关键词: 贡献因子, 样本属性, 径向基网络, 结构信息, 蛋白质二级预测

Abstract: At present the improvement of RBF (radial basis function) neural network mostly focuses on the hidden node selection, the function of learning rate or large-scale data organization, while the structure of the initial information in input samples is often ignored. It is also found that the different properties in input samples have a different influence on the classification capability and each property should have own category weight. After normalizing the sample, this paper studies the contribution factor to the classification for each property, then proposes to add the contribution factor to RBF network named CFRBF (contribution factor RBF) algorithm and uses the contribution factor to describe the importance of different properties. The paper chooses the prediction of protein secondary problem to validate the algorithm. The traditional secondary prediction is just to put the original samples into the network, and then classify them by hamming distance alone, so it loses a lot of information. For the new design model, the paper uses a new organization form to deal with the prediction problem. Experimental results show that after the introduction of new organizational forms, the net will be significantly improved, and its accuracy will be again increased by the CFRBF algorithm. At the same time, the contribution factor appears to reveal the irregular structure of amino acids between the state of relations, which also shows the importance of different properties in the samples.

Key words: contribution factor, sample properties, radial basis function network, structure information, protein secondary prediction