Jade: A New Approach to Video Streaming Quality

Video streaming has become a major part of our daily activities, especially with the rise of fast internet technologies like 5G. This has increased the demand for smooth, high-quality video experiences. However, ensuring that these videos are enjoyable for everyone can be challenging because different users have different preferences and experiences.

#The Challenge of Video Quality

To deliver a good video experience, streaming services use Adaptive Bitrate (ABR) algorithms. These algorithms adjust the quality of the video based on the speed of the internet connection. There are two main types of approaches:

1. Quality Of Service (QoS): Focuses on technical aspects like buffering and video quality.
2. Quality Of Experience (QoE): Aims to optimize how enjoyable the video is for users based on their feedback.

Most current approaches to improving QoE rely on average user ratings (known as Mean Opinion Scores or MOS). However, not all users rate videos the same way. This can lead to algorithms that don’t accurately reflect what users want, resulting in poorer experiences for some people.

#Introducing Jade

To address these issues, we introduce a new system called Jade. This system uses reinforcement learning with human feedback to create a more accurate model of user satisfaction. Jade learns from the preferences of different users, adjusting the video streaming accordingly.

#Understanding User Feedback

User feedback is vital in making improvements. In traditional approaches, the average of all users' ratings is taken into account. However, this method doesn’t capture the significant differences in how users rate the same videos. For instance, one user might give a video a high score while another might rate it low, even if both are watching the same content.

Jade tackles this problem by focusing on the relative feedback from users rather than the absolute scores. This way, if a user rates one session higher than another, it signals a better experience for that user.

#Building the Rank-Based Model

Jade uses a rank-based QoE model. This model looks at how users rank different video sessions in relation to each other. By focusing on the order of ratings rather than the exact numbers, Jade can create a more reliable system that reflects user satisfaction better.

Two types of models are implemented within Jade: a linear-based model and a Deep Neural Network (DNN)-based model. The linear model is simpler and more stable, while the DNN model can achieve high accuracy but may struggle with consistency.

#Training the System

Training Jade involves using feedback from users to develop the ABR algorithm. This process includes two key parts:

1. Smooth Training: In the early stages, the system uses the simpler linear model to establish a solid foundation. Later, it transitions to the more complex DNN model to fine-tune the scores.
2. Online Trace Selection: This approach uses real-time data to choose which network conditions to train on. It aims to select the best conditions that improve the learning process without prior assumptions.

#Performance Evaluation

Experimental tests show that Jade significantly improves the streaming experience in both slow and fast network conditions. It outperforms existing algorithms by a considerable percentage, indicating that it effectively caters to various user needs.

#Slow Network Performance

In conditions where the internet speed is slow, Jade's approach outperforms traditional ABR algorithms by more than 22%. This is a notable improvement compared to existing systems. The results highlight that Jade delivers better viewing quality and reduces buffering issues, which are common pain points for users.

#Fast Network Performance

In fast network conditions, Jade continues to shine. It enhances user satisfaction by at least 23% over comparable algorithms. The balancing act between video quality and buffering is well managed, ensuring users have a seamless experience.

#User Preferences Matter

The outcomes of the experiments show that different users have unique preferences. By recognizing this diversity, Jade successfully tailors the streaming experience to individual tastes. This is crucial in achieving high QoE since user satisfaction is not a one-size-fits-all situation.

#The Importance of Accurate Models

Reliance on simplified models can lead to poor experiences. For instance, when using imperfect models, algorithms can make poor decisions that result in user dissatisfaction. Jade’s combination of linear and DNN models ensures that it learns from both the general trends in data and the specific needs of individual users.

#Insights from User Feedback Analysis

The analysis of user feedback within the SQoE-IV dataset provided valuable insights. With thousands of user ratings, it became clear that there was significant variance in how scores were assigned. For some videos, scores could range drastically, indicating that users perceive video quality differently.

#Conclusion

Jade represents a significant advancement in the field of video streaming by recognizing and addressing user feedback variability. By incorporating reinforcement learning techniques and focusing on the ranks of user ratings, Jade effectively improves the video streaming experience for different users.

This innovative approach has the potential to set new standards for adaptive video streaming, aligning better with the varied expectations of users. Future developments could include real-time adjustments based on ongoing user feedback during live streaming scenarios, further enhancing the viewing experience.