Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts




НазваниеSupport for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts
страница3/35
Дата05.10.2012
Размер0.88 Mb.
ТипДокументы
1   2   3   4   5   6   7   8   9   ...   35

1.1Motivation


Dependencies and independencies are key properties of a distribution and have an essential effect on its behaviour. In a BN, they are represented by the graph structure. It is therefore not surprising that there is general consensus that the graphical structure of a network is its most important part [26] and determines whether the model is feasible. In order to make use of the robustness of the structure, it needs to be constructed properly. Thus, it is widely accepted that it is crucial to properly understand and continually assess the graphical modelling of the domain and how it represents its structural assumptions [44, 58, 96].


The main reasons for assessing graphical modelling will be discussed in detail in chapter 2. Briefly they are:

1. The importance of the influence of the structure on the inference in terms of tractability and on the output in terms of accuracy, simplicity and explanatory power.


2. The importance of the influence of the structure on the number of parameters that need to be assessed (the process of eliciting the local distributions). The structure can make the difference between a feasible model or not, and between a maintainable model or not [26, 57].


The process of graphical modelling relies on intuitions. Yet the process of the construction of a structure is typically based on a trial-and-error process. Once changes have been made to the structure, the probabilities regarding these changes need to be elicited and only then can the model be tested using sensitivity analyses. Reviewing changes in the structure and changes to the probabilities may need to be repeated time and again, which makes it an expensive process.


Although it has been argued that conditional independency assumptions can be inferred visually from the structure of the graph [72] (p.118), one needs to be an expert in graph manipulation in order to do so. However, most DEs who wish to use the technology are not BN experts and therefore will not know how to infer conditional independencies correctly. Yet the DE who constructs a structure must have an understanding of what the BN represents.


A review of the literature indicates that most articles focus on the elicitation of the numbers, i.e. the quantitative part. By contrast, there is very little to support the process of testing independency assumptions, i.e. the qualitative part. Moreover, there is no comprehensive guide for how to elicit and model the variables and independency-assumptions and existing tools do not provide the user with practical support for this process independently of the numbers. Given the importance of the qualitative part of the model and the necessity for the DE to understand what the BN structure represents, the dearth of literature on how to elicit graphical modelling decisions and the lack of methodology of how to assess these decisions is surprising.


Although it is widely accepted that it is important to review and assess the independency assumptions that are represented by the topology of the graph, possibly by using the d-separation criterion, the question of how to (efficiently) assess the graph topology is still open (section 2.3.2). Each of the current approaches can involve one or more of the following problems:

  • The need to be a BN expert to apply the method and/or to understand its output.

  • The methods produce too much data to process.

  • Using the methods is time consuming.

As a consequence of these problems, the available methods are not appropriate for a changing iterative process such as the construction of a BN [26, 44, 58].


To overcome these problems I propose an approach that is

  • Rich – can be understood by DEs who are not BN experts.

  • Interactive – enables users to make decisions with regard to what should be investigated, which is motivated by interest and not by random relationships. This prevents a flood of random information. In addition, an interactive approach enables the DE to make decisions regarding the output, thereby improving the DEs understanding of the output.

  • Integrative – is easy to use and less time consuming and, therefore, fits into an iterative process.


I propose that the materialisation of this approach be used for assessing the independency assumptions that are represented by the topology of the graph. The characteristics of this materialisation are as follows:

  • Rich –

    • Includes different functionalities to choose from.

    • Includes visualisation and verbal explanation that are intuitive to the user. This will provide understandable explications for DEs of the implications of the decisions they make by using visualisation and verbal explanation. Thus, it will be suitable for users who are not BN experts.

  • Interactive –

    • Includes an interface for the DE to choose nodes and relationships of interest to be explored.

    • Includes an interface for the user to decide on the details of the visualisation (colours and shapes) and on the verbal explanation (terminology and phrases).

    • Includes a methodology to guide the user in the investigation. The methodology is a structured execution of the functionality, so the investigation can be conducted in a systematic and exhaustive way. Alternatively, it can simply guide the user with questions such as: How to conduct the investigation? How to start the investigation? What relationships should be investigated? etc.

  • Integrative –

    • The approach is easy to re-apply to the whole structure or only to the latest changes to the structure, which makes it appropriate for the use within an iterative process.

    • Changes to the details of the visualisation and verbal explanations are saved, do not have to be entered again, and are stored for further use.

    • The details of the investigation are saved and can be reviewed.

    • The methodology is flexible. In order to investigate the structure, the users can follow the methodology as a whole, follow parts of the methodology, use it as a guideline or ignore it altogether and follow other rules.


The following figure illustrates the place of the proposed approach and materialisation and the implemented tool in the context of the knowledge engineering process of BNs.




Figure 1.1: The place of the proposed approach, materialisation and tool in the context of the knowledge engineering process of BNs.

(BNKE – Bayesian Network Knowledge Engineering)









1   2   3   4   5   6   7   8   9   ...   35

Похожие:

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconThis document is being distributed by the American International Health Alliance (aiha) / the EurasiaHealth Knowledge Network with support from the us agency
Саратовский областной Центр по профилактике и борьбе со спид и инфекционными заболеваниями
Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconStudents will demonstrate competence in; Engineering Curriculum; basic concepts such as significant digits, length, and mass; Engineering Drawing; the use of the computer as an engineering tool; designing a project

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconAn Object-Oriented Support Tool for the Design of Casting Procedures

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconRegional network of co-operation in Europe as a tool of efficient governance

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconEnabling Decision Support and Costing of Product Designs by using Visual Metaphors

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconRepresenting and using legal knowledge in integrated decision support systems: DataLex WorkStations

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconThis document contains a list of references to publications and reports about Bayesian Net technology, and especially Bayesian Net applications. The report will

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconNorwegian Network for the Support of the Indigenous Peoples of the Russian Arctic (nnsipra) Сеть Норвежских Организаций в Поддержку Коренных Народов Российского Севера

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconDevelopment of a Product model to support engineering change management

Support for Graphical Modelling in Bayesian Network Knowledge Engineering: a visual Tool for Domain Experts iconTool Use: Misuse of shop tools will result in the loss of tool privileges

Разместите кнопку на своём сайте:
Библиотека


База данных защищена авторским правом ©lib.znate.ru 2014
обратиться к администрации
Библиотека
Главная страница