Debugging Tabular Log as Dynamic Graphs

Chumeng Liang, Zhanyang Jin, Zahaib Akhtar, Mona Pereira, Haofei Yu, Jiaxuan You · 2025

Read Debugging Tabular Log as Dynamic Graphs by Chumeng Liang, Zhanyang Jin, Zahaib Akhtar, Mona Pereira, Haofei Yu, Jiaxuan You (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-2512-22903.txt.

Introduction

Chumeng Liang and co-authors tackle debugging of tabular logs, the structured records that track objects and events inside running systems. Inconsistencies in the real world show up as wrong log entries, but recent debugging tools lean heavily on large language models, which limits flexibility and scale. Their framework, GraphLogDebugger, instead rebuilds the system behind a log as a graph that evolves over time.

Objects and events become different kinds of nodes, linked by edges, so the log turns into a dynamic heterogeneous graph. On top of that representation, a comparatively simple dynamic graph neural network does the debugging. Experiments on real log datasets from computer systems and from academic publishing show the graph model debugs more accurately than language-model approaches. The result argues that recovering the right structure matters more than model size for this task.

Catalog record

CREATORChumeng Liang, Zhanyang Jin, Zahaib Akhtar, Mona Pereira, Haofei Yu, Jiaxuan You
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE52 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2512-22903

