Predicting process behavior meets factorization machines

Predictive business process monitoring methods use datasets of completed cases related to a process to predict the behaviour of running cases. To handle the large amounts of available event data, recent works have turned to deep learning techniques and have achieved fairly accurate results. However, results from these techniques are often difficult to interpret and explain. In the area of Recommender systems, factorization models have been an important class of predictive techniques due to its scalability and ability to infer latent features. Motivated by research in Recommender systems, this paper presents a predictive model that combines matrix factorization techniques from Recommender systems and knowledge from Business Process Management to learn interactions between latent features that can be used to predict the next event of an ongoing case. Evaluation on two real-life datasets from a Dutch Financial Institute and Volvo IT Belgium shows that the approach yields results that are comparable and at times superior to state-of-the-art techniques such as neural networks, yielding at most a precision of 0.87 for next event predictions.