Solving large-scale classification problem with approximate high dimensional indexing framework

Recently, linear classifier has been shown to be able to handle large-scale classification problem well. However, there are two main issues accompanied by large-scale classification problem. First, there may exist many unexplainable or noise instances in the datasets which will hurt the linear classifier’s performance. Second, when data is too large to load in memory, the linear classifier will spend much time on reading/writing between memory and disk. In this thesis, we propose an indexing optimization framework to solve these two issues simultaneously. We apply approximate indexing technique on high dimensional features space to help us efficiently retrieve the informative instances rather than outliers, and so that we can only load those instances into memory.
We conduct several experiments to compare our framework with the state-of-the art methods, and the results show that we have a better performance.