PaperlayerView exact revision on arXiv

Source-linked AI summary

Leveraging contextual events on structure-aware next activity prediction

Alessandro Mele, Claudia Diamantini, Domenico Potena

arXiv:2609.08622v1cs.LGcs.AI

TL;DR

Next activity prediction often overlooks the concurrent process instances that provide execution context. The paper represents cases with Instance Graphs, encodes contextual instances through multiple strategies, and evaluates them with Graph Neural Networks; experiments show that contextual modeling improves performance, especially when process and context are separated into graphs.

  • Problem

    Existing next activity prediction approaches rarely model the contextual process instances running concurrently, despite cases with the same state potentially evolving differently under different workloads.

  • Method

    The approach builds Instance Graphs and prefix-Instance Graphs, applies multiple encodings for contextual process instances, and provides the resulting representations to a Graph Neural Network.

  • Results

    The best overall performance is achieved by strategies separating prefix-Instance Graphs from contextual process instances, outperforming multiple graph-based approaches.

  • Takeaways & Limitations

    Explicitly separating process history from contextual information better captures contextual dependencies within the evaluated next activity prediction setting.

  • Takeaways & Limitations

    Contextual events are identified solely through temporal overlap, without considering resources, activity types, or causal relationships among process instances.

Abstract

from arXiv · show

Predictive process monitoring aims at forecasting various aspects of running processes. Among the different tasks, next activity prediction represents the most extensively investigated. However, only a limited number of existing approaches explicitly encode contextual information, i.e., the environmental conditions in which the process is executed, typically modeled through event log attributes or aggregated measures. In this paper, an approach based on the concept of Instance Graphs is introduced. To incorporate contextual process instances, several encoding strategies are proposed and evaluated by measuring their impact on prediction performance. For each encoding strategy, a set of prefix-Instance Graphs is generated and subsequently provided as input to a Graph Neural Network for the classification task. The proposed approach is evaluated on multiple real-world event logs, and the experimental results demonstrate that incorporating contextual process instances benefits prediction performance.

1 Introduction

The study addresses next activity prediction by modeling contextual process instances that may alter how otherwise similar cases evolve. It introduces Instance Graph-based encodings and evaluates their effect on Graph Neural Network performance.

  • Motivation: Contextual process instances can change the evolution of cases that share the same process state.The introduction illustrates this with help desk tickets whose delays depend on concurrent workload.
  • Research objective: The study investigates whether modeling contextual process instances benefits next activity prediction.Next activity prediction forecasts the next activity for a running process instance.
  • Approach: The proposed approach uses Instance Graphs and several encoding strategies to model contextual process instances.Their impact on Graph Neural Network performance is systematically evaluated.
  • Research questions: RQ1 asks which encoding strategy achieves the best prediction performance, while RQ2 compares the best strategy with graph-based approaches.These questions define the paper’s evaluation focus.

2 Related work

Prior work uses graph representations and engineered inter-case features to capture process structure and concurrent-case information. However, existing graph-based work does not investigate how alternative strategies for incorporating contextual instances affect performance.

  • Graph-based approaches: Graph-based predictive process monitoring represents control-flow and can capture parallelism and non-linear dependencies.Prior approaches use Instance Graphs, contextual aggregates, or heterogeneous graph representations.
  • Research gap: The impact of different strategies for incorporating contextual instances is not investigated in the cited graph-based work.This identifies the related-work gap addressed by the paper.
  • Inter-case information: Existing methods incorporate inter-case information through features derived from concurrently running cases.These features support predictions such as remaining time, trace, and runtime.

3 Methodology

The methodology converts event-log traces into Instance Graphs, enriches them with temporal and contextual features, derives prefix-Instance Graphs, and feeds alternative encodings to a Graph Neural Network for prediction.

  • Methodology pipeline: The pipeline starts from event-log traces and a mined process model, then builds Instance Graphs and enriches them with temporal and contextual features.The resulting prefix-Instance Graph representations are used for prediction.
  • Instance Graphs: An Instance Graph is a directed, acyclic graph whose nodes are events and whose edges model causal relations between process activities.The representation explicitly captures dependencies within a specific process execution.
  • Illustration: Table 1 provides an excerpt of the Helpdesk event log used to illustrate the graph construction.The corresponding Instance Graph is shown in Figure 1.
  • Illustration: Figure 1 contrasts the Instance Graph derived from Table 1 with a prefix-Instance Graph of size 3.The figure illustrates the full execution graph and its partial-process subgraph.

3.2 Feature engineering

Feature engineering enriches Instance Graphs with explicit temporal information and contextual perspectives derived from concurrently executing process instances. The contextual events are identified for events in the process under analysis.

  • Temporal preparation: Each event is required to have explicit start and end timestamps before temporal and contextual features are added.A procedure from prior work supplies missing timestamps when necessary.
  • Temporal features: Temporal features describe elapsed time since the predecessor, time since process start, and position within the working week.These features are derived from event timestamps and data payloads.
  • Contextual features: Contextual feature engineering draws on concurrently executing process instances to capture contextual perspectives.The formal setup uses the set of Instance Graphs, events, activities, and event timestamps.
  • Contextual-event identification: Figure 2 illustrates a case under analysis and the event for which contextual events are identified.The example connects the analyzed case with its contextual-event selection.

