Research Directions in Process Modeling and Mining Using Knowledge Graphs and Machine Learning

Abstract

Services Computing has seen a dramatic rise in the last twenty years. The foundation for services provided by enterprises is business processes, so progress in the development of effective and efficient processes is of utmost importance. The design or modeling of business processes is a challenging task. Over the years many research and development efforts have paid dividends, including languages and notations like the Business Process Executing Language and the Business Process Modeling Notation, along with supporting methodologies and tools. Research in Semantic Web Services and Processes showed promise for the automation of services discovery and composition (orchestration/choreography). The current large-scale deployment of enterprise knowledge graphs by many organizations coupled with huge advancements in machine learning (particularly deep learning) provides new opportunities for advancing this automation forward.

Appendix: COVID-19 as a Disease Process

A process model may be created for the COVID-19 disease as follows starting with a graph that represents the main states and state transitions.

Fig. 3.

State transition diagram for a $\textrm{SEIHURD}$ model

The states are Susceptible (S), Exposed (E), Infected (I), Hospitalized (H), Intensive Care Unit (U), Recovered (R), and Died (D). From patient data timestamps may be attached to each state. From these, durations may be attached to edges, e.g., positive test $\implies$ infected, and after say 6 d the patient is hospitalized, after 4 d, the patient is placed in the unit, etc. Similar time annotations may be used for aggregated data (at national and state levels) such as that provided by the Centers for Disease Control and Prevention (CDC), Johns Hopkins University (JHU), or Our World in Data (OWID).

Analytics may be performed by maintaining population counts for each of the vertices and transition rates (alternatively expected transition times) between vertices/states. A variety of modeling techniques may be applied, including Continuous-Time Markov Chains, Compartmental models (a system of ordinary differential equations), Vector Auto-Regressive (VAR) models, Staged Seasonal, Auto-Regressive, Integrated, Moving Average, Exogenous (SARIMAX) models, and several types of Deep Learning models.

A type of deep learning model that is showing potential for capturing much of the dynamics of the disease process is Graph Neural Network (GNN) model. A GNN can be built from Multivariate Time-Series (MTS) where each of the $n = 7$ variables forms a column in matrix Y. Given that these variables (and others) for COVID-19 have been recorded daily since January 2020, there are now about $m = 1000$ rows of data, so $Y \in \mathbb {R}^{m \times n}$. Each element in matrix Y, $y_{tj}$ where t is time/day and j designates which variable, e.g., $j = 2 \implies \textrm{I}$, forms a vertex in the GNN. Contemporaneous (at the same time) edges may be added based on the state-transition graph, they should be labeled as contemporaneous as these data/counts may not be available for forecasting. Structural learning should be performed to determine adjacency. Edges in the GNN can be added and removed by learning from data, for example, by associative or causal analysis between vertices using cross-correlation, Granger causality, or cross-mutual information. The most correlated or causal edges should be added (e.g., $\textrm{H} \rightarrow \textrm{D}$ preliminary analysis shows strong lagged effects at 8 and 14 d). A generic GNN for MTS forecasting is given in [11]; see Fig. 1 for a depiction of what such a GNN would look like. Research into the modeling and analysis of such processes can cross fertilize research into business processes.