Content Recommendation Considering Cache State
Yuto Murakami, Yuma Fukagawa, Noriaki Kamiyama · 2024
The demand for content such as music and videos has increased, and it drives up the Internet traffic. Content recommendation systems that suggest content based on user preferences have become a crucial component of content delivery services like Netflix, and it accounts for a significant portion of content requests. Moreover, cache delivery using CDN (content delivery network) has been widely used to reduce the delivery latency and network traffic. However, existing recommendation systems did not consider the status of caches, and they focused on just recommending contents which fitted the user interest. From the user perspective, recommending content tailored to individual preferences is desirable. On the other hand, from the network-providers perspective, recommending highly popular content to reduce delivery costs is preferred. Therefore, in this paper, we propose a content recommendation method aimed at improving user satisfaction as well as enhancing cache efficiency. The proposed method limits the targets of collaborative filtering to contents cached within cache servers. It recommends content that is physically closer to users and cached in nearby cache servers to reduce the network load and shorten access times. Additionally, to address the cold-start problem in collaborative filtering, we combine cache-limited collaborative filtering with a recommendation method based on reinforcement learning called $\epsilon$-greedy. We evaluate the effectiveness of the proposed method through computer simulations.