An evidential integrated method for maintaining case base and vocabulary containers within CBR systems

Abstract

Cases and vocabulary maintenance presents a crucial task to preserve high competent Case-Based Reasoning (CBR) systems, since the accuracy of their offered solutions are strongly dependent on stored cases and their describing attributes quality. The maintenance aims generally at eliminating two types of undesirable knowledge which are noisy and redundant data. However, inexpedient Case Base Maintenance (CBM) or vocabulary maintenance may not only greatly decrease CBR competence in solving new problems, but also reduce its performance in term of retrieval time. Besides, to provide a high maintenance quality, it is necessary to manage uncertainty within knowledge since "real-world data are never perfect" and stored cases within a CBR system's Case Base (CB) describe realworld experiences. Hence, we propose, in this paper, a new integrated method that maintains both of the CB and the vocabulary knowledge containers of CBR systems by offering a new alternating technique to properly detect noisiness and redundancy whether in cases or features. During the learning steps of our new integrated maintenance policy, which drives the decision making about cases and attributes selection, we manage uncertainty using one among the most powerful tools called the Belief Function Theory.

CIMMEP: constrained integrated method for CBR maintenance based on evidential policies

“…Actually, this imperfection could be handled using different theories such as fuzzy logic [27], rough sets [28], or other works on multigranulation [29]. However, in the integrated maintenance context, we find, unfortunately, only one work [30] that uses one among the most powerful tools for this matter called Evidence or belief function theory [3,4]. We aim, in this paper, to address a major limitation of all reviewed methods that figures in their disability to aid learning operations using prior knowledge.…”

“…Otherwise, they have the same objective function to be optimized in order to build the credal partition. The objective function of the non-constrained version of WCECM (WECM algorithm [30,36]) is defined such that:…”

“…To make difference between these two types, we compare the distance between cases and clusters' centers to a threshold (T h k ) which is set to the average of cases distances from a given cluster's center. As in [17,19,30], we used the Belief Mahalanobis Distance (BMD) [17] so as to take into account different shapes of distribution, along with managing uncertainty. We make difference, therefore, between Similar and Isolated cases as follows:…”

“…Thus, weights of features that have been flagged as Noisy, will be updated in form of penalization, during Step 3 of WCECTD. Since we choose to apply iterations of the two mentioned phases as times as the number of features, let update weights of features flagged as noises, from one iteration to another, in the following way [30]:…”