[PotW] Explainable and Effective Process Remaining Time Prediction Using Feature-informed Cascade Prediction Model

Paper of the week: “Explainable and Effective Process Remaining Time Prediction Using Feature-informed Cascade Prediction Model” by Na Guo, Cong Liu, Caihong Li, Qingtian Zeng, Chun Ouyang, Qingzhi Liu, and Xixi Lu. Check it out at: https://ieeexplore.ieee.org/abstract/document/10399950

Remaining time prediction is an important area of PPM as it allows to check if a running instance will meet the time constraints and to act timely. Deep learning models are generally applied due to their higher accuracy, however, at the cost of explainability.

To this end, the authors propose feature-informed cascade prediction framework. It uses a novel explainable automated feature selection strategy and a deep learning model that allows to correlate the effects between each input feature and its prediction results.

First, the effect of each feature is evaluated in a systematic manner, resulting in a so-called incremental feature tree. The tree is pruned using a threshold, and a node with the smallest MAE has the best feature set on the path to it from the root.

Second, the feature-informed cascade prediction model is constructed. Each feature gets a separate layer, and layers are trained sequentially. The input sub-layer receives its feature and the output of the previous layer. The hidden sub-layer includes a neural network,e.g. LSTM.

In their evaluation on 8 real-life event logs, they show that FCPM performs better (lower MAE) than other approaches on most logs while having medium training time. Additionally, they showcase an end-to-end application scenario with a detailed case study.

The proposed framework achieves both higher accuracy and greater explainability. In future work, the authors plan to not only use more advanced deep learning models for improved accuracy but to also develop a process-oriented feature selection approach.

I especially like the explainability of the approach,which is key to eventual acceptance of the PPM (or any other) technology. And I agree with the authors that predictive models in context of business processes require additional process-specific features. So go read the paper!