融合行为轮廓与Transformer的下一活动预测框架OA

Predicting the occurrence of the next activity in a process has become a key task in process intelligence.Existing research mainly focused on the temporal modeling of event sequences,with less attention paid to the modeling and utilization of structural behavioral relationships between activities.Therefore,a next activity prediction method was proposed.The method integrated the behavioral profile matrix and the Transformer structure,aiming to simultaneously mine the structural information and contextual semantic information in the process.The method extracted the behavioral profile matrix from the event log to describe the sequential and logical relationships between activities;used a convolutional neural network(CNN)to extract spatial pattern features from the matrix and employed a Transformer model to model the prefix event sequence;predicted the next activity through feature fusion and a multi-layer perceptron.Experiments on multiple real business process log datasets showed that the proposed method achieved superior prediction performance,verifying the effectiveness of the fusion modeling of structural information and temporal semantics.This paper provided a new modeling approach for improving prediction accuracy in complex process scenarios.