Fuzzy-belief k-nearest neighbor classifier for uncertain data

Abstract : Information fusion technique like evidence theory has been widely applied in the data classification to improve the performance of classifier. A new fuzzy-belief K-nearest neighbor (FBK-NN) classifier is proposed based on evidential reasoning for dealing with uncertain data. In FBK-NN, each labeled sample is assigned with a fuzzy membership to each class according to its neighborhood. For each input object to classify, K basic belief assignments (BBA's) are determined from the distances between the object and its K nearest neighbors taking into account the neighbors' memberships. The K BBA's are fused by a new method and the fusion results are used to finally decide the class of the query object. FBK-NN method works with credal classification and discriminate specific classes, metaclasses and ignorant class. Meta-classes are defined by disjunction of several specific classes and they allow to well model the partial imprecision of classification of the objects. The introduction of meta-classes in the classification procedure reduces the misclassification errors. The ignorant class is employed for outliers detections. The effectiveness of FBK-NN is illustrated through several experiments with a comparative analysis with respect to other classical methods.
Document type :
Conference papers
Complete list of metadatas

https://hal-onera.archives-ouvertes.fr/hal-01070480
Contributor : Véronique Soullier <>
Submitted on : Wednesday, October 1, 2014 - 2:43:46 PM
Last modification on : Friday, June 28, 2019 - 7:22:07 PM
Long-term archiving on : Friday, January 2, 2015 - 11:06:08 AM

File

DTIM14032.1404207568.pdf
Files produced by the author(s)

Identifiers

  • HAL Id : hal-01070480, version 1

Citation

Zhun-Ga Liu, Quam Pan, Jean Dezert, Grégoire Mercier, Yong Liu. Fuzzy-belief k-nearest neighbor classifier for uncertain data. Fusion 2014 : 17th International Conference on Information Fusion, Jul 2014, Salamanca, Spain. pp.1-8. ⟨hal-01070480⟩

Share

Metrics

Record views

476

Files downloads

340