Predictive End-to-End Enterprise Process Network Monitoring

Process mining (PM) is an established process analysis method in \(\hbox {BPM}\) that involves data-driven (process model) discovery, conformance checking, and enhancement of processes (van der Aalst et al. 2011a). PM’s general idea is to gain process transparency from event log data. It is thus an approach for process analytics, particularly focusing on ex-post process diagnostics. With the advent of predictive analytics, new potentials of gaining insights from event log data have been unlocked (Breuker et al. 2016). Using these methods, \(\hbox {PPM}\) has emerged as a new subfield of PM (Márquez-Chamorro et al. 2017). \(\hbox {PPM}\) provides a set of techniques to predict the properties of operational processes, which can be arranged into two general groups (Mehdiyev et al. 2020). The first group of techniques addresses regression tasks and refers to the prediction of continuous target variables, such as the completion time of a process instance (e.g., van der Aalst et al. 2011b; Wahid et al. 2019). In contrast, the second group tackles classification tasks and refers to the prediction of discrete target variables, such as the next activity (e.g., Mehdiyev et al. 2017; Breuker et al. 2016), process violations (e.g., Di Francescomarino et al. 2016), or process-related outcomes (e.g., Flath and Stein 2018; Kratsch et al. 2020). A branch of early PPM approaches augment discovered process models with predictive capabilities but require certain model structures to support prediction tasks. Thereby, the process model is transformed into a predictive model. For example, van der Aalst et al. (2011b) introduce a technique that uses an annotated transition system with the capability to predict process completion time based on historical event log data. Another example is Rogge-Solti et al. (2013), who mine a stochastic Petri net with arbitrary delay distribution from event log data. These approaches can be described as process-aware because they utilize “(...) an explicit representation of the process model to make predictions” (Márquez-Chamorro et al. 2017, p. 4).

However, real-world processes are usually more complex than the discovered process models (van der Aalst 2011). The process-model-dependence limits the predictive power (Senderovich et al. 2019). To overcome this restriction, another, more recent branch of \(\hbox {PPM}\) approaches proposes to encode sequences of process steps as features vectors for the straightforward use of ML models. This transforms the event log’s sequential process information into a predictive model without discovering a process model. Leveraging the generalization power of ML models, sequence-encoding approaches often outperform predictive models built on top of discovered process models (Senderovich et al. 2017).

The multi-layer perceptron (MLP) is a classic NN architecture (from the class of feed-forward DNN, Goodfellow et al. 2016). that has been leveraged for \(\hbox {PPM}\). The MLP does not explicitly model temporality. As a workaround, sequential data has a two-dimensional data structure. For example, Theis and Darabi (2019) used MLPs to predict the next activities. DNNs have been applied to \(\hbox {PPM}\), due to the conceptual similarities between next event prediction and natural language processing tasks (Evermann et al. 2016). DNNs can outperform statistical (e.g., Verenich et al. 2019) and traditional ML approaches (e.g., Kratsch et al. 2020; Mehdiyev et al. 2020; Evermann et al. 2016). DNNs perform multirepresentation learning, which “(...) focuses on extracting the multiple representations from the single view of data” (Zhu et al. 2019, p. 3) and are good at unveiling intricate structures in data (LeCun et al. 2015). A popular sub-class of DNNs are recurrent neural network (RNN) approaches (Rama-Maneiro et al. 2021), including \(\hbox {LSTM}\) and gated recurrent unit (GRU) neural networks, providing the capability to capture temporal dependencies within sequences (Rumelhart et al. 1985). Another \(\hbox {DNN}\) architecture, which allows the processing of temporal patterns across short time horizon (local temporal neighborhood), is the convolutional neural network (CNN) (Zhao et al. 2017). To leverage the potential of \(\hbox {CNN}\) for \(\hbox {PPM}\), a preprocessing of sequences from temporal to spatial structure is needed. Pasquadibisceglie et al. (2019) show the validity of such a sequence preprocessing for predicting the next process activity using the helpdesk event log and BPI challenge 2012 data. Graph neural networks (GNNs) have recently been used in \(\hbox {PPM}\) because the process control flow follows a graph structure (e.g., Stierle et al. 2021) and can directly be processed through \(\hbox {GNNs}\). Beyond the four general architectural types \(\hbox {MLPs}\), \(\hbox {RNNs}\), \(\hbox {CNNs}\), and \(\hbox {GNNs}\), extensions (e.g., transformer networks with dense layers like \(\hbox {MLPs}\); Moon et al. 2021) or combinations (e.g., long-term recurrent convolutional networks; Park and Song 2020) were proposed for \(\hbox {PPM}\).

