore, Hu and Jackson2011): (1) derivation or deduction rules, (2) integrity constraints, (3) transformation rules, and (4) reaction or event condition action (ECA) rules:
1. Derivation rules: generally relate to conditional relationships. For example, IF one statement or set of statements is true THEN a second statement or set of statements is also true.
2. Integrity constraints: in general use this rule type relates to validation checking predicated on truth by negation (if it is not true then conclude it is an error).
3. Transformation rules: relate to truth in, for example, one knowledge base or ontology with truth in another. For example, in a modified form (when used in combination with integrity constraints) transformation rules have been used in the matching and merging of ontologies (Lera et al., Reference Lera, Juiz and Puigjaner2010).
4. Reaction rules: relate to the notion of 〈action〉 where the IF component evaluates the rule 〈condition〉 resulting in an 〈action〉. The outcome can be either (1) a Boolean decision or (2) the firing of another rule.
The four rule strategies are based on the IF-THEN logic structure in the form: IF 〈condition〉 THEN 〈action〉; the context-processing rules (CPR) and fuzzy event condition rules algorithm (FECA) rules strategy is based on the ECA rule strategy using the IF-THEN logic structure. In actuality, RBS generally utilize a number of rule strategies. Viewed from an intelligent context-processing perspective, the final output is a Boolean decision; this clearly calls for a trigger rule type with a reaction rule strategy. However, as discussed in Moore et al. (Reference Moore, Jackson and Hu2010d) CS requires validation using integrity constraints to address, for example, implementation of security and access control policies. Where context modeling utilizes OBCM (Moore et al., Reference Moore, Hu, Zhu, Campbell and Ratcliffe2007), transformation or derivation rules strategies may be used.
RBS using the ECA rule strategy with the IF-THEN logic structure have been employed in diverse domains (Moore et al., Reference Moore, Hu and Wan2010b) including FRBS as discussed in Berkan and Trubatch (Reference Berkan and Trubatch1997). There are five principal fuzzy rule composition strategies identified in Berkan and Trubatch (Reference Berkan and Trubatch1997): (1) competitive rules, (2) weighted rules, (3) prioritized rules, (4) hierarchical rules, and (5) adaptive rules. Given the inherent complexity of context (Moore, Reference Moore2009) and the demands of CM with CS (Moore et al., Reference Moore, Jackson and Hu2010d) no single fuzzy rule composition strategy fulfills the requirements of CM (see Section 5 below). The FECA rules strategy (Moore et al., Reference Moore, Hu and Jackson2011) employs strategic functions drawn from a number of fuzzy rule composition strategies; for example, thresholds (competitive rules) and weighting (weighted rules). It is therefore clear that to address the domain-specific demands of RBS no one rule strategy is capable of addressing all RBS, each problem requiring a domain and problem-specific rules strategy based on the use of conventional logic with fuzzy logic using fuzzy extensions to reflect the need to address cases where conventional logic is not an adequate solution.
4.1 The Semantic Web and rule-based systems
There is an ongoing debate in Semantic Web circles surrounding the use of rules (Horrocks et al., Reference Horrocks, Patel-Schneider and van-Harmelen2003, Reference Horrocks, Patel-Schneider, Boley, Tabet, Grosof and Dean2004; Sheth & Ramakrishnan, Reference Sheth and Ramakrishnan2003; Parsia et al., Reference Parsia, Sirin, Grau, Ruckhaus and Hewlett2004) and in addressing the Semantic Web Rules Language (W3C, 2012) it has been observed that ‘it is commonly acknowledged that rules are next’ (Parsia et al., Reference Parsia, Sirin, Grau, Ruckhaus and Hewlett2004). An interesting aspect of this debate is the approach to reasoning and inference, Ontology Web Language (OWL), and the OWA versus the CWA. The CWA is predicated on negation-as-failure (i.e. ‘if it is not provably true, conclude that it’s false’) (Sheth & Ramakrishnan, Reference Sheth and Ramakrishnan2003). The Semantic Web is predicated on the OWA; however, the CWA is also useful in certain applications (Horrocks et al., Reference Horrocks, Patel-Schneider and van-Harmelen2003, Reference Horrocks, Patel-Schneider, Boley, Tabet, Grosof and Dean2004; Sheth & Ramakrishnan, Reference Sheth and Ramakrishnan2003). A discussion on the topic is beyond the scope of this article; however, an exposition with extensive references can be found in Sheth and Ramakrishnan (Reference Sheth and Ramakrishnan2003) and Sheth et al. (Reference Sheth, Ramakrishnan and Thomas2005). In summary, the proposed approach implemented using the CPA is predicated on the CWA to enable effective CM and CS with predictable decision support.
