Abstract
Techniques are described for multimodal simulation of a user browsing session through an ordered content feed. A feed is rendered into a visual representation, and recent engagement events are encoded into a live session state vector that includes an emotional state embedding inferred from behavior. A multimodal autoregressive model consumes the rendered feed and the session state to predict a trajectory across feed positions, including predicted interactions and predicted emotional-state shifts conditioned on prior positions. A planner and/or safety checker uses the predicted trajectory to select among candidate orderings or to block a slate and serve a fallback ordering when a risk condition is predicted. Training may include teacher-student distillation, counterfactual rendered orderings, and reinforcement learning using the simulator as an environment.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Anonymous, "Multimodal Session Simulation with Sequential User State Evolution for Content Ranking", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/10628