Graph-based Event Log Repair

1 st Sebastiano Dissegna, 2 nd Chiara Di Francescomarino, 3 rd Massimiliano Ronzani · 2025

Read Graph-based Event Log Repair by 1 st Sebastiano Dissegna, 2 nd Chiara Di Francescomarino, 3 rd Massimiliano Ronzani (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-05145.txt.

Introduction

Process mining reads the event logs that business software leaves behind, but real logs have holes: missing actors, times, or activity names. Three authors repair those holes with a graph neural network built for mixed data types. Each trace becomes a graph, and the network learns from complete traces how the missing pieces of a broken one should look.

Older repairs either need a model of the process or fix only a few kinds of fields. The graph approach aims at all event attributes at once, and the rich graph encoding captures how events in a trace relate. Tested against a leading autoencoder method on two synthetic logs and four real ones, with several kinds of values knocked out. The new model reconstructed the missing data very well across the board. The authors present it as a general repair tool for messy real-world logs.

Catalog record

CREATOR1 st Sebastiano Dissegna, 2 nd Chiara Di Francescomarino, 3 rd Massimiliano Ronzani
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE41 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-05145

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Graph-based Event Log Repair 1 st Sebastiano Dissegna, 2 nd Chiara Di Francescomarino, 3 rd Massimiliano Ronzani Abstract The quality of event logs in Process Mining is crucial when applying any form of analysis to them. In real-world event logs, the acquisition of data can be non-trivial (e.g., due to the execution of manual activities and related manual recording or to issues in collecting, for each event, all its attributes), and often may end up with events recorded with some missing information. Standard approaches to the problem of trace (or log) reconstruction either require the availability of a process model that is used to fill missing values by leveraging different reasoning techniques or employ a Machine Learning/Deep Learning model to restore the missing values by learning from similar cases. In recent years, a new type of Deep Learning model that is capable of handling input data encoded as graphs has emerged, namely Graph Neural Networks. Graph Neural Network models, and even more so Heterogeneous Graph Neural Networks, offer the advantage of working with a more natural representation of complex multi-modal sequences like the execution traces in Process Mining, allowing for more expressive and semantically rich encodings. In this work, we focus on the development of a Heterogeneous Graph Neural Network model that, given a trace containing some incomplete events, will return the full set of attributes missing from those events. We evaluate our work against a state-of-the-art approach leveraging autoencoders on two synthetic logs and four real event logs, on different types of missing values. Different from state-of-the-art model-free approaches, which mainly focus on repairing a subset of event attributes, the proposed approach shows very good performance in reconstructing all different event attributes. I Introduction Process Mining (PM) is an interdisciplinary field that combines data science and process management, enabling organizations to analyze and improve their operational processes using data extracted from event logs. However, the reliability and accuracy of any PM analysis fundamentally depend on the quality and completeness of the underlying event logs [ 1 , 2 ] . In practice, generating high-quality event logs is far from straightforward. Various challenges might arise due to the manual nature of certain activities, inconsistent data recording practices, or technical limitations during data acquisition. As a result, event logs often contain missing or incomplete information, especially related to individual event attributes. Hence the need for methods able to repair event logs and reconstruct the missing pieces of a trace. Traditional approaches to event log (or trace) repair in case of missing data generally fall into two main categories. The first one relies on the availability of a process model, which is then used to infer missing values through logical or statistical reasoning. The second utilizes Machine Learning (ML) or Deep Learning (DL) techniques to learn patterns from complete cases and apply them to repair incomplete traces. While these methods have shown promise, they often struggle to capture the full complexity of inter-attribute dependencies. This holds since most approaches only focus on a subset of event attributes, disregarding secondary ones, which might also contain relevant pieces of information. Recently, Graph Neural Networks (GNNs) have gained attention [ 3 ] for their ability to process data represented as graph structures that naturally capture the relationships and dependencies within the encoded trace attributes. In the context of PM, GNNs provide an expressive framework for representing execution traces, where events and their attributes can be modeled as nodes and edges in a graph [ 4 , 5 ] . In this work, we introduce \mathsf{{SANAGRAPH}} , which aims to tackle the problem of repairing event logs with missing information, with a focus on cases where all the attributes of an event are missing. This is achieved by leveraging a semantically rich encoding as a heterogeneous graph [ 6 ] , which, compared to its homogeneous counterpart, allows us to give a specific node type to each event attribute in the trace. This turns the repair problem into a node classification task, well-known in the GNN literature [ 3 ] . The Heterogeneous GNN (HGNN) model allows for information to circulate across the graph in input and enrich the nodes marked as empty, which represents the missing information. A set of linear layers, one for each attribute, classify then the empty nodes into their respective repaired values. We evaluate our approach on two synthetic and four real-world event logs, with various missing data patterns. In contrast to existing state-of-the-art methods, most of which consider only a subset of missing attributes, our model demonstrates strong performance in reconstructing the complete set of attributes, showing its potential for enhancing data quality in PM applications. The paper is structured as follows. In Section II , the background theory on PM and GNNs is introduced. Section III contains an analysis of the current state-of-the-art approaches. In Section IV , the heterogeneous graph encoding of traces, along with a description of the HGNN model employed and how it is trained is reported. Section V and Section 4 discuss how the evaluation was carried out and the obtained results, respectively. Finally, Section 4 concludes the paper and discusses future directions. II Background In this section we discuss the background theory on Process Mining and Graph Neural Networks. II-A Process Mining In Process Mining [ 7 , 2 ] , process executions can be represented in event logs, which are a description of the events in the process of interest. Definition 1 (Event Log) An event log \mathcal{E} is a collection of events that correspond to the selected process. Each event e\in\mathcal{E} has a series of attributes a\in\mathcal{AT} , where \mathcal{AT} is the set of all possible attributes. The three main attributes are as follows: \bullet Case_{ID} ( e^{id} ): this identifies the unique case a specific event belongs to; \bullet activity ( e^{act}\in\mathcal{A} ): this is the activity (in the set of activities \mathcal{A} ) performed in that event; \bullet timestamp ( e^{ts} ): this indicates the time at which the event occurred. CaseID Activity time:timestamp org:resource lifecycle:transition case:REG_DATE case:AMOUNT_REQ 173688 A_SUBMITTED 2011-09-30 22:38:44.546000+00:00 112 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 A_PARTLYSUBMITTED 2011-09-30 22:38:44.880000+00:00 112 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 A_PREACCEPTED 2011-09-30 22:39:37.906000+00:00 112 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 A_ACCEPTED 2011-10-01 09:42:43.308000+00:00 10862 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 A_FINALIZED 2011-10-01 09:45:09.243000+00:00 10862 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 TABLE I: Example of a trace belonging to the BPI 2012 dataset CaseID Activity time:timestamp org:resource lifecycle:transition case:REG_DATE case:AMOUNT_REQ 173688 - 2011-09-30 22:38:44.546000+00:00 112 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 A_PARTLYSUBMITTED - 112 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 - - - - - - 173688 A_ACCEPTED 2011-10-01 09:42:43.308000+00:00 10862 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 173688 - 2011-10-01 09:45:09.243000+00:00 10862 COMPLETE 2011-10-01 00:38:44.546000+02:00 20000 TABLE II: Example of a trace with missing attributes and events Along with the three aforementioned main attributes, each event may contain additional (categorical or numerical) attributes. Examples of attributes belonging to an event include the resource that has carried out the activity, e.g., the person responsible for it; the cost of executing a certain event; or information about how much time has elapsed since the start of the process or since the previous event occurred. An event log is a collection of traces that are

Read the complete text