WebFAST: Frequency-space Action Sequence Tokenization for Predictive Web Browsing Experience

We present WebFAST, adapting the frequency-space action sequence tokenization approach from robotics to web browsing behavior modeling. Inspired by recent advances in robot action tokenization (FAST) [1], we transfer frequency-domain analysis techniques from robotics to the domain of web browser user interaction modeling. Unlike existing approaches that rely on cloud-based processing or simple rule-based predictions, our method operates entirely locally within the Firefox browser while leveraging discrete cosine transform (DCT) to capture complex temporal patterns in user action sequences. In a preliminary case study, we demonstrate the feasibility of our approach through Synapse, a Firefox extension. On real-world browsing data, our evaluation shows a promising top-1 prediction accuracy of 48.6% using baseline models, validating the potential for local, privacy-preserving behavioral modeling. Our WebFAST tokenization achieves 5.0x compression ratio with 99.9% signal energy retention, enabling efficient learning of personalized browsing patterns and opening new research directions in privacy-preserving behavioral modeling and local machine learning integration in web browsers.Keywords: WebFAST, frequency-space tokenization, web browsing prediction, privacy-preserving ML, user behavior modeling, browser extensions