DDTR: Diffusion Denoising Trace Recovery

TL;DR

DDTR reframes the challenge of recovering deterministic traces from stochastically known process logs as an inverse problem solvable with guided diffusion. By extending Diffusion Denoising Probabilistic Models with model-free and model-aware denoisers, and by leveraging either latent process structure or its absence, DDTR achieves state-of-the-art accuracy and robustness under noise across real-world and synthetic datasets. The approach operates in the log-probability space and uses a dual-stream U-net with cross-attention to integrate trace information and process models, enabling conditioning on external signals and inverse problems. The results demonstrate up to 25% performance gains over baselines and show resilience to increasing uncertainty, signaling practical impact for reliable process mining in uncertain measurement environments.

Abstract

With recent technological advances, process logs, which were traditionally deterministic in nature, are being captured from non-deterministic sources, such as uncertain sensors or machine learning models (that predict activities using cameras). In the presence of stochastically-known logs, logs that contain probabilistic information, the need for stochastic trace recovery increases, to offer reliable means of understanding the processes that govern such systems. We design a novel deep learning approach for stochastic trace recovery, based on Diffusion Denoising Probabilistic Models (DDPM), which makes use of process knowledge (either implicitly by discovering a model or explicitly by injecting process knowledge in the training phase) to recover traces by denoising. We conduct an empirical evaluation demonstrating state-of-the-art performance with up to a 25% improvement over existing methods, along with increased robustness under high noise levels.

Figures (5)

• Figure 1: Illustration of the trace recovery process in DDTR. The recovery process begins on the left, with a random guess, denoted $\hat{x}_T$. The guess then goes through an iterative process of refinement and results in a highly accurate recovered trace $\hat{x}_0$. The trace denoted $x_0$ is the ground truth DK trace.
• Figure 2: Architecture of a U-net denoiser for trace generation. From left to right: the input trace is the result of the previous step in the reverse process, it goes through a series of convolutions and attention layers which generate features that are used to reconstruct a refined prediction of the DK trace.
• Figure 3: U-net architecture for trace generation conditioned on an SK trace. Each computation stream (blue for the denoised trace and orange for the SK trace) learns separate features, which are added in between blocks such that each block learns its features from the combined representation of the previous blocks. The bottleneck blocks do not combine representations to learn quality high level features for each stream separately. The blue computation stream on its own is identical to the one introduced in Figure \ref{['fig:unet_base']}.
• Figure 4: U-net architecture for trace generation conditioned on an SK trace and process model flow matrix. Each computation stream (blue for the denoised trace, orange for the SK trace and green for flow matrix) learns separate features which are added in between blocks such that each block learns its features from the combined representation of the previous blocks. Trace and matrix features are combined by cross attention.
• Figure 5: average accuracies of recovery methods across real-world datasets with added synthetic noise. As SK uncertainty increases, argmax accuracy sharply decreases while DDTR accuracy remains stable.

Theorems & Definitions (2)

• Example 1
• Example 2