Elsevier

Expert Systems with Applications

Volume 38, Issue 9, September 2011, Pages 10959-10965
Expert Systems with Applications

Online knowledge validation with prudence analysis in a document management application

https://doi.org/10.1016/j.eswa.2011.02.139Get rights and content

Abstract

Prudence analysis (PA) is a relatively new, practical and highly innovative approach to solving the problem of brittleness in knowledge based system (KBS) development. PA is essentially an online validation approach where as each situation or case is presented to the KBS for inferencing the result is simultaneously validated. Therefore, instead of the system simply providing a conclusion, it also provides a warning when the validation fails. Previous studies have shown that a modification to multiple classification ripple-down rules (MCRDR) referred to as rated MCRDR (RM) has been able to achieve strong and flexible results in simulated domains with artificial data sets. This paper presents a study into the effectiveness of RM in an eHealth document monitoring and classification domain using human expertise. Additionally, this paper also investigates what affect PA has when the KBS developer relied entirely on the warnings for maintenance. Results indicate that the system is surprisingly robust even when warning accuracy is allowed to drop quite low. This study of a previously little touched area provides a strong indication of the potential for future knowledge based system development.

Highlights

► We study prudence analysis (PA) as a method for online verification and validation. ► This study investigates the viability of prudence analysis in a real world domain. ► This paper also studies the affect on a KB when the system is trusted by the expert. ► Results indicate the system is robust even when warning accuracy is allowed to drop. ► This study provides a strong indication that PA is a viable approach to V&V.

Introduction

Knowledge based researchers have long battled with the issue of brittleness, which has either directly or indirectly led to the majority of methodological developments in symbolic based reasoning. Yet it is an issue that still remains in varying degrees in most systems today. In knowledge based systems (KBS) brittleness occurs when a system is asked to inference a situation that is beyond the knowledge captured in the knowledge base. This brittleness is particularly problematic in KBSs because the system does not display an error or crash when it occurs. To the user there is generally no discernable problem as the system still produces a response. Therefore, if the user is not sufficiently knowledgeable to be able to notice the response is flawed then they may treat it as correct. For instance, a nursing assistant may not challenge a response and thus not seek an expert’s opinion, resulting in an incorrect treatment of a patient.

A brittle inference tends to occur when a request from outside the systems knowledge domain occurs or the domain of operation is missing knowledge. The cause of such inadequacies is often seen as being due to the concentration of specialised knowledge in the target domain for the particular system (1991). The majority of research into resolving brittleness could be grouped into one of three areas:

  • Finding deeper levels of knowledge. For instance, methodologies, such as Knowledge Acquisition and Design Structuring (KADS) (Wielinga, Schreiber, & Breuker, 1992), were developed to help extract deeper forms of knowledge.

  • Forming a layer of general knowledge to use when the specialised knowledge was inadequate, such as Cyc (Lenat, 1995, Matuszek et al., 2006). This provides the opportunity to fall on levels of general knowledge when the domain specific knowledge falls short.

  • Measuring the completeness of a knowledge base through methods of verification and validation (V&V) (Preece, 2001).

All three approaches have, however, provided significant challenges. For instance, methods of finding deep knowledge still do not tell us when we have it all. At some point the system must be made available to users and will always run the risk of missing knowledge. Additionally, if there is any shift in the domain’s knowledge the system will require significant rework. Likewise, systems such as Cyc, fail to be able to adjust to the constant changes in one of the most contextually dynamic knowledge domains – general knowledge (Dazeley & Kang, 2008c). Thirdly, V&V methods tend to either perform with only known cases, therefore not checking unknown, or check all combinations of attributes which reveals many situations the system does not need to know about.

Fundamentally, validation is attempting to identify whether all the possible cases are covered by the KBS. Alternatively in the subfield of anomaly detection a system is analysed holistically to find structural anomalies, such as redundancies, conflicts or dead ends (Kusiak, 2000). These approaches are a static analysis of the system at a moment in time, usually during development, and are not generally applicable online while the system is in use. Prudence analysis (PA) is a form of dynamic online anomaly detection which uses actual cases as they are presented to determine if knowledge from outside the knowledge base is required. Therefore, PA validates real situations as they occur by detecting when an inferenced solution to a case may be wrong. The advantage of such an approach is that a system can provide a warning system for the user indicating that an inference goes beyond its knowledge domain. Essentially, PA uses a form of meta-knowledge to validate each inference. The expectation is that the user will then check the case being inferenced with a human expert that can validate the correctness of the inference by the system.

