Research On Improvement ELM Based Filling Approach Of Missing Data

The constantly development of Sensing technology, network transmission and storage technology brought a great improvement to people’s capability to produce and access data. Data acquisition tools also continue to progress, information analysis became convenient because there are enough data. However, many complex data have generated a lot of issues, among them; the data missing is the fundamental problem of data quality. How to handle the added missing data accurately? To make the good supplemented data to be ready for use? This will be the data quality research key point. Therefore, our research and data filling have theory and method relationship.First, this thesis introduces the ELM approach to fill missing data. Because we may not know the prior distribution of the data set that need to be filled, thus, it is difficult to select the activation function. If the data set to fill is very complicated, probably there is no convenient activation function. The thesis research key point focused on a named A-ELM approach to fill missing data, the key point of the method is to use the improved differential evolution algorithm to select the activation function.Second, because of the linear relationship of variables of other data set, therefore, to calculate the generalized inverse matrix will be very difficult in ELM. This thesis also focuses on the research of missing data filling called B-ELM, the key point of the method is to use the GCV algorithm to select parameter k. Then, figures out the AB-ELM-MD, it can fill some sets of data, we don’t know these data sets distribution function, or related dimension.Finally, this thesis used the real DNA datasets and traffic trace datasets as example by using three kinds of methods SVR, ELM, AB-ELM-MD for experiments, use the regression results to fill missing data. We proved that the proposed AB-ELM-MD can effectively fill missing data. It is has good effect to fill complex datasets, for example, the datasets with higher dimensions, each dimension between highly relevant, it is difficult to select the activation function, a large amount of datasets.