Research On Fast Incremental Classication Algorithms

As the popularization of equipments for collecting and editing data, it becomes more and more convenient for people to produce multimedia data such as images, graphics, audio, videos, cartoons and 3D models; Again, as the prevalence of Internet and large-scale storage technologies, no matter when and where, people can visit the multimedia data in the Internet. In general, the Internet data are isomerous, unstructured, high-dimensional and dynamic, these properties give birth to huge difficulties for the following work such as classification, clustering, data mining, data understanding and utilities. Aiming at these data properties from Internet, this paper mainly studies the dimensionality reduction methods of high-dimensional data and incremental classification methods for large-scale data.To reduce the high-dimensional data, this paper proposes a fast iterative method based on nonnegative matrix factorization. The proposed method uses the property of L1 normalization and the sparsity of relevant data matrix, and designs much simpler iterative updating formulas. The experimental results demonstrate that, this method can both fast reduce dimensionality and improve the following classification or clustering accuracy.Based on the matrix factorization described above, to classify large-scale data, this paper investigates an inverse matrix-free incremental learning method. To solve the inverse matrix problem in proximal support vector machine, this novel method designs new formulas for updating models without calculating inverse matrix, and obtains fast incremental learning. The experimental results show that, compared with the original method, under the condition of basically retaining prediction ac-curacy, this method has one order of magnitude less time complexity.Furthermore, to deal with the dynamic large-scale data, this paper proposes an incremental transfer learning method. This method derives a novel model which is appropriate for weighting samples. Then, prediction feedback mechanism is used to adjust the model parameters from auxiliary data. As a result, the proposed method is capable of handling the changes from data sizes and distributions. The experi- mental results demonstrate that, the proposed method can obtain better and faster effectiveness than non-transfer methods, and transfer knowledge bidirectionally.In a word, to handle large-scale, high-dimensional and dynamic data, this paper proposes three fast learning algorithms about dimensionality reduction and classifi-cation. Theoretical analyses and enough experimental results show that these novel methods achieve much higher efficiency and better prediction accuracy than their rival methods.