On the Identification and Prediction of Stalling Events to Improve QoE in Video Streaming

Monitoring the Quality of user Experience is a challenge for video streaming services. Models for Quality of User Experience (QoE) evaluation such as the ITU-T Rec. P.1203 are very promising. Among the input data that they require are the occurrence and duration of stalling events. A stalling even5 is an interruption in the playback of multimedia content, and its negative impact on QoE is immense. Given the idiosyncrasy of this type of event, to count it and its duration is a complex task to be automated, i.e., without the participation of the user who visualizes the events or without direct access to the final device. In this work, we propose two methods to overcome these limitations in video streaming using the DASH framework. The first method is intended to detect stalling events. For simplicity, it is based on the behavior of the transport layer data and is able to classify an IP packet as belonging (or not) to a stalling event. The second method aims to predict if the next IP packet of a multimedia stream will belong to a stalling event (or not), using a recurrent neural network with a variant of the Long Short-Term Memory (LSTM). Our results show that the detection model is able to spot the occurrence of a stalling event before being experienced by the user, and the prediction model is able to forecast if the next packet will belong to a stalling event with an error rate of 10.83%, achieving an F1 score of 0.923.