Efficient data reduction in multimedia data

摘要

As the amount of multimedia data is increasing day-by-day thanks to cheaper storage devices and increasing number of information sources, the machine learning algorithms are faced with large-sized datasets. When original data is huge in size small sample sizes are preferred for various applications. This is typically the case for multimedia applications. But using a simple random sample may not obtain satisfactory results because such a sample may not adequately represent the entire data set due to random fluctuations in the sampling process. The difficulty is particularly apparent when small sample sizes are needed. Fortunately the use of a good sampling set for training can improve the final results significantly. In KDD'03 we proposed EASE that outputs a sample based on its 'closeness' to the original sample. Reported results show that EASE outperforms simple random sampling (SRS). In this paper we propose EASIER that extends EASE in two ways. (1) EASE is a halving algorithm, i.e., to achieve the required sample ratio it starts from a suitable initial large sample and iteratively halves. EASIER, on the other hand, does away with the repeated halving by directly obtaining the required sample ratio in one iteration. (2) EASE was shown to work on IBM QUEST dataset which is a categorical count data set. EASIER, in addition, is shown to work on continuous data of images and audio features. We have successfully applied EASIER to image classification and audio event identification applications. Experimental results show that EASIER outperforms SRS significantly.