On-device Personalized Model Training via Local Evaluation

The patent segments the model training process into two parts: feature extraction is performed on user devices locally, while only the trained model parameters are transmitted to the server. This segmentation allows personalized modeling without transmitting raw user data, resolving the contradiction between model accuracy and privacy protection.

The patent creates a copy of the model training process that runs locally on user devices. Each device maintains its own copy of the personalized model trained on local user data, eliminating the need to transmit sensitive user information to the server while still achieving accurate personalized predictions.

The patent extracts the computationally intensive feature extraction and initial model training steps from the server and relocates them to user devices. Only the final model parameters and aggregated statistics are transmitted to the server, dramatically reducing server load while maintaining model accuracy through distributed computation.

User devices perform self-service by locally training personalized models using their own user data without requiring constant server intervention. This self-service approach reduces server load while maintaining high model accuracy through decentralized computation.

The patent segments the personalized modeling task across all user devices rather than concentrating it on the server. Each user device independently trains its own personalized model, allowing the system to scale with the number of users without increasing server load proportionally.

The patent introduces user devices as intermediaries that perform model training locally. This intermediary approach allows the server to serve many users without directly handling the computational burden of training personalized models for each user, thus scaling service coverage without linearly increasing server load.