Towards Accurate Context-aware User Simulation for Algorithm Auditing

Auditing context-aware mobile algorithms for transparency and fairness is crucial for the safety of online personalizing platforms. Current auditing methods primarily employ black-box interventions that simulate user interactions such as clicks, shares, or skips. However, these approaches overlook contextual signals, specifically, physical activities like exercising, taking public transportation, and working, that are increasingly utilized by online platforms. Therefore, we propose \name, the first con\textbf{T}ext-awa\textbf{R}e \textbf{U}ser \textbf{S}imula\textbf{T}ion framework designed to audit context-aware mobile algorithms. Our approach features a triple-layer user modeling framework comprising user group, role, and activity layers to comprehensively model user-activity relationships. We then employ Markov chains to model activity transitions, followed by an RNN-based variational autoencoder that generates realistic sensor-level activity signals. Adversarial training of the model ensures that the synthetic data accurately replicate the distribution of real sensor inputs. Experiments on 28$k$ human-collected real-world samples show that the generated data achieves 93\% accuracy in mobile activity recognition, approaching real-sample performance (98\%). The dataset for replication is available at \textcolor{blue}{\url{https://github.com/kokoroqwq/sensor\_dataset\_ADF5}}.