Active Learning Based On Locally Linear Reconstruction Coefficients

Recently, active learning has been a hot topic in the field of patternrecognition. Active learning aims to choose the most informative datapoints for labels to improve the performance of the classifier by iteration.However, the previous methods just take into consideration globalstructure while the local manifold structure is ignored. This paper mainlyproposes a novel active learning algorithm and this kind of algorithmtakes into account the local manifold structure of the data space. That isto say, each data point can be linearly expressed by its neighbors. Giventhe locally linear reconstruction coefficients of each data point, this kindof method is called Active Learning Based on Locally LinearReconstruction. By incorporating the manifold structure into the activelearning process, we can solve the problem which incorporates themanifold structure into the active learning process very well.In this paper, the in-depth research and improvement of activelearning based on locally linear reconstruction coefficients is conducted,the concrete work includes the following several aspects: 1、The concept of active learning and several classical active learningalgorithms are reviewed.2、By introducing the locally linear reconstruction coefficients on thebasis of Manifold Adaptive Experimental Design, the new algorithmcalled Manifold Graph Model is presented and the steps of the algorithmare provided.3、Extensive experiments conducted on benchmark image datasetsand text datasets demonstrate that Manifold Graph Model algorithmperforms well in term of accuracy while comparing with the previousclassical algorithms.