Case-based reasoning (CBR) is an artificial intelligence (AI) methodology with applications across diferent domains. Its reasoning is rooted in human intuition making this methodology universally accepted as interpretable. In contrast, black box models like neural networks have a widely accepted notion with high-performance accuracy and low interpretability. Despite low interpretability, there is a ready acceptance of black box models in academia and industry for automated decision-making. The focus of this research is investigating the relationship between performance and interpretability for CBR by exploring a variety of ways to note the impact of enhancing performance on CBR interpretability.
diferent ways to improve CBR performance (while maintaining its interpretability) would detail impact factors that can serve as a reference to future researchers and provide for the ready acceptance of CBR in high-stakes decision-making as an interpretable alternative to black-box methods.
Case-based Reasoning Case-Based Reasoning (CBR) is an AI methodology that uses
knowledge from past experiences to solve new problems. What distinguishes it from other reasoning
methodologies is its ability to identify the most “similar” experiences for solving a problem.
Since the solution of the experience(s) may not be fully applicable to the current problem, it
allows the methodology to tweak (or adapt) the past experience to solve the problem [
CBR has two foundational perspectives of understanding. The first perspective is as a process
cycle (called CBR cycle[
A set of approaches pertaining to the similarity container are reviewed. Wetterscherek and
Aha [
Interpretability Interpretability in A.I. stems from diferent notions inspired to gather a
holistic understanding of a machine learning model’s operability. The field has gained immense
popularity in the past decade primarily due to the increased usage of black box AI models
in high-stakes decision-making. Despite the stakes, the field lacks a unifying definition to
bind all afiliated notions [
Interpretability, as it stands today, is an amalgamation of diferent notions. Rudin [
An important subject of discussion is the relationship between performance and
interpretability for diferent AI methods. The DARPA XAI Program [
CBR methodology claims to have “natural” transparency (or inherent interpretability) [
Rudin [
This research aims to investigate methods to improve CBR performance. In addition, the impact of these methods on interpretability will be noted. Finally, the structure of the relationship between performance and interpretability will be analyzed and the level of sensitivity is observed in case a relationship is verified.
The first objective of this research is to improve CBR performance. The aim is to investigate a variety of methods and approaches to enhance CBR performance, where the approaches are limited to the retrieval step of the CBR cycle (or similarity container). The second objective is to note the impact of CBR approaches considered in the first research objective on interpretability. This objective would entail identifying the factors to govern CBR interpretability objectively and using these factors to witness changes in the approaches considered in the first research objective. Additionally, this objective aims to describe the structure of the relationship between performance and interpretability for the approaches considered in research objectives one and two. The aim is to identify the level of sensitivity between the two characteristics for approaches when an impact is noted.
The entirety of this research is divided into two phases. Phase 1 studies examine the literature on what has been proposed for performance in CBR and note increased performance as a direct consequence of a proposed alternative to similarity and retrieval-based approach.
This notion is explored via three studies. The first study incrementally increases the number of weights per feature to describe the relative relevance of features describing the cases. This includes methods involving clustering and weight-learning approaches. The second study uses the proposed methods in the literature for enhancing CBR performance. This comparison study would be comparing existing approaches in the literature by identifying datasets used across reviewed literature. The third study compares weight methodologies and their impact on performance. Where the focus is on increasing the number of weights in the first study, this study focuses on how weights are learned while keeping the number of weights constant. An additional aim of this phase is coming up with an approach built using approaches considered within the studies, that mimic the performance behavior of a neural network model.
Phase 2 investigates CBR interpretability notions to obtain factors that influence the
interpretability of CBR objectively. Consequently, this phase analyzes the relationships between the
performance and interpretability of CBR approaches and notes the sensitivity of the impact
one has on the other. Using the suggested concept of “cognitive chunks" discussed in Lage et al.
[
The first study for Phase 1 is ongoing. The baseline approach of gradient descent-based weight
learning (1 weight per feature) is altered by clustering cases according to some distance metric.
Once cases are clustered, each cluster has its own sets of weights (using gradient descent) to
separate the cases within. This gives every feature within a case represented by two weights
instead of one. One weight suggests the cluster the instance belongs to and the other weight
corresponds to the gradient descent method for the instance feature. This methodology was
evaluated using a 5-fold cross-validation split using the Credit Default Binary classification
Dataset [
The design for the second study is ongoing where the methods discussed in the reviewed literature, presented in the Related Works section, will be applied to note enhancements in performance.
This research is in its preliminary phases and no conclusions have been drawn yet. For limitations, the scope considers only the classification task. The investigations are, also, limited to the retrieval step and the similarity container.