Research for Remaining time Prediction Log Reduction Method
Jian‐Hong Ye, Yan Lin, YongJin Wu, HongKai Huang · Research Square · 2023
Abstract Information systems record a large amount of event log data, including low-value, redundant data, during the course of business operations. When performance prediction is performed directly on these logs, the efficiency of the prediction is reduced. How to simplify and compress this data while preserving the effective value has been an issue explored by researchers. Most of the existing approaches are considered from a data dimensionality perspective, where dimensionality reduction is performed by removing redundant and irrelevant features. There are also model-driven perspectives that use model simplification rules to reduce model parameters, transforming the original dataset into a new simplified dataset while preserving some valid values. However, relatively few studies have investigated the efficiency of execution before and after event log simplification. In this paper, we address the residual prediction goal of log simplification by first proposing a prediction point selection algorithm using a role based on a network of resource communities to avoid simplifying all similarly functioning points in the same resource community, which affects the efficiency of subsequent predictions. Then the sequence, or and self-loop structrues are selected to convert into a sim-plifiable part, and the deviation between the actual simplifiable and the original data prediction value is algorithmically optimised to avoid oversimplification. Experiments show that the simplified event log not only maintains its predictive performance, but in some scenarios even improves its predictive accuracy over the original event log due to reduced overfitting.