RESEARCH · RESEARCH · #1331
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%.
KEY POINTS
- 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%.
- SimTrace offers a privacy-preserving, shareable alternative to confidential logs for building virtual clients and training user-modeling and recommendation systems.
WHY IT MATTERS
SimTrace offers a privacy-preserving, shareable alternative to confidential logs for building virtual clients and training user-modeling and recommendation systems.