Abstract:In this paper, we propose a digital twin (DT)-assisted resource demand prediction scheme to enhance prediction accuracy for multicast short video streaming. Particularly, we construct user DTs (UDTs) for collecting real-time user status, including channel condition, location, watching duration, and preference. A reinforcement learning-empowered K-means++ algorithm is developed to cluster users based on the collected user status in UDTs, which can effectively employ the mined users' intrinsic correlation to improve the accuracy of user clustering. We then analyze users' video watching duration and preferences in each multicast group to obtain the swiping probability distribution and recommended videos, respectively. The obtained information is utilized to predict radio and computing resource demand of each multicast group. Initial results demonstrate that the proposed scheme can effectively abstract multicast groups' swiping probability distributions for accurate resource demand prediction.
What problem does this paper attempt to address?
This paper aims to solve the problem of resource demand prediction in multicast short - video streaming. Specifically, the authors propose a resource demand prediction scheme assisted by Digital Twin (DT) to improve the prediction accuracy. The following are the main problems that the paper attempts to solve:
1. **How to accurately and timely cluster users into multicast groups**:
- With the network dynamics and diverse characteristics of users, short - videos usually need to be transcoded into multiple bit rates on cloud or edge servers to reduce transmission delay. Therefore, a method is needed to quickly and accurately update the multicast group according to the real - time state of users (such as channel conditions, sliding behaviors, etc.).
2. **How to quantify the impact of users' sliding behaviors on resource reservation**:
- Users' sliding behaviors may lead to the pre - cached video segments not being played, resulting in excessive resource reservation. Therefore, a method is needed to quantify the impact of these behaviors on resource reservation in order to manage resources more effectively.
### Main contributions of the paper
1. **Construct user digital twins (UDTs) to collect user states**:
- By collecting the real - time states of users (including channel conditions, location, viewing duration, and preferences), user digital twins (UDTs) are constructed, and a two - step method is proposed to achieve accurate and fast multicast group construction.
2. **Abstract the sliding probability distribution from the viewing duration**:
- Using the viewing duration data stored in UDTs, the sliding probability distribution of the multicast group is abstracted, and this information is used to predict the resource demand of each multicast group.
### Specific steps of the scheme
1. **Multicast group construction**:
- Based on the user states stored in UDTs, analyze user similarity and group users with similar states into the same multicast group. User similarity is defined by the Euclidean distance between any two user states.
- Use a one - dimensional convolutional neural network (1D - CNN) to compress the time - series UDTs data.
- Adopt a double - deep Q - network (DDQN) to determine the number of groups, and then use the K - means++ algorithm for fast user clustering.
2. **Group - based resource demand prediction**:
- Once users are clustered into multicast groups, extract the corresponding group - level information, such as the sliding probability distribution and recommended videos.
- Use the users' viewing duration to update the sliding probability distribution of the multicast group, and update the recommended videos according to video popularity and user preferences.
- Based on the extracted group - level information, analyze the average participation time, video traffic, and computational consumption of the multicast group to predict radio and computational resource demands.
### Preliminary simulation results
- Simulation results show that the proposed scheme achieves a high prediction accuracy of up to 95.04% in radio resource demand prediction.
### Conclusions and future work
- This paper proposes a resource demand prediction scheme based on digital twins, which can effectively abstract the sliding probability distribution and recommended videos of multicast groups.
- Future work will focus on how to effectively reserve radio and computational resources based on the predicted multicast group resource demands.
Through these methods, this paper provides a new solution for resource management in multicast short - video streaming, which helps to improve resource utilization efficiency and user experience.