4.2 Logic systems and context
Conventional logic is generally characterized by notions based on a clear numerical bound (the crisp case) (Klir & Yuan, Reference Klir and Yuan1995) in which an element is defined as a member of a set based on numerical parameters in the discrete range {1,0}. Traditionally, logic (also referred to as Boolean logic) consists of three elements (Berkan & Trubatch, Reference Berkan and Trubatch1997): (1) truth values, (2) linguistic connectors, and (3) reasoning types. In Boolean logic truth values are in the discrete range {1,0} (truth or false, respectively).
In fuzzy logic truth is a matter of degree with truth values in the continuous range [1,0] (Klir & Yuan, Reference Klir and Yuan1995; Berkan & Trubatch, Reference Berkan and Trubatch1997). Fuzzy set theory enables a variable measure of membership of a set (expressed in terms of degrees of membership) (Klir & Yuan, Reference Klir and Yuan1995; Berkan & Trubatch, Reference Berkan and Trubatch1997) defined using a fuzzy distribution function (often termed a membership function in the literature) (Berkan & Trubatch, Reference Berkan and Trubatch1997) based on normalized values in the continuous range [1,0]. These mapping assumptions along with the related approach to defuzzification as discussed in Berkan and Trubatch (Reference Berkan and Trubatch1997), Moore and Pham (Reference Moore and Pham2012), and subsequent sections of this article are central to the FECA rules strategy.
To illustrate the need for variable degrees of membership of a set in the posited approach consider CM where the likelihood of a perfect match between the input and corresponding output context property values has a low probability (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Jackson2011; Moore & Pham, Reference Moore and Pham2012). This introduces the need for PM (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Jackson2011) where a matched context mapped to a normalized value of, for example {0.80}, has a defined degree of membership. CM with PM imposes issues similar to those encountered in decision support under uncertainty, which is possibly the most important category of decision problem (Shackle, Reference Shackle1961) and represents a fundamental issue for decision support.
Uncertainty is associated with vagueness (Klir & Yuan, Reference Klir and Yuan1995) and is considered in relation to FRBS by Berkan and Trubatch (Reference Berkan and Trubatch1997) where fuzzy logic, as it applies to decision support, is discussed. Space limits a discussion on the topic, however, a detailed exploration of fuzzy sets and fuzzy logic can be found in Klir and Yuan (Reference Klir and Yuan1995); a comprehensive discussion on fuzzy system design principles can be found in Berkan and Trubatch (Reference Berkan and Trubatch1997), where a number of classes of decision problem are identified and discussed. In summary, the CPA which is predicated on an FRBS, a strategy has been shown to provide the ability to arrive at decisions under uncertainty with high levels of predictability (Moore et al., Reference Moore, Hu and Jackson2011; Moore & Pham, Reference Moore and Pham2012).
4.3 Rule-based systems design
The use of an RBS raises questions in two (general) areas: (1) RBS design when viewed from a parametric design perspective with respect to context-aware systems and (2) the use of a rules engine—arguably a similar question with a different emphasis; however, the questions raised extend the RBS design parameters.
The use of a rule-based approach clearly involves the design and development of a suitable rule set serialized in a machine (and ideally human) readable formalism. In addition, an RBS is domain specific requiring domain-specific design where the primary considerations are the domain of interest and the problem being addressed; the domain-specific nature of RBS is shared with context-aware systems and can be summarized as form-follows-function. From a design perspective a significant potential issue for rule-based solutions is the temptation to apply an inappropriate solution to a problem. To address this issue the question Is a rule-based solution a suitable option? must be answered.
To attempt to provide an answer to this question the following design parameters have been identified to test the suitability of an RBS approach: (1) Is human problem-solving knowledge being replicated in the solution? (2) Is the problem-solving knowledge heuristic in nature? (3) Does the knowledge (or expertise) periodically change? (4) If expertise is involved, is the expertise fairly well understood and accepted? (5) Is the problem well understood? (6) Are the input data always complete and correct? (7) Can the problem be (better) solved by conventional database or programming solutions? (8) Does it pass the telephone test? (this test asks whether an expert, by speaking with someone over the telephone, can gather sufficient information to solve the problem). (9) Is the decision-making process merely the matching of data? (or are rules applied in the matching and decision-making process). (10) Is speed of operation a prime requisite for the system? (an RBS is arguably slower at runtime than conventional programming solutions). (11) Are partial solutions acceptable in the identification of individuals?
