PLDL: A Novel Method for Label Distribution
Learning
Venkatanareshbabu Kuppili,
Mainak Biswas, and Damodar Edla
Computer Science and Engineering, National Institute of
Technology Goa, India
Abstract: The nature, volume and orientation of data have been changed a lot in
the last few years. The changed situation has beckoned data scientists to
modify traditional algorithms and innovate new methods for processing new type
of high volume, extremely complex data. One of the challenges is label
ambiguity in the data, where the distribution of the significance of the labels
matters. In this paper, a new method named Probabilistic Label Distribution
Learning (PLDL) has been proposed for a computing degree of the belongingness.
It is based on a proposed new Label Probability Density Function (LPDF) derived
from Parzon estimate. The LPDF has been used in Algorithm Adoption K-Nearest Neighbors (AA-KNN) for Label Distribution
Learning (LDL). Probability density estimators are used to estimate this
ambiguity for each and every label. The overall degree of the belongingness of
unseen instance has been evaluated on various real datasets. Comparative
performance evaluation in terms of prediction accuracy of the proposed PLDL has
been made with Algorithm adaptation KNN, Multilayer Perceptron,
Levenberg-Marquardt neural network and layer recurrent neural for Label
Distribution Learning. It has been observed that the increase in prediction
accuracy for the proposed PLDL is highly statistically significant for most of
the real datasets when compared with the standard algorithms for LDL.
Keywords: Multi-label classification, data mining, label
distribution learning, probability density function.