Currently, PA has only been studied by a minority of researchers, all of whom have centred their studies on a single family of KBSs, referred to as ripple-down rules (RDR) (Compton & Jansen, 1988). The primary reason for this is that RDR is an incremental KA and maintenance methodology. It is RDR’s flexible and maintainable structure that makes it ideal for PA. Early work on PA such as WISE (Edwards, 1996, Kang, 1996), feature recognition prudence (FRP) (Edwards, 1996, Edwards et al., 1995, Edwards, Kang et al., 1995) and feature exception prudence (FEP) (Edwards, 1996, Edwards et al., 1995, Edwards, Kang et al., 1995) failed to deliver significant accuracy.

A new approach was taken by Compton, Preston, Edwards, and Kang (1996) of comparing cases with previously seen cases within context, and provided warnings if they differed in some unusual way. This simple method achieved a reasonably high level of accuracy on some datasets with significantly less false positives. This was subsequently followed by a Ph.D. thesis by Prayote (2007) which continued Compton et al.’s (1996) work by including a number of improvements, which allowed a reduction in the number of false positives. This study’s results indicate a significant reduction in false positives and more accurate warnings. One of the main problems with both of these approaches was their reliance on an attribute’s existence or absence in a case for the generation of warnings. This limits the methods ability to be applied primarily in domains with a controlled number of only relevant attributes. Domains with large amounts of irrelevant attributes such as free text classification will tend to produce a large amount of false positives.

The most recent study by Dazeley and Kang, 2008b, Dazeley and Kang, 2008d tried a new dynamic approach referred to as Rated MCRDR (RM) (Dazeley and Kang, 2003, Dazeley and Kang, 2009) by using a neural network approach to learning meta-knowledge. Dazeley and Kang, 2008b, Dazeley and Kang, 2008d concluded that the system was able to predict errors more accurately without increasing the false positives. More importantly though was that it contained two additional advantages over previous approaches. Firstly, it was versatile – previous approaches have a preset level of accuracy determined by the algorithms approach which cannot be altered, whereas RMs accuracy and number of false positives could be controlled. Secondly, if the system misses a case the system can still warn about similar cases in the future.

These early studies of RM however used simulated experts based purely on artificially generated knowledge, primarily using an inducted decision tree. This paper will provide details of a follow up study on this technique of PA by applying it in a domain using real data and human expertise. In this study we use an application referred to as MonClassifier which is part of a larger suite of applications called personalized web information management system (PWIMS) (Kim et al., 2004, Kim et al., 2004, Park et al., 2003, Park et al., 2004a, Park et al., 2004b). This application has been used to collect and classify a number of free text knowledge domains. For each of these domains a knowledge base has been incrementally built using human experts, such as the collection of eHealth articles used in this study. However, this process is exceptionally time-consuming, requiring the expert to read and check every article manually. This study provides an opportunity to empirically study the viability of using PA to significantly reduce the experts load, potentially opening up numerous application domains for the use of expert knowledge.

The following section discusses the PWIMS and MonClassifier applications, highlighting the fundamental problem with expert system development in such an application. It is this problem that PA is able to solve, which will be discussed in Section 3. In order to study PAs effectiveness in such a domain two experiments were performed, which are described in Section 4 with results and discussion in Section 5.