Statistical approaches in \(\hbox {PPM}\) (e.g., van der Aalst et al. 2011b; Rogge-Solti et al. 2013) start with the control flow information of event log data. This type of information is key for process predictions, as the control flow of processes describes their structure.

By using \(\hbox {ML}\), the scope of data is extended and \(\hbox {PPM}\) techniques can encode further event log information in feature vectors (e.g., Folino et al. 2012). This additional information is called process context information. It characterizes the environment in which the process is performed (Da Cunha Mattos et al. 2014; Rosemann et al. 2008), and represents, for example, information about the resource that performs an activity.

In recent years, \(\hbox {PPM}\) research has suggested \(\hbox {DL}\) architectures that integrate context information to improve prediction results (Rama-Maneiro et al. 2021). Current \(\hbox {PPM}\) approaches receive single event logs as input and do not leverage information from multiple data sources. Thereby, an event log can also contain several subprocesses, such as in the event log shared at the BPI Challenge 2012.³

Currently, there are no \(\hbox {PPM}\) techniques using multiple data sources to perform end-to-end enterprise process network predictions. Figure 2 differentiates published \(\hbox {PPM}\) techiques based on two dimensions, namely data scope and prediction method, to extract the research gap within scientific literature concerning end-to-end \(\hbox {PPM}\).

New time series forecasting techniques (e.g., Canizo et al. 2019; Mo et al. 2020; Wan et al. 2019) offer a promising way to realize such predictions through multi-headed \(\hbox {NN}\). These networks process data from each input head (e.g., from a machine sensor) individually and merge the heads’ outcomes subsequently. Motivated by this idea, we set out to adapt this method for end-to-end enterprise process networks.

The first phase specifies the problem by adapting the approach of Benscoter (2012), beginning with the problem identification at the business department or enterprise process network layer. Their approach to “identify and analyze problems in your organization” (Benscoter 2012) has a particular focus on identifying a situation’s impact on processes and workers as well as problem-relevant metrics. Subsequently, the establishment of an understanding of the interdependent processes and data sources is crucial. Within an organization’s layers, all relevant processes and data sources, which can add value to the predictive analysis task, should be identified. Then, their dependencies should be understood to identify common denominators for synchronizing heterogeneous data sources and how they relate to the organizational problem or situation. Based on this process and data understanding, the method prescribes to define the organizational objective and the type of predictive task (regression or classification).⁴

Having identified relevant processes and data sources, we next acquire and prepare input data for the \(\hbox {MH-NN}\). Data acquisition relates to activities seeking to obtain the heterogeneous data. This data is analyzed to gain insights about the data source and subsequently prepare it for the \(\hbox {MH-NN}\). The network processes each data source individually, without the need for prior aggregation and combination. We apply some standard preparation techniques (Han et al. 2011) but more generally follow the \(\hbox {DL}\) recommendation of focusing on standard \(\hbox {DL}\) architectures for feature extraction and limiting extensive preparation (LeCun et al. 2015).

As a crucial step of data preparation, \(\hbox {PPM}\) requires appropriately encoded events and sequences. Events can be encoded based on the attributes’ type. Sequences of events can be encoded as feature-outcome pairs (Van Dongen et al. 2008), n-grams of sub-sequences (Mehdiyev et al. 2020), feature vectors derived from Petri nets (Theis and Darabi 2019), or weighted adjacency matrices (Oberdorf et al. 2021a).

