Rongze Xia

A BP neural network text categorization method optimized by an improved genetic algorithm

The back propagation(BP) neural network is widely used for text categorization and could achieve high performance. However, the greatest disadvantage of this network is its long training time. The genetic algorithm is often used to generate useful solutions for optimization. In this paper we combined the genetic algorithm and the back propagation neural network for text categorization. We use the genetic algorithm to optimize weights of connections in the back propagation neural network instead of back-propagating. At the same time, we improved the genetic algorithm to increase its efficiency. Through this method, we overcome the traditional disadvantage of the BP neural network. Our experiments show that our method outperforms the traditional method for text categorization.

A New Combination Sampling Method for Imbalanced Data

Imbalanced data is commonly in the real world and brings a lot of challenges. In this paper, we propose a combination sampling method which resamples both minority class and majority class. Improved SMOTE (ISMOTE) is used to do over-sampling on minority class, while distance-based under-sampling (DUS) method is used to do under-sampling on majority class. We adjust the sampling times to search for the optimal results while maintain the dataset size unchanged. Experiments on UCI datasets show that the proposed method performs better than using single over-sampling or under-sampling method.

Ensemble multi-label learning based on neural network

Multi-label classification problem refers to predict each instance to be one or more labels in a given label set. It is very common in the real world, e.g. image annotation. Based on a comprehensive analysis of existing researches, we propose a new ensemble learning method for multi-label classification problems. AdaBoost and multi-label neural network are integrated to enhance the generalization ability of the method. Experiments on three standard datasets show that the proposed method performs well.

Imbalanced Chinese Text Classification Based on Weighted Sampling

Traditional text classification methods assume that dataset is balanced. But, in the real world, there are plenty of imbalanced data on which traditional classification methods could not get satisfactory results. Based on comprehensive analysis of existing researches on imbalanced data classification, we propose a data rebalance method based on weighted sampling. The method assigns weights to each class by calculating the ratio between different categories. Then, each class is sampled with different ratios using weighted sampling methods. Experimental results on real Chinese text data set show that the proposed method can effectively improve the classification accuracy.