The design parameters and issues 1–8 have been derived from Gonzales and Dankel (Reference Gonzales and Dankel1993), parameters 9–11 representing three additional parameters which are relevant to intelligent context-aware systems (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Jackson2011).
4.4 Rule-based systems and context processing
The use of a rule-based approach clearly involves the creation of rules and a suitable data structure. In the posited approach the rules are encoded in Java; the rules updating and accessing a Jena TDB data set using SPARQL and SPARQL Update (SPARQL, 2007; Jena, 2012; SPARQL Update, 2012; W3C, 2012). The data structure is a semantic context-modeling ontology created using OWL DL (W3C, 2012) based on OBCM as discussed in Moore et al. (Reference Moore, Hu, Zhu, Campbell and Ratcliffe2007). Space restricts consideration of the semantic context-modeling ontology, however, the ontology is presented with a discussion, illustrative implementations, and proof-of-concept in Moore et al. (Reference Moore, Hu and Wan2010b). In summary, to define the context properties and the related data values the RDF/S triple structure 〈S〉 〈P〉 〈O〉 (the subject, predicate, and object (the literal value), respectively) is used. RDF/S and OWL serialization can be cryptic and is not particularly human friendly, however, the availability of development tools such as Protege 3.4 and 4.1 (Protege, 2012) which automatically creates the RDF/XML serialization generally resolves this issue.
The design parameters 1–8 have been derived from Gonzales and Dankel (Reference Gonzales and Dankel1993); parameters 9–11 representing three further design considerations relevant to the CM problem under investigation. A comprehensive discussion on the design and engineering of knowledge-based systems can be found in Gonzales and Dankel (Reference Gonzales and Dankel1993), however, in summary, a positive response to parameters (1, 2, 3, 4, 5, 8, 11) indicates that an RBS does potentially provide an effective solution. A positive response to parameters (6, 7, 9, 10), however, implies that a rule-based approach is not potentially a viable solution. Applying the design parameters to the CM problem points to the conclusion that an RBS does provide a viable and effective basis for realizing a context-aware decision support in an intelligent context-aware system.
5 Context processing
Context processing utilizes the context-processing algorithm that implements two algorithms: (1) the context-processing rules and (2) the fuzzy event condition rules algorithm, which is an extension of the CPR algorithm. The CPR and FECA algorithms (Moore et al., Reference Moore, Hu and Wan2010c, Reference Moore, Hu and Jackson2011; Moore & Pham, Reference Moore and Pham2012) are designed to implement CM with PM and CS. Figure 1 graphically models the CM problem; shown is the input and output context(s) with the PM problem and the implementation of the membership function (the decision threshold) the FECA algorithm is designed to facilitate (Moore et al., Reference Moore, Hu and Jackson2011); see Section 8 (below) where there is an overview of the design for the membership function. In certain cases, for example, context-aware health-monitoring systems as discussed in Moore et al. (Reference Moore, Jackson and Hu2010d) and Ashford et al. (Reference Ashford, Moore, Hu, Jackson and Wan2010), the proposed approach has the capability to implement multiple decision thresholds (Moore & Pham, Reference Moore and Pham2012) where there are number of possible decisions relating to a patient’s medical condition and prognosis.
5.1 The context-processing algorithm
The CPR implemented in the CPR and FECA algorithms are predicated on three elements: (1) event, (2) condition, and (3) action. The basic structure of the rules is {ON 〈event〉 IF 〈condition〉 THEN 〈action〉}. An event relates to an input which can be either (1) the 〈event〉 and the context data or (2) the output resulting from the firing of a rule (which triggers a subsequent rule). A condition (a rule antecedent/body) relates to the evaluation of the context properties in which comparisons are drawn between the input property(s) and the output context properties (a potential solution properties and related property values derived from the data structure) as discussed in the context-matching algorithm. An action (a rule consequent/head) is the result derived from the processing of the rule condition(s) (the rule antecedent/body) and results in either (1) the triggering of a related simple rule, a complex or composite rule, or another rule set or (2) a Boolean decision based on CM arrived at using CP. In (1) there is a pre-defined sequence incorporating built in stopping criteria setting out termination and/or CS criteria. In (2) the decision is based on the relationship between the context match and a decision threshold.
5.1.1 The context-processing rules algorithm
Using the illustrative context properties a1, b1, b2, c1, c2 modeled in Figure 1 the CPR algorithm is as follows where:
∙ e is the numerical evaluation of an individual property match {1,0} (true/false, respectively);
∙ n is the sum of the individual property evaluation matches;
∙ x is the value for a perfect match; that is, if all context matches are true and the computation equates to a normalized value of [1];
∙ m is the normalized value for the context match: (m=n/x);
∙ t is the preset threshold value in the range [0…1].
