Structurally Valid Log Generation using FSM-GFlowNets

TL;DR

This paper introduces a novel framework combining FSMs and GFlowNets to generate structurally valid, diverse, and realistic synthetic event logs, improving over existing methods in coherence and controllability.

Contribution

The paper presents a new FSM-constrained GFlowNet approach for generating valid and diverse event logs, with a hybrid reward system and dynamic action masking, applicable across symbolic sequence domains.

Findings

Produces realistic, structurally consistent logs with low divergence metrics.

Outperforms GPT-4o and Gemini in distributional similarity metrics.

Synthetic logs enable effective downstream classification tasks.

Abstract

Generating structurally valid and behaviorally diverse synthetic event logs for interaction-aware models is a challenging yet crucial problem, particularly in settings with limited or privacy constrained user data. Existing methods such as heuristic simulations and LLM based generators often lack structural coherence or controllability, producing synthetic data that fails to accurately represent real world system interactions. This paper presents a framework that integrates Finite State Machines or FSMs with Generative Flow Networks or GFlowNets to generate structured, semantically valid, and diverse synthetic event logs. Our FSM-constrained GFlowNet ensures syntactic validity and behavioral variation through dynamic action masking and guided sampling. The FSM, derived from expert traces, encodes domain-specific rules, while the GFlowNet is trained using a flow matching objective with a hybrid reward balancing FSM compliance and statistical fidelity. We instantiate the framework in the context of UI interaction logs using the UIC HCI dataset, but the approach generalizes to any symbolic sequence domain. Experimental results based on distributional metrics show that our FSM GFlowNet produces realistic, structurally consistent logs, achieving, for instance, under the real user logs baseline, a KL divergence of 0.2769 and Chi squared distance of 0.3522, significantly outperforming GPT-4o's 2.5294/13.8020 and Gemini's 3.7233/63.0355, alongside a leading bigram overlap of 0.1214 vs. GPT 4o's 0.0028 and Gemini's 0.0007. A downstream use case intent classification demonstrates that classifiers trained solely on our synthetic logs produced from FSM-GFlowNet achieve competitive accuracy compared to real data.