Never judge a Case by its (unreliable) neighbors: Estimating Case Reliability for CBR

It is known, in the Case-based Reasoning community, that the effectiveness of a Case-Based Reasoner is heavily dictated by the quality of cases in its Case Base, and therefore the presence of poor quality cases can adversely influence its predictions. While it is common practice for a domain expert to periodically check the cases in the case base of a reasoner, it often becomes a time-intensive exercise in practice. Existing literature provides potential measures of reliability of cases in the case base, however, they fail to provide robust estimates of the reliability of a case when its neighborhood comprises of mixed quality cases. In this work, we propose RelCBR which builds upon a circular definition of case reliability - a case is reliable if it is well-aligned with its reliable neighbors. This formulation allows us to arrive at more robust estimates of reliability and results in streamlining the case base maintenance process by drawing the attention of the expert to cases that are more likely of being incorrect. In addition, these reliability values can also be used to discount the contribution of unreliable cases in the reasoning process that would consequently boost the performance of the reasoner.