Abstract
Current approaches to avoid the drop in engagement with notifications have a high overhead, are slow to roll out, and lack content personalization. This disclosure describes techniques, implemented with user permission, to automatically select and manage content for push notifications with an automated closed-loop self-healing personalized architecture. The approach can be implemented with a Contextual Multi-Armed Bandit (CMAB) Reinforcement Learning (RL) model with reward signals configured to trigger a Large Language Model (LLM) to generate, validate, and inject fresh personalized notification strings in the content pool. The pipeline to generate structured content to replace previously pushed notification content can be automatically triggered based on relevant engagement metrics. The dynamic notification strings generation can be based on user profile, contextual information, and/or historical interaction signals. The notification serving policy can be dynamically updated based on evaluation of metrics of user interaction. Notification strings can be selected by considering long-term user experience and retention by employing an asymmetric, weighted reward function to disfavor content likely to trigger immediate clicks but create longer term annoyance and fatigue.
Creative Commons License
This work is licensed under a Creative Commons Attribution 4.0 License.
Recommended Citation
Jha, Sumit Kumar; Gambhir, Ishan; and Kumawat, Shubham, "Notification String Optimization Using Reinforcement Learning and LLMs", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/11528