Designing the multi-headed \(\hbox {NN}\), we follow recent work on \(\hbox {PPM}\) methods, which move from explicit process models and traditional \(\hbox {ML}\) approaches to \(\hbox {NN}\)-based approaches (Mehdiyev et al. 2020). Yet, for some scenarios, the sequential structure of these \(\hbox {NN}\)s is not sufficiently flexible such as, if data from different sources with different dimensions are required to explain the output variable. Following Chollet (2018, p. 301), the proposed architecture for these cases is a multi-head \(\hbox {NN}\). Architectures with multiple heads use independent single-channel input heads to process each input individually. With this approach, each data source can be processed, according to its data type and structure. Head outputs are then concatenated and further processed to obtain a prediction in the output layer.

For the design of the multi-headed \(\hbox {NN}\), the method facilitates the use of a multitude of architectures (Fig. 4). In general, it distinguishes customized and state-of-the-art architectures.

For customized architectures, a combination of \(\hbox {NN}\) layers can be selected (Sect. 2.1). Following Goodfellow et al. (2016), combining various layers in a task-specific manner enables the implicit extraction of valuable features. To this end, distinct properties of architectures can be leveraged, such as the particular suitability of \(\hbox {LSTM}\) layers to process time-series or \(\hbox {CNN}\) layers for matrix data. These properties can even be combined to process time-series, such as a combination of \(\hbox {LSTM}\) and \(\hbox {CNN}\) layers (Brownlee 2017).

In addition to the customized architectures, the method taps into recent advances in the \(\hbox {DL}\) domain by incorporating established architectures. There are state-of-the-art architectures for the various domains such as image, text, or signal processing. As the numbers of available architectures are constantly changing, we suggest checking for currently available state-of-the-art networks during a model’s design phase to build on recent research advances.⁵ Figure 4 provides an overview of currently established state-of-the-art methods for various tasks. Depending on the data type, we show current \(\hbox {DL}\) solutions for problems, such as sentiment analysis (Jiang et al. 2019), language modeling (Brown et al. 2020), text, time-series, audio, image, or graph classification (Lin et al. 2021; Horn et al. 2020; Verbitskiy and Vyshegorodtsev 2021; Dai et al. 2021; Zhang et al. 2019), as well as link prediction (Wang et al. 2019), or community detection (Jia et al. 2019) in networks.

The common denominator for such models is that they consist of complex \(\hbox {DL}\) architectures with many hidden layers and trainable parameters. Because the training of such models is computationally demanding, they are usually provided with pretrained weights, which can then be leveraged for the prediction task at hand or even fine-tuned based on the task’s specific data.

The method next requires to consider aspects of model evaluation. For this purpose, we follow Brownlee (2020)’s approach, including the generation of a validation set and the use of performance metrics to assess a model’s performance. The evaluation of the resulting model is crucial for the selection of a proper configuration. It reveals whether the model is suitable to estimate the desired target variables. To this end, test and validation sets are artificially generated through validation methods. In particular, in the field of \(\hbox {PPM}\), selecting an appropriate validation set method is challenging. There are three established validation set generation methods (Fig. 3). In addition to the validation set generation, it is common to keep a holdout set containing exclusive data for a final model evaluation.

The most common method used is a straightforward strategy, referred to as a train-test split procedure (James et al. 2017, p.176–178). An alternative evaluation procedure is k-fold cross-validation for estimating the prediction error (James et al. 2017, p.181–186). It splits the data set into k folds, uses \(k-1\) of folds for training and the other fold for validation.

In some settings, regular k-fold cross-validation is not directly applicable. This is the case for time-series data, where observations are samples with fixed time intervals. The constraint is the temporal components inherent in the problem. Here, a time-series split is an appropriate method, where in the \(k^{th}\) split, the first k folds are used as a train set, and the \((k+1)^{th}\) fold is used as a test set. Time-series splits have the drawback that there is overlap between the training and testing data. This limitation can be resolved by forward testing techniques where the model is automatically retrained at each time step when new data is added (Kohzadi et al. 1996).

