Research
Joint Assortment and Cache Planning for Practical User Choice Model in Wireless Content Caching Networks
Y. Fu, X. Xu, H. Liu, Q. Yu, H. -N. Dai and T. Q. S. Quek, "Joint Assortment and Cache Planning for Practical User Choice Model in Wireless Content Caching Networks," in IEEE Transactions on Mobile Computing, vol. 23, no. 5, pp. 4709-4722, May 2024, doi: 10.1109/TMC.2023.3297987.
In wireless content caching networks (WCCNs), a user's content consumption crucially depends on the assortment offered. Here, the assortment refers to the recommendation list. An appropriate user choice model is essential for greater revenue. Therefore, in this paper, we propose a practical multinomial logit choice model to capture users’ content requests. Based on this model, we first derive the individual demand distribution per user and then investigate the effect of the interplay between the assortment decision and cache planning on WCCNs’ achievable revenue. A revenue maximization problem is formulated while incorporating the influences of the screen size constraints of users and the cache capacity budget of the base station (BS). The formulated optimization problem is a non-convex integer programming problem. For ease of analysis, we decompose it into two folds, i.e., the personalized assortment decision problem and the cache planning problem. By using structure-oriented geometric properties, we design an iterative algorithm with examinable quadratic time complexity to solve the non-convex assortment problem in an optimal manner. The cache planning problem is proved to be a 0-1 Knapsack problem and thus can be addressed by a dynamic programming approach with pseudo-polynomial time complexity. Afterwards, an alternating optimization method is used to optimize the two types of variables until convergence. It is shown by simulations that the proposed scheme outperforms various existing benchmark schemes.
Machine Learning-as-a-Service: A Distribution Channel Perspective
(In progress)
Machine Learning-as-a-service (MLaaS) aims to make artificial intelligence (AI) products more accessible and affordable by streamlining and automating the customization process. However, this innovative solution has led to different channel structures in the industry. In this paper, we develop game-theoretic models to study distribution channel choices under various MLaaS business models. Under Model-I where MLaaS is operated by the downstream Infrastructure-as-a-Service (IaaS) provider, we show the upstream hardware supplier could retain a high wholesale price to deter the MLaaS channel if MLaaS is inefficient. When MLaaS is efficient, we find Model-I is a win-win business model compared to the benchmark where no MLaaS is offered because the hardware supplier prefers to lower the wholesale price in dual channels. Inspired by the practice of Nvidia DGX Cloud, we introduce a novel buyback selling format in the hardware supplier-operated MLaaS model (Model-H), in which the hardware supplier leases servers from the IaaS provider to facilitate its MLaaS channel. Under Model-H, the MLaaS channel is more likely to exist in equilibrium when the wholesale price is rather high. In response, when MLaaS efficiency is intermediate, the hardware supplier would charge a higher wholesale price than the benchmark to induce dual channels, which could be detrimental to the IaaS provider. Furthermore, we compare these two business models and highlight: Model-H is preferable to both parties when MLaaS efficiency is moderately high; and the MLaaS provider paradoxically prefers the other party to run MLaaS when its efficiency is sufficiently high. We also consider an agency MLaaS model (Model-A) as an alternative to Model-H and find it could always avoid win-lose consequences if the commission rate is fine-tuned. Our results provide useful insights and practical guidelines for the operations of hardware suppliers and cloud service providers in the growing AI industry.