Applying these functions the algorithm process is as follows:
∙ Step 1: evaluate the context match (e) for each individual context property [1,0] (true, false).
∙ Step 2: sum the values for (n) derived from the matching process for each individual property (e).
∙ Step 3: calculate the value (x) for a perfect match.
∙ Step 4:calculate the normalized value for (m) in the range [0…1].
∙ Step 5: using the preset threshold (t) value (in the range [0…1]) determine if the output (potential solution) context is a suitably qualified match with the input (problem) context.
5.1.2 Evaluation and proof-of-concept
An example implementation of the CPR algorithm where: e(a1)=[1], e(b1)=[0], e(b2)=[1], e(c3)=[1], e(c4)=[0] and given that: x=[5] results in a normalized value for m of {0.60}. As (m) is less than the preset threshold value (t=0.75) the context match is false and the individual is therefore not a suitably qualified individual.
5.2 The fuzzy event condition action rules algorithm
The FECA rules algorithm extends the RDF/S triple structure (used in the CPR algorithm) with a weighting 〈W〉 component to give a triple plus comprising 〈S〉 〈P〉 〈O〉 〈W〉. The FECA rules algorithm is predicated on a quantitative approach and is based on fuzzy logic using degrees of membership of a set as discussed in Moore et al. (Reference Moore, Hu and Jackson2011) and Moore and Pham (Reference Moore and Pham2012) and Section 4.2 (above). In summary, the 〈W〉 component in the numerical range [0.01…1.00] is applied to each context property to implement a prioritizing bias into the CP to accommodate CS (or at least mitigate CS violations).
5.2.1 Implementation
The FECA rules algorithm is set out below with an example implementation to demonstrate CP and provide proof-of-concept. Using the illustrative context properties {a1, b1, b2, c1, c2} modeled in Figure 1 the FECA rules algorithm is as follows:
∙ w [0.10…1.00] is a weight applied to each context property to reflect its relative priority.
∙ e {1,0} is the numerical (Boolean) evaluation for each context property match true=[1] or false [0].
∙ av (e×w) is the computation of the actual value for each context property match evaluation in the range [0.00…1.00].
∙ sav (av (a1…an)) is the sum of the av.
∙ mpv (av (a1…an)) is the maximum potential value. This represents a state in which all context property matches (av (a1…an)) are true [1]. The mpv assumes that a perfect context match has been identified.
∙ rv (sav/mpv) is the resultant value. This value represents the degree to which the overall context match is true in the range [0.00…1.00]. The (rv) value is used in step 7 to test if (rv) is greater than (t).
∙ t is the value for the decision boundary (threshold) in the range [0.10…1.00]. The (t) is used in step 7 to test if (rv) is greater than (t). Note: the posited approach enables multiple thresholds (t).
Applying these functions the algorithm process is as follows:
∙ Step 1: evaluate the context match [1,0] for each individual context property, for example: (IF (a1(input)).equalTo(a1(output)) THEN e=[1] or IF (a1(input)).notEqualTo(a1(output)) THEN e=0).
∙ Step 2: obtain the pre-defined property weighting (w) in the range [0.1…1.0] for each context property.
∙ Step 3: apply the weighting (w) to the value as derived from (2): (IF e(a1)={1,0} THEN av=(e×w)).
∙ Step 4: sum the values derived from the matching process: (sav=(av (a1)+av (b1)+av (b2)+av (c1)+av (c2))).
∙ Step 5: compute the mpv for the CM process: (mpv=(w(a1)+w(b1)+w(b2)+w(c1)+w(c2))).
∙ Step 6: compute the rv for testing against threshold value (t): (rv=(sav/mpv)).
∙ Step 7: using the preset threshold value (t) determine if the output (potential solution) context definition is a suitably qualified match with the input context: IF (rv) is greater than or equal to (t) THEN context match=true [1] or IF (rv) is less than (t) THEN context match=false [0].
5.2.2 Evaluation and proof-of-concept
The following illustrative implementation of the FECA rules algorithm demonstrates the CM process. Applying the algorithm using the context properties modeled in Figure 1 and the following example values for (w), (e), and (t) results in a computation as follows:
∙ the values for (w) are (w(a1)=0.7, w(b1)=0.8, w(b2)=0.7, w(c1)=1.0, w(c2)=0.9);
∙ the values for (e) are (e(a1)=[1], e(b1)=[0], e(b2)=[1], e(c1)=[1], e(c2)=[1]);
∙ the threshold value (t) is {0.75}.
