Extending MAP-independence for Bayesian network explainability

Resumen

In the past years there has been an increasing interest in explainable AI (XAI), since it can be a potential solution to the performance, ethical and legal concerns of the new obscure complex models such as neural networks. Selecting transparent models over top performing ones can be a better option in terms of both performance and explainability. As such, in this work we use Bayesian networks. Stability is a desirable property of explanations which consists of avoiding that a minimal change in the input data leads to a significant modification in the explanation. We review MAP-independence [3] as a measure related to stability for Bayesian network explanations and formulate properties that improve computational efficiency and expand the concept to continuous domains. In Section 2, we review the basic concepts concerning our proposal, which is presented in Section 3. In Section 4, conclusions about our work are drawn.