SimTrace: open-source framework to generate grounded multimodal synthetic clickstreams (arXiv:2609.38397v1)
SimTrace is a framework that anonymizes real interactions, builds a simulated twin of a web environment, and generates fine-grained multimodal synthetic clickstreams grounded in that data; it is released as an open-source package in arXiv:2609.38397v1. In an e-commerce evaluation SimTrace outperforms baselines on 7 of 8 fidelity metrics, produces models whose downstream performance is comparable to models trained on real data, and augmenting real data with SimTrace synthetic traces improved next-action prediction accuracy by 11.0%.