Using these values the CM process is as follows: sav=((av(a1)=([1]×0.7)=[0.70])+(av(b1)=([0]×0.8)=[0.00])+(av(b2)=([1]×0.7)=[0.70])+(av(c1)=([1]×1.0)=[1.00])+(av(c2)=([1]×0.9)={0.9})). Thus, sav=3.30.
mpv=(av (a1…an))=4.10.
rv=(sav/mpv), rv=(3.30/4.10), rv=0.80.
As (rv) is 0.80 and t=0.75 (rv) is greater that (t). The result of the CM is therefore true [1] and the individual is a suitably qualified individual.
An analysis of the implementation supports the following conclusions:
∙ The evaluation of the context properties (step 1) and their respective values will initially result in a Boolean result in the range (e)={1,0}.
∙ The Boolean value for truth [1] or falsity [0] derived for (e) (step 1) is retained when the weight 〈W〉 applicable to a context property (w) is applied (e×w) (step 3).
∙ The implementation of a prioritizing bias to reflect the relative importance of specific context properties is achieved.
∙ CM with PM and CS is realized using the FECA rules strategy.
In actuality, implementation of the proposed approach employs a four-stage process. The use of the 〈W〉 parameter is utilized for context properties in stages 2 and 3 of the CM process:
∙ Stage 1: the evaluation of control properties to validate the inputs, apply error checking, apply relevant constraints and policies, and implement rule assignment as discussed in Berndtsson and Lings (Reference Berndtsson and Lings1995) and Moore et al. (Reference Moore, Hu and Wan2010b).
∙ Stage 2: an evaluation of the input context properties with the corresponding output context properties (CM) to identify if an output context (an instance of an entity where such an instance is a member of a set of entities) is a suitably qualified recipient for personalization.
∙ Stage 3: the application of the context properties that represent constraints and the implementation of priorities to avoid (or at the very least mitigate) violations of CS.
∙ Stage 4: the implementation of the decision; there are two courses of action, either (a) the delivery of the service if the CM is true (equates to [1]) or (b) in the event of the CM being false (equates to [0]) the system terminates.
5.3 The context-processing rules and fuzzy event condition rules algorithm strategy
The CPR and FECA rules strategies are predicated on the ECA rule strategy (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Jackson2011; Moore & Pham, Reference Moore and Pham2012) in the form: {ON 〈event〉 IF 〈condition〉 THEN 〈action〉}. The 〈event〉 rule component implements either (1) a system input event such as a location change or login event or (2) an event resulting from the firing of a previous rule. The 〈condition〉 (or antecedent) component of the rule employs a Boolean structure similar to Horn style rules in a simple rule structure using a single context property with input and output literal values. In operation, the condition rule component executes a conditional comparison of the input and output literal values to arrive at a Boolean result: the truth [1] or alternatively the falsity [0].
The 〈action〉 (or consequent) rule component reacts to the result of the 〈condition〉 rule component, the output being either a Boolean decision or the firing of a subsequent rule or rule set. The condition component of an ECA rule (the antecedent) employs the IF-THEN rule strategy commonly found in reaction rules (Berndtsson & Lings, Reference Berndtsson and Lings1995) which relates to the notion of 〈action〉 where the IF component evaluates the rule 〈condition〉 resulting in an 〈action〉. The 〈action〉 in the proposed approach can be either (1) a Boolean decision or (2) the firing of another rule.
To address the PM issue the posited approach applies the principles identified in fuzzy logic and fuzzy set theory with a defined membership function (Moore & Pham, Reference Moore and Pham2012) as discussed in Section 4.2 (above). Conventional logic is generally characterized using notions based on a clear numerical bound (the crisp case); that is, an element is (or alternatively is not) defined as a member of a set in binary terms according to a bivalent condition expressed as numerical parameters {1,0}. Fuzzy set theory enables a continuous measure of membership of a set based on normalized values in the range [0,1].
These mapping assumptions are central to the CM algorithm (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Jackson2011; Moore & Pham, Reference Moore and Pham2012). A system becomes a fuzzy system when its operations are ‘entirely or partly governed by fuzzy logic or are based on fuzzy sets’ and ‘once fuzziness is characterized at reasonable level, fuzzy systems can perform well within an expected precision range’ (Berkan & Trubatch, Reference Berkan and Trubatch1997). Consider a use-case where a matched context mapped to a normalized value of, for example [0.80], has a defined degree of membership. This measure, while interesting, is not in itself useful when used in a decision-support system (Moore & Pham, Reference Moore and Pham2012); in the CPA this is addressed using a distribution function, more generally referred to in the literature as a membership function (Berkan & Trubatch, Reference Berkan and Trubatch1997) to implement the essential process of defuzzification as discussed in Moore et al. (Reference Moore, Hu and Jackson2011), Moore and Pham (Reference Moore and Pham2012), and Berkan and Trubatch (Reference Berkan and Trubatch1997).