Based on the validation set and performance metrics, the model is trained and tuned. For effective and efficient tuning of training parameters, several software packages such as Hyperopt (Komer et al. 2019), keras-tuner (O’Malley et al. 2019), or auto-sklearn (Feurer et al. 2019), can be used. These tools instantiate intelligent search procedures (Bergstra and Bengio 2012; Snoek et al. 2012). Finally, the tuned models are tested and the learning curves evaluated, to ensure a robust model for the prediction task.

In the last phase, the method describes aspects for \(\hbox {MH-NN}\) application. This includes the operationalization of data acquisition and preparation as well as the deployment of an evaluated \(\hbox {MH-NN}\). Of particular importance is the live connection to the enterprise process network and the data sources. Instead of training on historical data, the \(\hbox {MH-NN}\) must handle live data to provide real-time predictions. Thus, besides model performance, runtime performance becomes particularly relevant during model deployment.

We collaborated with a medium-sized German manufacturing company. The firm has multiple distributed production and assembly lines for highly customized mechatronics products. Competitive pressure necessitates the firm to offer high-quality products with (mass) customization options. This combination can lead to fairly complex production processes. Here, disruptions⁶ where a worker has to interrupt work, are not uncommon.

To efficiently handle such disruptions, our cooperation partner has deployed a disruption management system (Oberdorf et al. 2021b). The system automates responder notification for solving a disruption.⁷ As a disruption is solved through the responding agent, the agent provides the system additional information, such as one of 32 disruption reasons (types). We identified the disruption’s type as a central component of the problem specification. If the type was already known, an agent could already prepare the solution process (e.g., bringing relevant tools or documentation), which reduces the disruption associated downtime.

In parallel, the production processes have been analyzed with PM techniques to identify optimization potentials. However, due to the enterprise process network’s complexity, interrelations, and dependencies, the respective analyses are very time-consuming. Consequently, the realization horizon of possible benefits is long. Striving for immediate benefit with minimal analysis effort, we adopt the \(\hbox {PPNM}\) method and provide an end-to-end \(\hbox {PPNM}\) solution. Thereby, the \(\hbox {MH-NN}\) is integrated into the organizational enterprise process network. The organizational objective is to improve the production process through better disruption handling, resulting in reduced downtime. We do so by predicting the disruption type and providing a solution suggestion to a notified agent based on the prediction. Accurate predictions are essential for meaningful notifications and suggestions.

We engaged with various departments (digitalization, logistics, and production) to evaluate the \(\hbox {PPNM}\) method in practice. Thereby, we elaborated on each department’s process event log and related databases.⁸

We compute basic statistics and advanced event log characteristics such as sparsity, variation, or repetitiveness (Heinrich et al. 2021; Di Francescomarino et al. 2017) to better understand the production and logistics event log data used (Table 1) as well as the disruption context information (Table 2). The descriptives demonstrate the high complexity of the semi-structured event logs with many unique process variants and activity types. Furthermore, we combine both event logs and obtain the combined production event log, which contains information about the logistics and production process, its control flow, and context information.

The disruption log is closely related to the intra-logistics and production departments and processes, as disruptions occur in both departments. It contains information about historical disruptions with features such as the disruption hardware id and timestamp. This way disruptions can be mapped to a workplace through the hardware device database. This enables us to retrieve product information from the respective data sources, which we can also leverage as features for the predictive task.

We follow the \(\hbox {PPNM}\) method to design a multi-head \(\hbox {NN}\): We start with the data preparation for the disruption log. Concerning the hardware id, we include additional workstation and product information using one-hot encoding. Besides, we can extract time features, such as days, weekdays, hours, and minutes, from the disruption-associated timestamp, which we subsequently normalize.

By aggregating the logistics and production log, we obtain a process event log with context information. To transform the event log into valuable features, we follow Oberdorf et al. (2021a) and select process instances within a time window, which we subsequently transform into a matrix representation. Thereby, rows and columns relate to specific workstations and the value of a distinct cell to the production quantity within the time window. For \(\hbox {NN}\) preparation, we scale each matrix by the maximum production quantity of all matrices. This process is used for the control flow data (process matrices) as well as for the context data (context matrices).

