Proposition and learning of some belief function contextual correction mechanisms

A new evidential K-nearest neighbor rule based on contextual discounting with partially supervised learning

“…In future work, it will be interesting to find out whether this richer information can be exploited for, e.g., classifier combination. Beyond contextual discounting, other mass correction mechanisms such as introduced in [27], [23]- [24] and [28] could also be investigated. Require: {η k , γ k } c k=1 , ξ, α, {pl ik , pl ik , p ik } n,c i,k , {β ijk } n,K,c i,j,k , {d ij } n,K i,j…”

Combination of Classifiers With Optimal Weight Based on Evidential Reasoning

“…This contextual discounting operation can properly redistribute the masses of belief according to the reliability vector. The set of tools has been enlarged in [29] to deal with the contextual knowledge about the source quality in terms of relevance and truthfulness based on belief function theory. The practical means to learn the contextual knowledge from available labeled data are also introduced in [29], and it makes the correction mechanisms interesting and useful in practice.…”

Quality of Information Sources in Information Fusion

“…The discounting operation proposed by Shafer [28] to integrate the reliability of information sources is a special case of transformation (9), recovered for q = 1: it corresponds to a partially relevant source that is truthful. The negating operation [26] is also a special case recovered for p = 1, which corresponds to a partially truthful source that is relevant. In particular, the negation of a mass function is obtained for p = 1 and q = 0: it corresponds to a relevant source that is lying 2 .…”

“…As shown in [26, Proposition 1], if s asserts x 2 A for some A ✓ X , and is assumed to be in statè B , for some B ✓ X , then one must deduce that x 2 (A \ B) [ (A \ B). We refer the reader to [26] for examples where such states`B may be relevant.…”

“…This suggests again to consider states corresponding to weaker assumptions on the lack of truthfulness. Two states are particularly interesting, as shown in [26]. The first one, denoted p B , corresponds to the assumption that a source is (positively and negatively) truthful for all These states also fit our approach since, e.g., the transformations associated to states n B can be represented by a multi-valued mapping G A from H = {n B |B ✓ X } to X such that G A (n B ) = A [ B, for all B ✓ X .…”