Streaming Pattern Based Feature Extraction for Training Neural Network Classifier to Predict Quality of VOD services
- Title
- Streaming Pattern Based Feature Extraction for Training Neural Network Classifier to Predict Quality of VOD services
- Authors
- Pandey, S.; Hong, J.W.-K.; Yoo, J.-H.; Choi, M.J.
- Date Issued
- 2021-05
- Publisher
- Institute of Electrical and Electronics Engineers Inc.
- Abstract
- Netflix, Amazon Prime, and YouTube are the most popular and fastest-growing streaming services globally. We first studied the characteristics of the streaming patterns of these three services and then utilized these characteristics to extract several features for a quality assessment of the stream. Any streaming traffic has three main characteristics including 1) Adaptive Bit Rate (ABR) streaming 2) On-Off Cycle, 3) Buffering, and Steady-state phases. We observed that streaming providers vary in their ABR strategies and rate-controlling mechanisms. The amount of data they download in the buffering state and steady-state varies. Their On-Off cycle length and data block size vary as well. The quality of their services will depend on their streaming characteristics. Therefore we extracted 12 unique features from the streams based on these streaming patterns. We then trained a perceptron based neural network model for video quality assessment. The model was tested with 50 streams of each service, captured at varying access network bandwidth ranging from 75kbps to 30Mbps. The model could successfully identify a good and a bad stream with an accuracy of 0.929 for YouTube, 0.857 for Amazon Prime, and 0.933 for Netflix. At last, we analyzed the importance of each feature for these three services. Our approach can be used to compare and contrast the streaming services strategies and fine-tune their ABR and flow control mechanisms. ? 2021 IFIP.
- URI
- https://oasis.postech.ac.kr/handle/2014.oak/109916
- Article Type
- Conference
- Citation
- 17th IFIP/IEEE International Symposium on Integrated Network Management, IM 2021, page. 551 - 557, 2021-05
- Files in This Item:
- There are no files associated with this item.
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.