AI News

Virtual clients can support A/B testing, recommender-system development, and interface evaluation, but useful online user trajectories are difficult to share. Proprietary logs raise privacy concerns, while public datasets often omit fine-grained interactions or remain small and platform-specific. SimTrace addresses this gap by anonymizing real interactions, building a simulated twin of a target web environment, and using a computer-use client agent to generate synthetic multimodal clickstreams grounded in both the trajectories and the environment. Each action is paired with web observations and user context. In an e-commerce evaluation, SimTrace exceeded competing baselines on seven of eight fidelity metrics. Models trained on its synthetic data performed comparably to models trained on real data for purchase prediction and recommendation. For next-action prediction, adding synthetic data to real data improved accuracy by 11.0% over training on real data alone. The authors release SimTrace as an open-source package for research on online behavior modeling.