Very large-scale data classification based on K-means clustering and multi-kernel SVM

When classifying very large-scale data sets, there are two major challenges: the first challenge is that it is time-consuming and laborious to label sufficient amount of training samples; the second challenge is that it is difficult to train a model in a time-efficient and high-accuracy manner. This is due to the fact that to create a high-accuracy model, normally it is required to generate a large and representative training set. A large training set may also require significantly more training time. There is a trade-off between the speed and accuracy when performing classification training, especially for large-scale data sets. To address this problem, a novel strategy of large-scale data classification is proposed by combining K-means clustering technology and multi-kernel support vector machine method. First, the K-means clustering method is used on a small portion of the original data set. The clustering stage is designed with a special strategy to select representative training instances. Such method reduces the needs of creating a large training set as well as the subsequent manual labeling work. K-means clustering method has two characteristics: (1) the result is greatly influenced by the cluster number k, and (2) the optimal result is difficult to achieve. In the proposed special strategy, the two characteristics are utilized to find the most representative instances by defining a relaxed cluster number k and doing K-means repeatedly. In each K-means clustering step, both the nearest and the farthest instance to each cluster center are selected into a set. Using this method, the selected instances will have a representative distribution of the original whole data set and reduce the need of labeling the original data set. An outlier detection method is applied to further delete the outlier instances according to their outlier scores. Finally, a multi-kernel SVM is trained using the selected instances and a classifier model can be obtained to predict subsequent new instances. The evaluation results show that the proposed instance selection method significantly reduces the size of training data sets as well as training time; in the meanwhile, it maintains a relatively good accuracy performance.