为提升短视频内容分发的精度,分析用户所属社交群体的兴趣倾向和对短视频内容的个性化需求,在基于主动推荐方式的短视频应用场景中,以视频内容提供商利润最大化为优化目标,设计了一种短视频内容分发策略.首先,基于联邦学习,利用用户群本地相册数据训练兴趣预测模型,提出用户群兴趣向量预测算法并得到用户群的兴趣向量表示;然后,以用户群的兴趣向量作为输入,基于组合置信上界(CUCB)算法实时设计相应的短视频内容分发策略,从而使视频内容提供商获取的长期利润最大化.所提策略获得的平均利润相对稳定且明显优于单纯基于CUCB的短视频分发策略得到的平均利润;与置信上界(UCB)策略和随机策略相比,所提策略使得视频内容提供商获得的总利润分别提高了12％和30％.实验结果表明,所提短视频内容分发策略能有效地提升短视频分发的精度,从而进一步提高视频内容提供商获取的利润.

With the development of 5th generation (5G) wireless communication networks and the popularity of short video applications, there has been a rapid increase in short video traffic in cellular networks. Device-to-device (D2D) communication-based short video sharing is considered to be an effective way to offload traffic from cellular networks. Due to the selfish nature of mobile user equipment (MUEs), how to dynamically motivate MUEs to engage in short video sharing while ensuring the Quality of Service, which makes it critical to design an appropriate incentive mechanism. In this paper, we firstly analyze the rationale for dynamically setting rewards and penalties and then define the rewards and penalties setting dynamically for maximizing the utility of the mobile edge computing server (RPSDMU) problem. The problem is proved NP-hard. Furthermore, we formulate the dynamic incentive process as the Markov Decision Process problem. Considering the complexity and dynamics of the problem, we design a Dynamic Incentive Mechanism algorithm of D2D-based Short Video Sharing based on Asynchronous Advantage Actor-Critic (DIM-A3C) to solve the problem. Simulation results show that the proposed dynamic incentive mechanism can increase the utility of mobile edge computing server by an average of 22% and 16% compared with the existing proportional incentive mechanism (PIM) and scoring-based incentive mechanism (SIM). Meanwhile, DIM-A3C achieves a higher degree of satisfaction than PIM and SIM.

With the rapid popularity of short video applications, short videos sharing based on D2D multicast communication is considered as a promising technology to improve the quality of service. In this paper, we propose an adaptive clustering strategy in a short videos distribution scenario to group short video users into clusters. Specifically, a cluster heads (CHs) control algorithm based on deep reinforcement learning is proposed to determine the number of CHs adaptively for different system states. After that, we propose the bisecting K-means based on the physical-interest distance algorithm (PIBK-means) and select the CH of each cluster by analyzing the distance-power weight. Simulation results verify the effectiveness of the proposed adaptive clustering strategy and show that the proposed strategy can increase the degree of satisfaction by an average of 40% than the K-means algorithm.