In this article, the edge caching policy in fog radio access networks (F-RANs) is optimized via deep learning. Considering that it is hard for fog access points (F-APs) to collect sufficient data of massive content features, our proposed edge caching policy only utilizes the number of requests and user location. In an offline phase, we propose to learn the corresponding popularity prediction model for every content popularity trend class and user location prediction models to make the popularity prediction accurate, adaptive and targeted. Moreover, we develop a loss function to avoid overfitting and increase sensitivity to high popularity for popularity prediction models. In an online phase, we propose a reactive caching scheme to react to user requests. In order to guarantee that classification can improve the popularity prediction accuracy in both phases, deep learning and k-Nearest Neighbor (kNN) are combined to classify popularity trends. Besides, a joint proactive-reactive caching policy is proposed to maximize the cache hit rate. The proposed policy is able to promptly track the various popularity trends with spatial-temporal popularity, trend and user dynamics with a low computational complexity. Extensive performance evaluation results show that the cache hit rate of our proposed policy approaches that of the optimal policy.

中文翻译：

---

雾无线接入网络中基于深度学习的边缘缓存

在本文中，雾无线电接入网络 (F-RAN) 中的边缘缓存策略通过深度学习进行了优化。考虑到雾接入点（F-AP）难以收集足够的海量内容特征数据，我们提出的边缘缓存策略仅利用请求数和用户位置。在离线阶段，我们建议为每个内容流行趋势类别学习相应的流行预测模型和用户位置预测模型，使流行预测准确、自适应和有针对性。此外，我们开发了一个损失函数来避免过度拟合并增加对流行度预测模型的高流行度的敏感性。在在线阶段，我们提出了一种反应式缓存方案来对用户请求做出反应。为了保证分类在两个阶段都能提高流行度预测精度，结合深度学习和k-最近邻（kNN）对流行度趋势进行分类。此外，提出了联合主动-反应缓存策略以最大化缓存命中率。所提出的策略能够以较低的计算复杂度及时跟踪具有时空流行度、趋势和用户动态的各种流行趋势。广泛的性能评估结果表明，我们提出的策略的缓存命中率接近最优策略的缓存命中率。所提出的策略能够以较低的计算复杂度及时跟踪具有时空流行度、趋势和用户动态的各种流行趋势。广泛的性能评估结果表明，我们提出的策略的缓存命中率接近最优策略的缓存命中率。所提出的策略能够以较低的计算复杂度及时跟踪具有时空流行度、趋势和用户动态的各种流行趋势。广泛的性能评估结果表明，我们提出的策略的缓存命中率接近最优策略的缓存命中率。