A growing and pruning strategy-based resource allocation neural network for text categorization
首发时间:2013-10-28
Abstract:In this paper, we propose a novel learning classifier which utilizes growing and pruning strategy-based resource allocation neural network (GPRAN) for text categorization. In order to avoid the limitation of local-optimal and reduce the sensitivity of the input dataset, GPRAN uses an agglomerate hierarchical k-means method to effectively initialize the centers of hidden layer and adopts a least square method to initialize the weight from hidden layer to output layer. The algorithm determines the structure and complexity of the GPRAN by dynamically growing and pruning the hidden centers. Then least square method is uesd to enhances the network's ability for classifying. Such a dynamic approach builds a compact network structure which decreases the computational complexity and maintains the higher convergence rate. In order to solve the issue of text organization, we utilize a semantic similarity approach which reduces the input scales of neural network and reveals the latent semantics between text features. In experiments, the results reveal that the developed learning process enhances the predicted precision and decreases its computational complexity.
keywords: text categorization resource allocation network novelty criteria least square method
点击查看论文中文信息
基于增加和剪枝策略的资源分配神经网络的文本分类
摘要:本论文提出一种新的利用基于增加和剪枝策略的资源分配神经网络(GPRAN)进行文本分类。为了避免局部最优限制和降低初始数据的敏感性,GPRAN采用分层凝聚的K-means算法来初始化网络的隐层中心,并且使用最小二乘法初始化隐层到输出层的权值。本算法通过动态添加和删除隐层中心来确定网络的结构和复杂度,采用最小二乘法进一步提高网络的分类能力。这样的方法能够创建一个紧凑的网络结构,减少计算的复杂性,并且维持较高的收敛速率。为了解决文本组织问题,本算法采用语义相似度方法降低神经网络的输入规模并且揭示文本特征间的潜在相似性。实验结果表明其动态的学习过程不仅提高了分类的精度,而且降低了计算的复杂度。
基金:
论文图表:
引用
No.4564749927030138****
同行评议
共计0人参与
勘误表
基于增加和剪枝策略的资源分配神经网络的文本分类
评论
全部评论0/1000