URL Categorization Classifier With Noise-Elimination Training

The URL categorization system is divided into multiple independent components: a training module that processes training data and generates training outputs, a classifier module that performs actual URL classification, and a policy enforcement module that applies security policies. This segmentation allows each component to be optimized independently and deployed separately, reducing overall system complexity while maintaining high classification accuracy through specialized processing in each module.

The patent introduces intermediary components such as a data preprocessing layer that cleans and prepares training data before it reaches the classifier, and a post-processing layer that refines classification results before policy enforcement. These intermediaries act as buffers that decouple the complex training process from the classification task, allowing the system to achieve high accuracy without proportionally increasing the complexity of the core classification engine.

The system performs preliminary actions by pre-processing training data before classification training, including cleaning, normalization, and feature extraction. This preliminary preparation of data ensures that the classifier receives high-quality input, improving convergence speed and final accuracy while reducing the number of training iterations needed. The training outputs are also pre-processed and structured in advance to facilitate faster model training.

The patent employs parameter changes by adjusting training hyperparameters, data sampling strategies, and model architecture parameters to optimize the training process. By dynamically modifying training parameters such as learning rates, batch sizes, and data selection criteria, the system achieves high classification accuracy with reduced training time and computational resource consumption compared to fixed-parameter training approaches.