6 Context-processing and large-scale distributed systems
As discussed in Section 2 (above) the computing landscape is frequently characterized by LSDS; such systems are dynamic, complex, inherently context aware, and generally function under uncertainty. Addressing these often conflicting characteristics places high demands on context middleware and the processing of data derived from diverse local and distributed sources including ‘mashups’ (Lopez et al., Reference Lopez, Bellas, Alberto Pan and Montoto2011) and cloud-based systems (Rimal et al., Reference Rimal, C. and Lumb2009). The posited approach to CP has been developed as a generic system capable of processing contextual information from a diverse range of distributed data sources with implementation in a broad range of domains, systems, and technologies. As such, the CPA incorporates the ability to manage the inherent complexity that characterizes context, context-aware systems, and LSDS. In addition, the data structure employed in the posited approach and implemented using the Jena Java API (Jena, 2012) incorporates the ability to handle both in memory and persistent data storage along with updating using distributed data processing employing local processing in, for example, a mobile device using a thin client and centralized processing in a server.
Decentralized systems will increasingly feature local processing of captured data into information useful in CP. For example, (1) there may be a preprocessing stage in the overall processing of data captured from a sensor(s) in which processing capacity may be incorporated into a sensor(s) (Thomas et al., Reference Thomas, Shah, Moore, Rayson, Wilcox, Osman, Evans, Chapman, Athwal, While, Pham and Mount2012) before submission to the context middleware or (2) in an assisted living environment (in, e.g., a health-monitoring system; Ashford et al., Reference Ashford, Moore, Hu, Jackson and Wan2010; Moore & Pham, Reference Moore and Pham2012) local processing of data in an intelligent context-aware decision-support system using a combination of sensors and mobile technologies in a PAN may be employed. Space precludes a detailed discussion on the topic; however, the relative merits and demerits, while complex, are well covered in the literature and range from technical considerations (Thomas et al., Reference Thomas, Shah, Moore, Rayson, Wilcox, Osman, Evans, Chapman, Athwal, While, Pham and Mount2012) to cryptography using digital aggregate signature schemes with bilinear mapping (Boneh et al., Reference Boneh, Gentry, Lynn and Shacham2003a, Reference Boneh, Mironov and Shoup2003b).
In summary, the design issues are twofold: (1) data storage and (2) context processing. For general location and identity-based systems a distributed approach using a thin client solution with local data storage has potential advantages. The increased complexity of intelligent context-aware systems, however, presents increased data storage and processing demands with related computational overhead and tractability issues. These issues, when taken with the multiple data capture modes in LSDS, wireless ad hoc networks, and mobile systems present significant challenges in the area of data security. While the centralized server approach may offer greater potential security the growing ubiquity of wireless mobile systems demands a domain-specific approach to implementation.
Saha and Mukherjee (Reference Saha and Mukherjee2003) in addressing pervasive computing have noted that all the basic component technologies exist today in hardware, we have mobile devices, sensors, and even smart appliances. Thomas et al. (Reference Thomas, Shah, Moore, Rayson, Wilcox, Osman, Evans, Chapman, Athwal, While, Pham and Mount2012) concur observing components such as sensors, wireless mesh architectures, cloud services, and data brokerage/processing are all currently available and widely researched. The challenges lie in the development of context middleware; the posited approach incorporates the ability to manage these challenges in LSDS.
7 Discussion
This article has considered personalization, PSP, and context with consideration of the domains and systems to which context has been applied and the nature of the contextual data. An overview of the evolving computing landscape has been presented with its relationship to context. Rules and rule strategies have been discussed with logic systems and an overview of the OWA versus the CWA and the relationship with the Semantic Web. The design parameters for RBS have been considered; the discussion supports the conclusion that an RBS represents a suitable solution to enable effective context processing. The event-driven rule-based approach (which forms the basis upon which intelligent context processing can be realized) has been presented with an evaluation and proof-of-concept.