3.3 Data encoding

The approach derives prefix-Instance Graphs from partial process executions and encodes contextual events using multiple graph-organization strategies.

  • Data encoding: Each prefix-Instance Graph contains the first k events of a process instance and is labeled with the activity at position k + 1.The label represents the next activity to predict.
  • Data encoding: Contextual events are represented either within the prefix graph or in a separate context graph.The first two strategies use one graph, while the remaining strategies use separate prefix and context graphs.

Encoding strategy 1 (E1)

The first encoding strategy augments the prefix-Instance Graph with contextual-event nodes linked from its first node, while the second starts those contextual events at the current prefix endpoint.

  • Encoding strategy 1 (E1): E1 adds one node for each contextual event and connects it from the first node of the prefix-Instance Graph.This represents contextual events as starting simultaneously with the prefix's first event.
  • Encoding strategy 1 (E1): E2 connects contextual nodes from nodes preceding the last prefix node, treating contextual events as starting simultaneously with the current event.E2 is more constrained than E1 because contextual events are assumed to start simultaneously with ek.

Encoding strategy 4 (E4)

E4 extends a context graph with complete contextual prefix-Instance Graphs, while retaining a connected-component representation and explicitly modeling temporal features relative to prior work.

  • Encoding strategy 4 (E4): E4 adds each contextual case's prefix-Instance Graph to the context graph and links it from the first event of the current prefix.When several contextual events belong to one case, only the prefix containing all the others is retained.
  • Encoding strategy 4 (E4): The context graph is represented as a single connected component, and temporal features are explicitly modeled compared with the approach in.These are the two stated differences from the related approach.

3.4 Graph Neural Network

The model uses spatial Graph Convolutional Neural Networks for graph classification, combining node features and graph structure into graph-level representations for prediction.

  • 3.4 Graph Neural Network: Each graph encodes node activities with one-hot features and three temporal features, while its adjacency matrix represents graph connectivity.The graph is described by a feature matrix X and an adjacency matrix A.
  • 3.4 Graph Neural Network: Graph convolution performs message passing between neighboring nodes using learnable self and neighbor transformations.The neighborhood is represented by a random sample of each node's neighbors, and the activation uses a Rectified Linear Unit.
  • 3.4 Graph Neural Network: A readout function produces a global graph representation that is passed to an MLP and Softmax classification layer.For separate graphs, prefix and context embeddings are produced separately, concatenated, and supplied to the MLP.

4 Experiments

The experiments compare contextual-instance encoding strategies and graph-based competitors across real-world event logs. E4 is selected for comparison, and the proposed approach outperforms the selected graph-based approaches on the reported metrics.

  • Experimental setup: The evaluation uses chronological case splits, Optuna-based GNN tuning, and five independent runs per configuration across publicly available real-world event logs.The first 67% of cases are used for training, the remaining 33% for testing, and the last 20% of training cases for validation.
  • RQ1: Encoding strategies: No single encoding strategy clearly dominates all datasets, while E2 performs worst and E1 improves over E2 by 4.95% in Acc, 2.89% in F1-M, and 4.58% in F1-W on average.E3 is almost always second best, whereas E4 achieves the best overall performance with only negligible improvement over E3.
  • RQ1: Encoding strategies: E4 is selected for competitor comparison because it achieves the lowest validation loss on three of five datasets.E4 models the entire history of contextual events.
  • RQ2: Competitor comparison: The selected encoding strategy outperforms both the baseline and the aggregated-context approach.The comparison uses Accuracy, Macro F1-score, and Weighted F1-score reported as means and standard deviations over five runs.
  • RQ2: Competitor comparison: Compared with, the proposed approach improves average Acc by 2.70%, F1-M by 4.57%, and F1-W by 3.16%.Compared with heterogeneous-graph approach, the average improvements are 7.75% in Acc, 11.76% in F1-M, and 7.90% in F1-W.
  • RQ2: Competitor comparison: Excluding prefix-IG sizes retaining only 5% of samples, the proposed approach achieves F1-W performance equal to or better than.F1-W becomes increasingly unstable for larger prefix-IG sizes because the dataset contains fewer large prefix-IGs.

5 Conclusions and future work

The paper presents a graph-based next activity prediction approach that models process executions and contextual process instances through multiple encoding strategies. Results favor separating prefix-IGs from contextual process instances, while contextual-event identification remains limited to temporal overlap.

  • Conclusions: The approach models process executions and contextual process instances for next activity prediction using multiple encoding strategies.The strategies are evaluated for their impact on prediction performance.
  • Conclusions: No single encoding strategy consistently dominates across all datasets.The best overall performance is achieved by strategies that represent prefix-IGs and contextual process instances as separate graphs.
  • Conclusions: Separating process history from contextual information better captures contextual dependencies within the reported experiments.This conclusion follows from the strongest performance of strategies using two separate graphs.
  • Future work: Contextual events are identified solely through temporal overlap, without considering resources, activity types, or causal relationships among process instances.Future work includes studying thresholds for selecting contextual events and cases.
Loading 2609.08622v1…