Browse the full library · About the Tower of Babel

Full text

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

Debugging Tabular Log as Dynamic Graphs Chumeng Liang, Zhanyang Jin, Zahaib Akhtar, Mona Pereira, Haofei Yu, Jiaxuan You Abstract Tabular log abstracts objects and events in the real-world system and reports their updates to reflect the change of the system, where one can detect real-world inconsistencies efficiently by debugging corresponding log entries. However, recent advances in processing text-enriched tabular log data overly depend on large language models (LLMs) and other heavy-load models, thus suffering from limited flexibility and scalability. This paper proposes a new framework, GraphLogDebugger, to debug tabular log based on dynamic graphs. By constructing heterogeneous nodes for objects and events and connecting node-wise edges, the framework recovers the system behind the tabular log as an evolving dynamic graph. With the help of our dynamic graph modeling, a simple dynamic Graph Neural Network (GNN) is representative enough to outperform LLMs in debugging tabular log, which is validated by experimental results on real-world log datasets of computer systems and academic papers. 1 Introduction Tabular log data plays a crucial role in representing and tracking the state and evolution of real-world systems. These logs are structured as rows of log entries, each capturing an event involving certain objects and their attributes at a specific time point. Common examples include system logs recording computing services ( 86 ) , research logs tracking scientific publication activities ( 9 ) , and interaction logs from multi-agent systems powered by large language models (LLMs) ( 79 ) . Debugging of tabular logs is essential: it allows practitioners to detect anomalies in the original systems through efficient inspection of associated log records. Log anomaly detection ( 23 ) has therefore been a long-standing research field in different niche areas, where data distributions are invariant or have little change. Existing frameworks ( 11 ; 43 ; 82 ; 50 ; 20 ; 7 ) benefit from manually defined data structures or templates for log parsing which are often tailored to certain domains and thus yield absolute success in specific areas like computer system log or financial event log. However, due to this domain-specific principle, designing a general-purpose log debugger always remains challenging. Efforts to overcome this challenge have led to two main lines of work, as shown in Figure 1 . One stream focuses on graph modeling of the log data ( 8 ; 75 ; 47 ) , where information in log is gathered in a unified data structure: the graph, such as constructing knowledge graphs or text-rich dynamic graphs for computer system log ( 60 ; 37 ) . Although these methods are both efficient and powerful, many of them lack flexibility: they still customize static graph structures for certain domains. Another stream explores LLM-based solutions, such as LLM prompting ( 74 ; 51 ; 49 ) or retrieval-augmented generation (RAG) ( 45 ; 77 ; 63 ) pipelines. While these methods demonstrate general capabilities in text-based reasoning, thus showing potential of generalization, they often come with significant drawbacks: high computational costs, slow inference, and difficulty scaling to long log streams or resource-constrained settings. Inspired by the idea to unify multimodal information in dynamic graphs ( 17 ) , we propose GraphLogDebugger , a general and efficient framework for debugging tabular logs through dynamic graph modeling. Our core idea is to interpret tabular log entries as the evolving state of a hidden system, which can be reconstructed as a dynamic heterogeneous graph. We treat objects and events as different types of nodes with text embeddings empowered by modern language embedding models, and use the tabular structure to generate time-stamped connections between them. As new log entries arrive, they incrementally update the dynamic graph, capturing both structural and temporal dependencies. This formulation allows us to apply a lightweight dynamic Graph Neural Network (GNN) to perform online anomaly detection by evaluating the likelihood of new connections. Our approach avoids reliance on heavy LLMs while still capturing rich semantic and relational information in the data. Experimental results on real-world datasets from computer system logs and scientific publication logs validate the effectiveness of our approach. Despite its simplicity, our dynamic GNN framework outperforms LLM-based baselines in both accuracy and efficiency, demonstrating that dynamic graph modeling is a highly expressive yet lightweight alternative. Our contributions can be summarized as follows: • We introduce a novel view of tabular logs as dynamic heterogeneous graphs, bridging the gap between structured attributes and semantic reasoning, and redefine the framework of online log anomaly detection, where object-event connections in each incoming log are evaluated through link prediction on the evolving graph. • We propose a lightweight GNN-based debugger that can efficiently and accurately detect anomalies without using LLMs, and validate its performance on real-world datasets with diverse modalities. Figure 1: Comparing GraphLogDebugger with two existing lines of works . Processing log with domain-specific graphs requires custom text parsing, which lacks flexibility. LLM-based log processing overcomes this shortcoming by the general comprehension skills of LLMs, but suffers from poor efficiency. GraphLogDebugger combines the advantages of graph representation and those of LLMs and balances well generalizability and scalability. 2 Related Works Tabular Log Processing. Many real-world logs include structured, time-stamped tabular attributes alongside annotated text fields. Examples come from financial prices paired with event series ( 61 ; 54 ; 10 ) , scientific publication metadata ( 9 ; 28 ) , healthcare records ( 27 ) , and computer ( 86 ) and multi-agent system reports ( 79 ) . A key challenge in processing tabular logs lies in capturing multi-attribute correlations while maintaining comprehension of their semantics ( 66 ) . One common approach integrates main attributes recognized by human priors into structured data ( 70 ; 84 ; 30 ) , and then subsequently augments the representation by retrieval ( 32 ; 67 ) . This modeling achieves good performance in domain-specific data, but lacks flexibility and generalizability for adaptation to other fields ( 18 ) . An emerging alternative leverages Large Language Models (LLMs) ( 4 ) , which have demonstrated strong generalizability in understanding, predicting, and generating tabular data ( 38 ; 80 ; 16 ; 65 ) . By parsing diverse logs into a unified format with LLMs ( 85 ) , these models can be applied to downstream tasks that require reasoning capabilities, such as predicting stock prices ( 74 ) , electricity demand ( 64 ) , and future events ( 58 ; 72 ) . However, LLM-based approaches often suffer from high overheads, complex deployment, and limited throughput. There remains a strong need for lighter-weight alternatives with comparable performance. Dynamic Graphs. Graph Neural Networks (GNNs) have become a foundational paradigm for learning on graph-structured data ( 29 ; 21 ; 19 ) . Static GNN models have benefited from advances in message passing ( 3 ) , architectural depth ( 35 ; 14 ) , and inductive scalability ( 21 ) . However, many real-world systems are dynamic, motivating models that capture both structural and temporal dependencies. Early approaches used recurrent layers or time-aware embeddings ( 36 ; 57 ) to extend static GNNs to dynamic settings ( 48 ; 55 ; 31 ) . Recent methods have embraced memory modules ( 53 ) and temporal encoding

Read the complete text