It is posited that the proposed approach has the capability to implement intelligent CP with predictable decision support. In addition, the reported results in Moore et al. (Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Jackson2011) and Moore and Pham (Reference Moore and Pham2012) when taken with the evaluation presented in this article supports the conclusion that the posited approach provides an effective and predictable basis upon which CS can be accommodated (or at least the potential violations of CS can be mitigated). A discussion on the CPA, logic systems, and rule strategies with the related conditional relationships for intelligent context-aware systems can be found in Moore et al. (Reference Moore, Hu and Jackson2011) and Moore and Pham (Reference Moore and Pham2012). For an in-depth discussion on CM and PM with example implementations and a data set evaluation, see Moore (Reference Moore2009) and Moore et al. (Reference Moore, Hu and Wan2010b).
While the research has resolved many crucial issues, a number of outstanding issues and challenges have been identified (Moore et al., Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Wan2010c, Reference Moore, Hu and Jackson2011; Moore & Pham, Reference Moore and Pham2012) including (1) the design of the membership function, (2) data capture and data description, (3) persistent storage, (4) centralized versus distributed systems, (5) the approach to the setting of the 〈W〉 parameter, and (6) alternative approaches to context processing.
Item 4 has been discussed in Sections 2 and 6 (above); the conclusions drawn are that the posited approach provides an effective basis upon which the inherent complexity and data-intensive nature of LSDS can be accommodated. Items 2 and 3 have been addressed in Moore et al. (Reference Moore, Hu and Jackson2010a, Reference Moore, Hu and Wan2010b, Reference Moore, Hu and Jackson2011). Items 1, 5, and 6 are considered in the following sections.
7.1 The design of the membership function
The design of a membership function is domain and application specific (Berkan & Trubatch, Reference Berkan and Trubatch1997; Moore & Pham, Reference Moore and Pham2012). The membership function currently uses a decision boundary (or threshold) solution to enable defuzzification in CM as discussed in Sections 4.2 and 5 (above). This approach, while effective, is recognized as a suboptimal solution. Current research is investigating (1) using semantic representation and (2) how heuristics are applied using the available a priori knowledge related to the domain of interest and the problem being addressed. For a comprehensive discussion on the design of the membership function with consideration of the relative impact of points (1) and (2) on the location of the decision boundary(s) (t1 t2), see Moore and Pham (Reference Moore and Pham2012). Figure 2 graphically models the CM and PM (the degree of membership as discussed in Section 5 (above)) mapped onto the solution space (shown as the CM (X) axis).
A conceptual model of the context matching (CM) mapped onto the solution space. Shown are the decision boundaries (thresholds) t1 and t2 used in the defuzzification process as discussed under the design of the membership function and in Sections 4.2 and 5
Points (1) and (2) represent important design parameters for a membership function. These parameters when taken with the relative contribution each context property (in a set of context properties that combine to create a context) makes to the overall context definition (Moore & Pham, Reference Moore and Pham2012) are being used in ongoing research to attempt clarify the approach to the identification for the optimal values for the decision boundaries. The results derived from the research will be analyzed to attempt to identify the optimal solution for the membership function design generalizable to a range of domains and applications. The identification of the optimal solution represents an open research question.
7.2 Setting the weight value for the 〈W〉 component
Currently, setting the literal value for the 〈W〉 parameter value is set manually (to reflect the domain of interest and the problem being addressed). Ideally, the setting of the 〈W〉 value would be initially set manually using default values and updated dynamically as the context-aware system gathers information on users’ system usage and changing context. However, a potential side effect of dynamic updating for the 〈W〉 parameter is the induction of unknown and emergent dynamic change to the preset design parameters and policy priorities; this may impact the ability to enable CS with predictability and consistency. Identifying solutions to the dynamic updating issue represents an open research question.
7.3 Alternative approaches to context processing
The current approach proposed in the article can be viewed in terms of a deterministic search for a suitably qualified user (u) who is a member of a set of users (U). This process is essentially a linear ‘brute-force’ search of the hypothesis space (U), to improve the efficiency of context processing a number of alternative approaches have been investigated including Bayesian methods to enable inferred conclusions to be drawn (i.e. the Boolean truth) as to a user(s) (u) suitability for personalization.
7.3.1 Bayes theorem and the context-matching problem
Bayesian reasoning provides a probabilistic approach to inference. However, a practical issue (when using Bayesian methods) is that they typically require knowledge of prior probabilities. When these probabilities are not known they are (generally) estimated based on empirical background information, previously available data, and assumptions about the underlying distributions. A detailed discussion on Bayes theorem and its fundamental principles with examples and applications can be found in Mitchell (Reference Mitchell1997). In summary, Bayes theorem applies a priori probabilities as set out below. In actuality, Bayes theorem is quite general and can be applied to any set of mutually exclusive hypothesis (h) in a set of hypotheses (H) whose probabilities sum to 1.