We choose a three-headed \(\hbox {DNN}\) architecture (Fig. 6 in the Appendix, available online via http://​link.​springer.​com). The disruption vector is the first input for the multi-head \(\hbox {NN}\) and is processed with an MLP (head), including a batch normalization. For both input matrices (weighted adjacency and context matrices), we use \(\hbox {CNN}\) architectures, consisting of stacked \(\hbox {CNN}\) and fully connected (FC) layers. For the context information, we apply a \(\hbox {CNN}\)-FC architecture to perform best in combination with the other heads. It consists of three \(\hbox {CNN}\)-layers and a subsequent FC layer. The third head’s design – the process event head – posts a more challenging task. We tried the architecture used for context information and appended the adjacency matrices to the context matrices in the fourth dimension.⁹ However, none of these approaches delivered satisfactory results. For this reason, we leverage process knowledge in the definition of the \(\hbox {CNN}\) kernel sizes. Basically, multiple sequential \(\hbox {CNN}\) layers extract features with distinct kernels.¹⁰ After feature extraction, both matrix head outputs have a 4D shape. To combine both with the disruption head’s output vector, we flatten the matrix head outputs. The flattened features are subsequently processed by a dense layer and the final output dense layer for the multi-class classification task.

For the quantitative evaluation, we classify the type of each disruption event with the constructed \(\hbox {MH-NN}\). In addition, we compare traditional aggregation-based approaches, where we append the disruption input vector with engineered (process) adjacency list features and, in addition, a vector of context information. Instead of 24 disruption vector features, we use 291 input features for adjacency list combination. In combination with the 267 additional adjacency list features, we use a total of 558 features.

We perform a five-time repeated five-fold cross-validation with random initialization. To prevent the \(\hbox {DNN}\) models from overfitting, we integrate an early stopping rule for validation accuracy. We store the best-performing models during each training cycle and used a Bayesian optimization algorithm (O’Malley et al. 2019) for hyperparameter tuning. Our tuning objective is the validation accuracy with a maximum retrial of 50 configurations.

For the tuned \(\hbox {FC}\), \(\hbox {CNN}\), and multi-headed (MH) models, we first compare the validation loss (Fig. 5) at the stopping time. The multi-headed approach’s loss clearly outperforms the other \(\hbox {DNN}\) architectures. In addition, it reaches a solid model with fewer epochs compared to the \(\hbox {CNN}\) or \(\hbox {FC}\) architecture with flattened feature inputs.

The final models are subsequently evaluated on the hold-out set, resulting in the metrics summarized in Table 3, where we compare basic benchmark approaches such as most frequent (mFreq) or k-nearest-neighbor (KNN) methods, as well as more advanced machine learning, deep learning, and the multi-headed architectures. All evaluated algorithms, \(\hbox {ML}\), and \(\hbox {DNN}\) models outperform the naive benchmark in terms of BMACC as well as the (weighted) F1-score, Precision, and Recall-score. We observe that the \(\hbox {FC}\) architecture benefits from the additional adjacency list features. However, we also see that the additional context list features lead to a decrease in predictive power, indicating that the \(\hbox {FC}\) architecture cannot completely prevent overfitting.

A comparison of \(\hbox {CNN}\) with only adjacency matrix features shows that they contain some basic information. However, this performance does not match the \(\hbox {FC}\) architecture with disruption and adjacency list features. The proposed multi-headed \(\hbox {NN}\) approach outperforms all benchmark architectures. Besides the better training behavior of the multi-headed \(\hbox {NN}\) approach, the higher aggregation of the data seems to result in this information loss. Due to the matrix properties, the \(\hbox {CNN}\) can identify patterns in the data that lead to improved results. Note that the resulting multi-class accuracy refers to a 32-class classification problem. Accordingly, the 81% \(\hbox {MH}\) accuracy is a good result, allowing a reliable solution suggestion.

