Article Preview
TopIntroduction
As the internet continues its relentless expansion, there has been an exponential surge in data traffic. To cater to users' communication demands, major telecommunications operators have deployed densely packed small cell stations. However, this has significantly burdened the backhaul links (X. Wang et al., 2014). In emergency scenarios such as natural disasters or other crises, certain base stations may be damaged and complex geographic terrains complicate matters further, and the intricacies and vulnerabilities of communication systems are amplified. To address these challenges, the caching of popular content on UAVs integrated into the cellular network has emerged as a compelling research topic. This approach not only mitigates transmission latency but also alleviates the strain on backhaul links during peak hours (Li et al., 2019; S. Zhang et al., 2019).
Research by Navarro-Ortiz et al. (2020) forecasts a global increase in smartphone users from 6.3 billion to 12 billion between 2020 and 2030, accompanied by a 10 to 100-fold surge in global mobile communication traffic. This colossal surge in data traffic poses a substantial challenge to existing communication systems. In recent years, UAVs have found increasing applications in wireless communication. For instance, Chen et al. (2017) introduced a UAV deployment function with caching capabilities in a cloud access network, aiming to minimize UAV transmission distances. Liu et al. (2019) employed a genetic algorithm-based K-means (GA-K-means) method to partition users into cells and subsequently proposed a Q-learning-based deployment algorithm for UAVs. Park et al. (2019) optimized the placement of multiple UAVs in a base station-aided communication system, considering user demands to maximize service throughput. J. Yang et al. (2020) introduced a UAV collaboration scheme for caching in cognitive radio networks, enhancing CRN's transmission capacity while reducing redundant traffic loads. Zeng et al. (2022) presented a layered caching solution for different UAVs, caching specified layers of video based on Scalable Video Coding (SVC).
However, none of these solutions consider the user's preferences when requesting resources. Zhao et al. (2019) and T. Zhang et al. (2020) focused solely on caching popular content on UAVs. Traditional caching policies such as LFU, LRU, and FIFO, while effective in scenarios with consistent object sizes (Xu et al., 2018), struggle to adapt adequately to the wireless network environment due to a lack of consideration for factors like network topology and user preferences (Han et al., 2021).
On another front, the explosive growth of multimedia content has led to the pervasive problem of information overload. As one of the critical methods for alleviating this issue, recommendation systems provide users with services tailored to their interests. The evolution of recommendation system models can be broadly categorized into three phases: traditional shallow models, general neural network models, and graph neural network models. Early models relied on computing the similarity of user/item historical data directly in order to capture collaborative filtering (CF) effects (Koren et al., 2009; Su & Khoshgoftaar, 2009). However, these shallow models struggle to handle complex user behaviors or data inputs. With rapid advancements in deep learning research, neural network-based recommendation models emerged as an upgrade to shallow models. The Neural Collaborative Filtering (NCF) model, for instance, employs a Multi-Layer Perceptron (MLP) instead of the dot-product function in matrix factorization models (Kipf & Welling, 2016).