Using the data-augmented heterogeneous graph neural networks to identify risk of complex R&D projects considering project similarity

Purpose Identifying potential risks is important in project risk management, especially in complex R&D projects that are commonly implemented as project portfolio (PP). However, most of the existing data-driven risk prediction and identification methods focus on individual projects and specific risks, and there is limited research on risk prediction and identification methods that also consider the similarity between risk propagation and projects. This study aims to provide a data-driven approach for risk identification in complex R&D projects, expanding the tools used for risk prediction. Design/methodology/approach This study proposes a similarity measurement framework for R&D projects. A relational graph conventional network based on Node2vec, referred to as Node2vec-RGCN, is then utilized for data augmentation in the project risk association network, facilitating risk identification. Finally, the model is validated on real data. Findings The test results indicate that the average accuracy of our model is 70.2%, the call rate is 73.4% and the AUC indicator is 71.9%, which enables better identification of potential risks and analysis of their possible sources. Research limitations/implications In this study, for the first time, link prediction based on graph neural networks is used in project risk networks to replace guesswork in risk identification by data-driven approach. However, there are still some limitations. The first limitation is that projects have a long life cycle, and risks may occur in different project phases. Therefore, to consider the dynamics of risks and project phases, the concept of time can be added to graph neural networks. The second limitation is that there can be complex interactions between different risks, and one risk can trigger the occurrence of other risks. However, in our study, risks were treated as independent and interactions between risks were not considered. In the future, risk identification can incorporate risk interactions to make the risk identification model more comprehensive. Practical implications Specifically, it assists program managers in making decisions across three components: (1) During the project initiation phase, subsequent to the acquisition of data from the enterprise’s risk register and the project’s historical case base, it may, to a certain degree, supplant risk identification that relies on expert opinion, thereby facilitating the project in identifying potential risks and their fundamental characteristics. (2) During the project planning phase, it is also possible to conduct a risk assessment based on the risk scores generated through deep learning. This process enables the prioritization of risks and the allocation of risk response resources to address those of higher significance. (3) During the project execution phase, an updated list of potential risks and their prioritization may be generated by revising the model data according to newly generated dynamic risks. Finally, although our model is for R&D projects for risk prediction, except for the calculation of project similarity, which is mainly for R&D projects, all other data are standard features common to PP, so as long as the similarity calculation is done for projects in different industries for model adjustment, it can be applied to PP in various industries for risk identification and prediction, and it has strong transfer ability. Social implications The contribution of this study is mainly in three aspects. First, our proposed model considers the risk propagation caused by the PIs. Most existing prediction studies for project risk identification and analysis ignore the impact of complex relationships between projects on risk prediction. Therefore, our model results make the decision-making more reliable and objective. Second, the model is an effective risk management tool based on ML that can assist managers in decision-making. This study uses a real data set from a company that can help project managers identify potential risks by using data-driven instead of guesswork with records of risks that have occurred in the past and project similarities. Originality/value This study develops a hybrid risk identification model that integrates project similarity, Node2vec and RGCN, applying graph neural networks to capture risk impacts from other projects in the risk propagation of project portfolios. The results of the study replace project decision-makers’ guesses about potential risks with a data-driven approach.