Section snippets

Application domain

The personalized web information management system (PWIMS) provides support for dynamic and personal web portals in a simple suite of applications. The platform contains three main components:

  • a Web monitoring agent

  • a storage management (or knowledge management) component

  • a knowledge sharing agent

The Web monitoring agent monitors a number of user-specified websites for newly uploaded pages. When a new page is uploaded the system retrieves the page and stores it in the database. Therefore, this

Methodology

A PA system for the MonClassifier application must be reasonable accurate – missing very few incorrect cases, while minimising the amount of incorrect warnings. The approach must also be able to operate in a free text environment where there are numerous irrelevant symbols. Therefore, the attribute based approaches of Compton et al., 1996, Prayote, 2007 would not be expected to operate well – producing too many false positives. Additionally, neither of these approaches has been used in multiple

Experiments

One of the greatest difficulties in KA and KBSs research is how to evaluate the methodologies developed (Compton, 2000). This is because any evaluation requires people to actually build the system. Furthermore, the same experts would ideally be used to compare two systems. Clearly, for this comparison to be meaningful they should also be built around the same domain. Therefore, the expert will have accrued some experience when building the first system and may provide better quality knowledge

Results and discussion

The results for the first test are shown in Table 1. In this study accuracy is calculated as TP/(TP + FN) which is the approach taken in earlier prudence work (Compton et al., 1996, Dazeley and Kang, 2008b, Dazeley and Kang, 2008d, Kang, 1996, Prayote, 2007). This approach is used because it is vastly more important to keep FN to a minimum and less important to reduce FP, which the standard method of measuring accuracy does not take into account.

These results show that RM can sustain a very high

Conclusion

Brittleness has long been a problem in knowledge based systems that has often prevented their use in a number of domains. This paper has discussed a method of online anomaly detection referred to as prudence analysis and applied this innovative approach to a real world knowledge acquisition task. In this study it was shown that such an approach can be very accurate and that on the rare case that a warning was missed that it was often picked up with a later case. This feature of the approach

References (32)

  • G. Edwards et al.

    Prudent expert systems with credentials: Managing the expertise of decision support systems

    International Journal of Biomedical Computing

    (1995)
  • Beydoun, G. (2000). Incremental knowledge acquisition for search control heuristics. Unpublished Ph.D. Thesis. Sydney:...
  • Compton, P. (2000). Simulating expertise. In Paper presented at the proceedings of the 6th pacific knowledge...
  • Compton, P., & Jansen, R. (1988). Knowledge in context: A strategy for expert system maintenance. In Paper presented at...
  • Compton, P., Edwards, G., Kang, B., Lazarus, L., Malor, R., & Menzies, T., et al. (1991). Ripple down rules:...
  • Compton, P., Kang, B. H., Preston, P., & Mulholland, M. (1993). Knowledge acquisition without analysis. In Paper...
  • Compton, P., Preston, P., Edwards, G., & Kang, B. H. (1996). Knowledge based systems that have some idea of their...
  • Dazeley, R., & Kang, B. H. (2003). Rated MCRDR: Finding non-linear relationships between classifications in MCRDR. In...
  • Dazeley, R., & Kang, B. H. (2008a). An approach for generalising symbolic knowledge. In Paper presented at the 21st...
  • Dazeley, R., & Kang, B. H. (2008b). Detecting the knowledge boundary with prudence analysis. In Paper presented at the...
  • Dazeley, R., & Kang, B. H. (2008d). The viability of prudence analysis. In Paper presented at the 2008 Pacific Rim...
  • R. Dazeley et al.

    Epistemological approach to the process of practice

    Journal of Minds and Machines

    (2008)
  • R. Dazeley et al.

    Generalising symbolic knowledge in online classification and prediction

  • Edwards, G. (1996). Reflective expert systems in clinical pathology. Unpublished M.D. Thesis. Sydney: University of New...
  • Edwards, G., Compton, P., Kang, B., Preston, P., & Lazarus, L. (1995). New paradigms for expert systems in healthcare....
  • Kang, B. H. (1996). Validating knowledge acquisition: Multiple classification ripple down rules. Unpublished Ph.D....
  • Cited by (5)

    • Using a domain expert in semi-supervised learning

      2014, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    • Run-time validation of knowledge-based systems

      2013, Proceedings of the 7th International Conference on Knowledge Capture: "Knowledge Capture in the Age of Massive Web Data", K-CAP 2013
    • Detection of CAN by ensemble classifiers based on ripple down rules

      2012, Lecture Notes in Computer Science (including subseries Lecture Notes in Artificial Intelligence and Lecture Notes in Bioinformatics)
    • Prudent fraud detection in internet banking

      2012, Proceedings - 2012 3rd Cybercrime and Trustworthy Computing Workshop, CTC 2012
    View full text