Learning Fully-Connected Neural Network for Skyline Computation

Abstract

The skyline operation is used to extract data that is not inferior to other data in the multidimensional data and such feature makes the skyline operation actively used in the recommendation system. The existing skyline calculation research includes BNL, Divide & Conquer, and Bitmap. These methods compare new data with all the skylines found to date for skyline discrimination. Therefore, there exists a weakness of increased comparison operation with increased number of skyline. There was a precedent study which learned virtual skyline representing the skyline to reduce such comparison operations, and the study actually decreased the comparison operation amount. In the precedent study, model was constituted with the partially connected artificial neural network and learned the virtual skyline using the backpropagation algorithm. In this study, we used TensorFlow, a Python machine learning library, to bring the studied method in the precedent study and shortened the time. The study also attempted fully connected artificial neural network instead of partially connected artificial neural network for efficient virtual skyline learning. Lastly, the research has been conducted to compare the artificial neural network model in this study to the existing search method. The model with Tensorflow and without Tensorflow, the partially connected neural network and fully connected neural network model were also compared and their performances were evaluated.