The experimental results of the multi-headed architecture are in line with recent research in computer vision (He et al. 2016) in general and predictive process monitoring (Rama-Maneiro et al. 2021) in particular. The \(\hbox {DL}\) algorithms show superior performance for the specific use case of multi-class classification. However, the superiority of the \(\hbox {MH-NN}\) architecture in terms of predictive power is tied to some drawbacks regarding implementation and training time. Compared to the standard \(\hbox {ML}\) models, that are readily implemented using libraries such as Scikit-learn (Pedregosa et al. 2011), finding and implementing optimal \(\hbox {NN}\) architectures for each network head is a complex and time-consuming task. Additionally, the training of the multi-headed \(\hbox {NN}\) takes significantly more time.¹¹ Clearly, this is a limitation of the \(\hbox {MH-NN}\) model. For our use case, however, the prediction duration is more relevant, which is acceptable and facilitates the application of the model.

In the last phase of the \(\hbox {PPNM}\) method, we deploy data acquisition and preparation as well as the identified best model. The method’s resources are deployed on a standard commercial virtual machine with Linux OS. It is connected to the organizational enterprise process network through an MQTT connection, which enables the live interaction with the disruption management system. Whenever a disruption occurs and the worker triggers the notification process, the disruption data is transmitted through the MQTT connection and triggers the prediction process. Recent production and intra-logistic event log data are automatically obtained, and all data are prepared as well as forwarded to the \(\hbox {MH-NN}\). The prediction result is then transmitted to the disruption management system and improves the information, which a responding agent receives as part of the disruption notification. Therefore, better preparation for the disruption task at hand is possible, which ultimately reduces disruption downtimes and associated costs.

To provide an evaluation based on the real-world setting, we follow the approach described by Kraus et al. (2020) and evaluate the prediction error costs (\(\mathrm {c}_{\rm err}\)). The costs originate from the downtimes for solving a disruption. We calculate the costs based on the production environment setup across the production lines with a mean disruption rate of 1.3% per produced part and report it in a relative monetary unit (MU). To do so, we leverage a previously established study that analyzes the prediction accuracy with respect to the resulting downtimes (Oberdorf et al. 2021b). Based on our quantitative study, increasing model accuracy results in decreasing downtimes due to better information and thus preparation of the notified agents. Further, an increasing accuracy, such as for the \(\hbox {MH-NN}\), results in reduced prediction error costs. While, for example, the basic benchmark approach mFreq creates prediction error costs of about 3,246 MU, the \(\hbox {MH-NN}\) comes to prediction error costs of 695 MU.

In addition, we interviewed a data scientist and a project manager. According to the data scientist, the collaboration facilitated the awareness for the great interdependence of the processes. Clearly, processes affect each other, even across organizational borders, which the employees were aware of. However, combining these heterogeneous data sources meant great efforts. The proposed method provides a valuable tool for structured data combination across departments.

Of course, we are aware of interdependent processes, but leveraging the data was usually not practical. The multi-headed NN approaches bridge this gap, as we can further combine data without the downside of extensive aggregation. And due to the deployment, even without first searching and collecting the data.

(Data Scientist)

We presented the initial results to data scientists, project managers, and managers of the cooperation partner and discussed the practical implications. Aligned with the data scientist’s perspective, the project manager depicts the potential on an organizational scale. Beyond the digitalization, production, and logistics departments, applications to financial and controlling are of particular focus. Connections to the customer resources management (CRM) system or website user statistics may enable a better prediction of incoming orders, leading to improved production planning. In addition to better predictions, the deployment is then of special importance.

We do not just want to have the [multi-headed NN] approach, but really looked forward to deployment of services. Without deployment, we can not generate the desired value.

(Project Manager)

Through the first and last phase with particular focus on the organizational layers, we enable end-to-end analyses. Leveraging the multi-headed \(\hbox {DNN}\) architecture provides a scaleable solution to combine multiple data sources from across the organization and processes, each with specialized input heads. For the case study, we applied \(\hbox {PPNM}\) to a real-world use case and designed a three-headed \(\hbox {DNN}\) architecture with multi-log and context data input heads. Based on the numerical evaluation, combined with the employees’ feedback, we can summarize that the \(\hbox {PPNM}\) method helps guiding the development of predictive end-to-end enterprise process network monitoring.

