Kernel-based Principal Components Analysis on Large Telecommunication Data

Linear Principal Components Analysis (LPCA) is known for its simplicity to reduce the features dimensionality. An extension of LPCA, Kernel Principal Components Analysis (KPCA), outperforms LPCA when applied on non-linear data in high dimensional feature space. However, on large datasets with high input space, KPCA deals with a memory issue and imbalance classification problems with difficulty. This paper presents an approach to reduce the complexity of the training process of KPCA by condensing the training set with sampling and clustering techniques as pre-processing step. The experiments were carried out on a large real-world Telecommunication dataset and were assessed on a churn prediction task. The experiments show that the proposed approach, when combined with clustering techniques, can efficiently reduce feature dimension and outperforms standard PCA for customer churn prediction.