∙ P(h) denotes the initial probability that hypothesis (h) holds. P(h) is generally termed the prior probability of (h). In problems where no prior knowledge of probabilities exists the same probability may be assigned to each potential hypothesis (h).
∙ P(D) denotes the prior probability that the training data will be observed independent of knowledge relating to (H) where (H) is a set of hypothesis (h).
∙ P(D−h) denotes the probability that (D) will be observed given some world in which (h) holds.
∙ P(h−D) denotes that (h) holds given (D) and is generally termed the posterior probability of (h).
Intuitively, there is an observed similarity between Bayes theorem hypothesis where (h) is a member of a set of hypotheses (H) and the CM problem where a user (u) is a member of a set of users (U). An analysis of Bayesian methods when applied to the CM problem has, however, identified a number of issues. Before addressing these issues consider the following relationships: (1) an instance of an individual (u) in the CM problem generally correlates with an hypothesis (h) in the Bayes theorem and (2) the input context data (C) in the CM problem generally equates to the data (D) in the Bayes theorem. Given these relationships and that according to Bayes theorem the sum of the probabilities must sum to [1]; consider the following.
In the CM problem the input data (C) (which describes, e.g., a resource being distributed or a requested collaboration query) will always hold to be true independent of (u) as the input data (C) will in all cases be correct. The prior probability P(C) is therefore estimated to be universally [1]. All output data (context properties that describe an individual(s) (u) context) will always hold to be true independent of (C) as the data for (u) will in all cases be correct. The prior probability P(u) for (u) is therefore estimated to be universally 1. In the CM problem the probability of a true (or alternatively a false) context match P(u−C) is unknown before CP given the highly dynamic and variable nature of the potential inputs (C) and output(s) (u). Thus, estimating a prior probability for (u) is not realistic. However, given these assumptions, it may be argued that there is an equal probability of an individual (u) being a true or alternatively a false match. Therefore, a prior probability is estimated to be 0.5 for each potential user (u) universally as set out in Bayes theorem. In considering the application of these estimated probabilities, a number of conclusions have been drawn.
As the probabilities are applied universally to every context property the overall effects on context processing for each context property, and therefore the resulting CM, will universally be nil. There are observed negative side effects on the context-matching process in that the application of the estimated probabilities will intuitively reduce the normalized result for the context match universally to a value lower than the value for the threshold (t). Addressing this issue will require additional steps in the proposed algorithm with respect to the value for (t). There will also be an unknown effect on the 〈W〉 parameter with potential consequences resulting in potential compromises in the ability to implement priorities and therefore enable effective CS.
The issues identified above in relation to probabilities in the application of Bayes theorem call into question the potential for Bayesian methods to address the demands of CP and the CM problem. An additional issue for Bayesian methods is the potential for significant computational overhead with the resulting tractability issues. Notwithstanding the factors discussed, given the observed similarities context processing using Bayesian methods remains an interesting direction for future research.
7.3.2 Inference and reasoning in context matching
The proposed approach, as currently constituted, does not utilize inference and reasoning, however, the semantic context-modeling ontology created using the OBCM approach using OWL DL along with the posited rule strategy has been designed to incorporate the ability to implement inference and reasoning in CP where the domain-specific design requirements call for such an approach. The proposed rule strategy uses logic functions such as AND, OR, and NOT in simple and complex rule structures where the 〈condition〉 may potentially incorporate multiple statements. Alternatively, multiple simple rules employing a Boolean structure similar to Horn style rules may be used. This capability and the portability of the proposed approach enables implementation in a range of rule engines using differing rule strategies. For example, the rule strategy adopted may be implemented using Java and Common Lisp in the Jess rule engine for the Java platform (Jess, 2010). These observations support the conclusion that inference and reasoning can be achieved based on the proposed approach where the domain requires it.
8 Conclusion and future work
While the research has addressed the requirements specification and the original design goals as discussed in Moore et al. (Reference Moore, Hu and Wan2010b), it has also identified a number of outstanding challenges and issues along with suboptimal solutions which remain open research questions as briefly discussed in the discussion. The approaches considered are thought to represent potentially useful directions for future research either singly or in hybrid solutions where personalization and PSP is required.
The difficulties in addressing the challenges and open research questions are not underestimated, however, resolving the challenges may provide an effective basis upon which intelligent context processing can be implemented with greater effectiveness and increased levels of computational intelligence.
