MMAD: Markov Model-based Adaptive Denoising Method for User Interaction Logs

Real-world user interaction logs record various actions between users and applications, which can be used to identify patterns and optimize processes in robot process automation systems.However, they often contain a lot of noise, posing a substantial challenge to the performance and accuracy of robotic process automation mining systems.Existing noise removal approaches typically recognize infrequent actions as noise, and prior knowledge is needed to determine thresholds artificially.These issues result in poor generality of existing methods, making them ineffective in removing log noise across different scenarios.This paper proposed an adaptive denoising method for log templates based on a Markov model, called MMAD.It first identifies user-specific actions and then extracts them to generate a general log template.Then, it constructed the template association graph, and employed the Markov model to analyze state transition relationships.Finally, an adaptive threshold is calculated based on the truncated normal distribution probability statistical method.Experiments proved the effectiveness of MMAD in denoising two real datasets in two scenarios, without relying on prior knowledge of the logs.The average precision of this method is 94.31%, the average recall is 97.71%, and the average F1-score is 95.64%.Compared to the baseline methods, there was a 3.3% increase in the average F1-score.