A NEW Cluster-histo-regression ANALYSIS FOR INCREMENTAL LEARNING FROM TEMPORAL DATA CHUNKS

Abstract

In scenarios where data chunks arrive temporally, a good algorithm for exploratory analysis should be able to generate the knowledge and with the next chunk of data arriving, the process should be the one of just updating online by accumulating the knowledge derived from the recent chunk. Such an incremental learning process in most of the cases indent a lot of memory requiring to carry all earlier data in the process of updating the knowledge successively. In this research work we propose to employ a novel Cluster-Histo-Regression analysis of the chunk to extract the knowledge for the temporal instant and fuse this knowledge through Histo-Regression-Distance analysis with the already accumulated knowledge. We have designed a methodology which (i) discards all those data samples from the chunk which have participated in the knowledge generation process (ii) indents minimum amount of memory to carry the accumulated knowledge and (iii) proposes to carry forward only those limited data samples (referred to as hard samples) which could not contribute to knowledge generated at that moment. Knowledge of each cluster is represented in the form of a histogram for each dimension of the clustered data and is transformed to regression line for the compact representation of the knowledge. The regression line parameters of the clusters obtained by incremental augmentation have shown an accuracy of up to 100% for some of the data sets that are considered for experimentation.