Improved techniques in vector auto regression for time-series link prediction

• Main Author: CO, JAN MILES A.
• Format: text
• Published: Archīum Ateneo 2017
• Subjects: Regression analysis > Computer programs; Regression analysis > Data processing; Time-series analysis; Autoregression (Statistics); Computer networks; Computer algorithms; Machine learning.
• Online Access: https://archium.ateneo.edu/theses-dissertations/43; http://rizalls.lib.admu.edu.ph/#section=resource&resourceid=1191832099&currentIndex=0&view=fullDetailsDetailsTab
• Institution: Ateneo De Manila University

Link Prediction is an area in network analysis that deals with determining the existence of hidden links or the emergence of new links. There are two approaches to the link prediction problem. The static approach uses one network snapshot, while the dynamic or time-series approach uses the current and some previous configurations of the network for the prediction of future links. Based on a previous work, the Vector Auto Regression (VAR) technique has been shown to be one of the best for time-series based link prediction. In this study, we were able to improve the VAR technique by several approaches. Our proposed methods were investigated on different datasets such as DBLP, Enron, and several synthetic datasets. Our proposed techniques were able to significantly improve the baseline VAR with 2 lags. For the DBLP dataset, the baseline AUC-ROC of 84.96% was increased to 86.32% by using our proposed technique that uses SVM with 5 lags. Furthermore, the baseline was also increased to 86.69% and 86.00% by using our improved VAR technique that uses our proposed time-based features and by using our improved VAR technique that uses our proposed one sided noise reduction technique. Based on the results, we discovered that there are several approaches that can significantly improve the baseline VAR for our datasets. Our proposed methods for improved link prediction can be applied to other related applications of the VAR technique outside the context of the link prediction problem.