Bidirectional LSTM for Continuously Predicting QoE in HTTP Adaptive Streaming

Continuous monitoring of quality of experience (QoE) has increasingly become an important mechanism in quantifying the user's satisfaction over video streaming services. Such mechanism requires approaches with the ability to model the complex temporal dependencies and time-varying characteristics of the QoE. Long-Short Term Memory (LSTM) related approaches carried out by recent research efforts have shown highly potential results since they can leverage past events occurring during a streaming session. However, the unidirectional LSTM structure only consider forward dependencies, leading to high possibility of missing out useful information. In this paper, a novel model for continuous QoE prediction is proposed to consider not only forward dependencies but also backward dependencies. The proposal utilizes a Bidirectional Long Short Term-Memory (BLSTM) model to process inputs obtained from perceptual video quality algorithms, rebuffering, and memory-related temporal data. Comparisons with other state-of-the-art models indicate that the proposed model achieves very promising performance in terms of accuracy.