An Initial Study on Temporal Loss Functions for Remaining Time Models

EasyChair Preprint 5808

Abstract

Guided by two business goals (earliness and accuracy), this initial study investigate the performance of five different loss functions, across event-log data from four different domains (healthcare, public administration and IT services). Three different temporal losses are proposed for improvement of earliness-performance. The results show that MAE is either outperformed or tied with the temporal losses in terms of both earliness and accuracy. Based on the results from the experiments, the optimal weighting of the temporal penalty vary based on the characteristics of the event log. However, the proposed MAE_MtD loss proved to perform well in most cases, in terms of both accuracy and earliness.

Keyphrases: Event Log Data, LSTM, Predictive Process Monitoring, Remaining time models, earliness, process monitoring, temporal loss

@booklet{EasyChair:5808,
  author    = {Mike Riess},
  title     = {An Initial Study on Temporal Loss Functions for Remaining Time Models},
  howpublished = {EasyChair Preprint 5808},
  year      = {EasyChair, 2021}}