Moreover, there are standard procedure models for data mining, such as CRISP-DM (Wirth and Hipp 2000), that someone may compare to our engineered method. Even though these procedure models work well for numerous use-cases in practical settings, they lack specifications and instructions for guiding the actual model design or combining multiple data sources, particularly considering the complex design process of a multi-headed neural network in an organizational context. For this purpose, the engineered \(\hbox {PPNM}\) establishes a more specialized perspective on defining the problem in the enterprise process network and particularly considers the combination of data sources in the design of a \(\hbox {MH-NN}\) with dedicated \(\hbox {NN}\) input heads.

Finally, considering the \(\hbox {MH-NN}\), architecture alternatives may enhance predictive power. Thus, it may be worth comparing multiple architectures for the same input. We did so during the \(\hbox {MH-NN}\) design, resulting in the design with three customized heads. However, with ongoing advances in \(\hbox {NN}\) development, new layers or even (pre-trained) state-of-the-art methods may emerge. Thus, the chosen \(\hbox {MH-NN}\) should be regularly reviewed.

The fifth phase consists of the final step of model integration and operationalization in the enterprise process network. It comprises the final online deployment, where (live) data sources are fed into the trained model for real-time predictions. With respect to the results, the prediction time of the MH model is worse compared to DL and ML or bencharmk approaches. However, for the current use-case the prediction time is satisfying, whereas it may be optimization potential for future research. Once the predictive model has been put into production, it draws on the knowledge from the historical data used for training. Deployed models inevitably face the phenomenon of structural changes in data over time, which is referred to as concept drift and usually leads to a deterioration of the prediction performance. Maisenbacher and Weidlich (2017), Denisov et al. (2018) and Spenrath and Hassani (2020) mention respective observations in various organizational \(\hbox {PPM}\) contexts. Yet, the concept drift problem is neither limited to \(\hbox {PPM}\), but also known in the more general fields of PM (Adams et al. 2021; de Sousa et al. 2021) and \(\hbox {ML}\) (Widmer and Kubat 1996).

For valid process predictions and analyses, the phenomenon of concept drift has to be detected and counteracted at an early stage. Currently the \(\hbox {PPNM}\) method, does not account for concept drift. To detect a concept drift, multiple methods are known (Seidl 2021; Kahani et al. 2021), such as local outlier detection, which can initiate retraining of the model with updated data to avoid wrong predictions and achieve temporal stability (Teinemaa et al. 2018).

A common phenomenon of traditional enterprises with hierarchical organizational structures is silo thinking. The symptoms of it are weak collaboration throughout the organization. As a result, isolated process analysis within departmental boundaries is often observed, as there is little responsibility for end-to-end processes (Eggers et al. 2021). Nevertheless, a holistic view of the organization is necessary as processes often span several departments. Connected through information systems, inter-departmental information about processes is available. In this regard, digitalization and emerging technologies, such as \(\hbox {PM}\) or \(\hbox {PPM}\), enable end-to-end insights into processes and a holistic view on the heterogeneous IT-landscape of enterprises (Armengaud et al. 2020). Both \(\hbox {PM}\) and \(\hbox {PPM}\) provide tools for generating insights on processes on an organizational scale, as they can process large amounts of data. For example, Lorenz et al. (2021) provide an end-to-end perspective for \(\hbox {PM}\) to improve the productivity in make-to-stock manufacturing processes, and Eggers et al. (2021) show how management decisions can drive an end-to-end perspective on process data by creating new process owner positions. However, the capability of end-to-end process analysis is hardly considered in research as well as in practice.

Our proposed \(\hbox {PPNM}\) method contributes to this field of research by integrating the enterprise process network with all its interrelations and dependencies. In addition, for \(\hbox {PPM}\) as a subcategory of \(\hbox {PM}\), our research has shown the benefits of taking an end-to-end view of processes for predictive tasks. The \(\hbox {PPNM}\) method and the fusion of inter-departmental data sources significantly increase the predictive power. This is already a first contribution, but it should not be the end of the research. Our approach for end-to-end \(\hbox {PPNM}\) is only an avenue towards general approaches for end-to-end \(\hbox {PM}\). Therefore, future research should focus on leveraging the resources of the enterprise process network for \(\hbox {PM}\) and derive end-to-end insights.