Explanation in Case-Based Reasoning–Perspectives and Goals | Sørmo, Cassens & Aamodt
This paper provides an overview of explanation and explanation goals in case-based reasoning. It's quite long and I only have time to skim it for now.
Oriented on explanation goals:
We will present a framework of explanation based on important explanation goals, and discuss how they place limitations on each other and how different kinds of systems may be better suited to fulfill different goals. (110)
Philosophical and Cognitive Accounts of Explanation
Skipped for now.
Explanations in Expert Systems
- Reasoning trace
- Not that interesting to user and often incomprehensible
- Justification for reasoning step by reference to deeper background knowledge
- Strategic: explaining the reasoning strategy rather than the specific instance
- Terminological: defining and explaining terms and concepts in a domain
Requirements:
- Fidelity: "the explanation given should mirror the knowledge used by the system in its reasoning" (122)
- Low construction overhead; justifiable resource usage
- Understandable to user
- Sufficient: "enough knowledge must be represented in the system to answer the question the user may have" (122)
Fidelity may be controversial and one could say that the proces of giving explanations can be a separate process that does not need to mirror the workings of how the system came to an outcome.
Wick and Thompson (1992) propose three explanation goals:
- "Verification is the goal of the knowledge engineer in verifying that the system works as it should. A successful verification explanation would accurately and precisely convey the knowledge of the system on the knowledge level." (122)
- "Duplication is to help the domain expert examine the knowledge of the system. The system should not only expose its own knowledge, but help the user learn the methods and knowledge used in the problem solving process."
- "Finally, the goal of ratification is to increase the end user’s confidence in the system’s conclusion." (122)
If the goal is ratification, "the explanation process should increasingly be decoupled from the reasoning process in order to provide explanations that focus on the solution." (14)
Explanation in CBR
In practice, CBR is often used in systems that do support user decisions, so there is a lot of overlap in the use context and explanation goals of expert systems.
It has long been an article of faith in the CBR community that displaying an earlier solved case that represents a situation similar to the present problem situation can serve as a good explanation for adopting the solution of the previous case. (124)
Explanation goals
Explanation goals identified by this paper:
- Transparency: How the system reached the answer
- Justification: Explain why the answer is a good answer
- Relevance: Explain why a question asked / strategy chosen is relevant
- Conceptualization: Clarify the meaning of concepts
- Learning: Teach the user about the domain
Case as explanation
Basic tenet: provide a similar case to your current problem, so that the solution to that case can be investigated. This does assume that it's possible for the user to easily see the similarity.
Both of these views depend on the user’s ability to understand the case and to confirm the similarity assessment. In general, for the retrieved case to serve as an explanation to the user, the similarity between the retrieved case and current problem must be obvious to him. The difficulty for the user in comparing cases increases as the case structure becomes more complex and the similarity measures more convoluted. It also increases with the use of more complex adaptation techniques where the retrieved case may not be the most similar but one that fits the adaptation process, as suggested e.g. by Smyth and Keane (1998). (127)
Knowledge containers
"Richter (1995) describes the knowledge sources used in problem solving as knowledge containers." (128)
Expert systems have facts and rules as "knowledge containers". CBR systems have four of those:
- Case base
- Similarity measure
- Adaptation knowledge
- "allows the reasoner to change the solution of a previous case to better fit a new problem" (128)
- Vocabulary
- essentially domain knowledge and structure of the domain e.g. ontology
Some interesting points:
Richter points out that given a complete case coverage of the problem domain, the similarity measure and adaptation problems become trivial since any problem can simply be looked up. Similarly, if we have perfect adaptation knowledge so that any previous solution can be adapted to a perfect new solution, the process only requires a starting position for the adaptation so that there is little need for cases or a similarity measure. Finally, if the system is always able to order the cases so that the cases with the correct solution are ranked highest, a classification system only needs a case representing each solution class, and there is no need for adaptation knowledge. (128-129)
- If case coverage is complete, you can always look up the problem solution
- If you can always perfectly adapt a case, you don't necessarily need a similarity measure
- If you can rank cases with the highest solution, you don't need adaptation knowledge because ... ?
Survey of explanation in CBR
- Showing the most similar case
- Is not necessarily the most convincing or informative!
- More difficult with ensemble classifiers
- Multi-level explanations
- Similar case as top level
- More detail e.g. by explaining based on feature values or decision boundaries
- Visualization
- e.g. of similarity of cases
- or on which attributes cases differ
- Separate explanation model/system to produce explanations
- e.g. IBP or CATO system (Ashley involved in both) combine "model-based reasoning (...) with CBR" (131)
- e.g. backwards after-the-fact reasoning/chaining using causal models to infer what must have been the cause of the outcome
- etc.
- Reasoning trace
- Feasible if producing explanations is already part of the reasoning process
- Case space awareness
- The competence of a CBR systems relies on the availability of good cases in the case base
- Do not hide contrasting evidence
- "displaying the most similar case(s) that are not of the proposed solution class to the user" (134)
- "The Stamping Advisor (Leake et al., 2001a) mentioned in Section 5.5 displays cases which are close to each other but with different findings." (134)
- " Compound confidence measures can also be calculated as for instance in (Cheetham and Prince, 2004)." (134)
- "McSherry’s ProCon system (McSherry, 2004) identifies which attributes of the input case support the suggested solution and which attributes oppose it. The attributes are identified as opposers or supporters of a solution based on how the attributes affect the probability of the solution. This allows the system to present justifications that are not only simpler to understand than possibly complex case similarity measures, but also help the user to identify what attributes are important to the conclusion." (134)
- Simplified problem solving strategy
- Concept maps
- Machine learning induction
- e.g. see ProCon example
Challenges
- Maintaining transparency in complex systems
- Providing justification to novice users
- Connecting cases to general knowledge in tutoring
- Scope of explanation efforts
Conclusions
The goals that an explanation is required to achieve vary with the domain, system, and user. It can be hard to model these dynamically for the system itself, but the designer of the system can often make assumptions about the goals and capabilities of prototypical users of the system. We also believe that explicitly formulating such explanation goals facilitates the discussion of possible conflicts between goals and makes clear how different approaches tend to favor different types of goals. (139)