Proxy-based Web Prefetching Exploiting Long Short-Term Memory

Abstract

We propose an intention-related long short-term memory (Ir-LSTM) model based on deep learning to realize web prediction. This model draws on an LSTM model and skip-gram embedding method, and we expand the input features with user information. To maximize its potential, we propose a real-time dynamic allocation module that detects traffic bursts in real time and ensures better utilization of server resources. Experiments demonstrated that Ir-LSTM can improve the hit ratio by approximately 27% rather than hidden Markov model (HMM) and pure LSTM.