Neural Network App Inference From Session and Time Data

The patent replaces traditional rule-based or statistical methods with a neural network-based intelligent system. The neural network processes session information including application identifiers, time stamps, and user interaction data to predict next applications with high accuracy. This substitution of mechanical/algorithms with intelligent systems resolves the contradiction by achieving superior prediction accuracy while accepting increased system complexity.

The patent transforms raw session information into embedding vectors that capture temporal and contextual patterns. By changing the parameter representation from discrete application IDs to continuous embedding spaces, the system enables the neural network to learn complex usage patterns and predict applications with higher precision, thereby resolving the accuracy-complexity contradiction.

The patent performs preliminary actions by collecting and preprocessing session information in advance, organizing it into structured formats with embeddings before actual prediction is needed. This pre-processing and pre-embedding of data reduces the computational burden during real-time prediction, allowing high recommendation quality without excessive training time during deployment.

The patent implements a hybrid approach where the neural network is trained on partial datasets initially to achieve baseline performance, then progressively refined with more data. This partial training strategy enables the system to provide quality recommendations sooner while continuing to improve over time, balancing training time investment with recommendation quality returns.