TL;DR
This paper introduces a novel "Effect" measure to provide meaningful explanations for changes within Bayesian inferencing systems. It addresses the critical need for users to understand why a system reaches a conclusion, moving beyond simple control explanations to delve into the knowledge base content itself.
What Problem Does This Paper Address?
Expert Systems (ES) are increasingly used in critical domains like medicine, where human users bear ultimate responsibility. For these systems to be trusted and effective, they must not only provide recommendations but also explain how they arrived at them. While existing ES frameworks offer explanations related to control flow (e.g., "why this rule fired"), they often fall short in explaining the underlying knowledge base content. This is particularly challenging in systems that handle uncertainty, where the semantics of inferencing can be complex and inconsistent. This paper tackles this gap by focusing on generating explanations for changes in variable probabilities within Bayesian inferencing systems, specifically addressing the "why" behind these probabilistic shifts.
Methods
The core contribution of this paper is the proposal of an "Effect" measure, denoted as $Ef(H,E)$, designed to quantify the impact of an evidence event $E$ on a hypothesis variable $H$ within a Bayesian inferencing system. This measure is calculated in three steps:
- Correlation Measure: The difference between the joint probability of $H$ and $E$, and the product of their marginal probabilities ($Pr(H \cap E) - Pr(H)Pr(E)$). This captures the degree of correlation.
- Evidence Probability Change: The difference between the updated probability of the evidence event $E$ and its prior probability ($Pr'(E) - Pr(E)$). This quantifies how much the evidence itself has changed.
- Hypothesis Probability Change: The difference between the updated probability of the hypothesis $H$ and its prior probability ($Pr'(H) - Pr(H)$). This quantifies how much the hypothesis has changed.
The final "Effect" is the product of these three components: $Ef(H,E) = [Pr(H \cap E) - Pr(H)Pr(E)] \times [Pr'(H) - Pr(H)] \times [Pr'(E) - Pr(E)]$
This measure is designed to be intuitively pleasing, increasing with the magnitude of correlation and probability changes. The paper also discusses desirable properties for such a measure, including symmetry ($Ef(H,E) = Ef(E,H)$) and insensitivity to the negation of an event ($Ef(H, \neg E) = Ef(H,E)$), which the proposed measure satisfies. The authors implement this measure within the Generalized Bayesian Inferencing System (GBI), a framework supporting Bayesian hierarchical inferencing.
Datasets
Not discussed in the paper. The paper uses a sample LEG Network (Figure 1) for illustrative purposes, which is a hypothetical construct rather than a dataset.
Key Insights and Ideas
- Beyond Control: Explaining Knowledge Base Content: The paper argues that true user trust in ES requires explanations that go beyond just the inferencing control flow and delve into why the knowledge base itself is changing.
- The "Effect" Measure: The proposed $Ef(H,E)$ measure provides a mathematically grounded way to quantify the influence of one event on another within a Bayesian network, considering correlation and probability shifts.
- Symmetry is Key: The requirement for symmetry in an explanation measure is crucial for Bayesian systems, ensuring that the explanation is consistent regardless of whether the evidence or hypothesis is the starting point.
- Local vs. Global Explanations: The paper distinguishes between "local" explanations (within a single LEG) and "global" explanations (tracing causal chains across multiple LEGs), offering different levels of detail for users.
- Causal Structure as Control Information: While GBI doesn't inherently understand causality, the paper suggests that explicitly defining causal structures can serve as valuable control information for generating more targeted explanations.
Key Results
The paper presents the theoretical framework for the "Effect" measure and its implementation within the GBI system. It demonstrates how this measure can generate both detailed (for knowledge engineers) and simplified (for typical users) local explanations. Figure 2 and Figure 4 provide examples of these local explanations, showing how changes in one variable's probability are explained by correlated changes in other variables within the same LEG. Figure 6 illustrates a global explanation, tracing a causal chain of effects. The paper emphasizes that the measure is the basis for the current GBI explanation capability and that it can be extended to multivariate explanations.
Limitations and Future Work
The authors acknowledge several limitations and suggest future directions:
- Univariate Explanations: The current implementation primarily focuses on univariate explanations (effect of one event on one variable). The paper expresses plans to extend this to multivariate explanations, where joint events might have a stronger influence.
- Natural Language Generation: The current explanation output is described as "tedious fill-in-the-blank." A significant future enhancement would be to integrate better natural language generation capabilities.
- Graphical Display: A graphical display of the LEG Net and causal structures is suggested to improve user interface convenience and make explanations easier to understand and direct.
- Handling Special Relationships: The system could be enhanced to recognize and handle special relationships between variables, such as mutual exclusion and implication.
- Control Information Integration: While the paper focuses on knowledge base explanations, it notes that integrating knowledge of control mechanisms would further enrich explanations, detailing how effects propagate through the system.
- Likelihood Ratio as an Alternative: The paper briefly discusses the likelihood ratio as an alternative measure but finds it less suitable for direct substitution into their proposed framework due to its non-negative nature and range.
Summary updated