User generated content (UGC) makes up a significant portion of Internet traffic. As opposed to other content, UGC has so far been left outside over-the-top providing network operators content distribution networks (telco-CDN) due to the difficulty in determining optimised placement of such content. The side effect of this is that UGC content is not placed close to end users and therefore occupy unnecessary network resources. The difficulty in determining optimal placement of UGC stems from the different geographical and dynamic behaviour of the content generators, and a further complication is that with UGC, it is necessary to place content in real-time which this has an impact on performance optimality. Even though CDNs have been widely studied in the literature, little attention has been given to the challenging case of UGC placement. In this paper, we propose an on-line placement algorithm and compare its performance with the off-line counterpart based on integer programming, both under the assumption that the popularity of content is known to the algorithms. In order to determine the popularity, we present an on-line learning model to predict spatial patterns in content requests. Furthermore, we couple the model with an algorithm for learning the early popularity of content, i.e., shortly after the content becomes known. We show that together, these approaches enable service providers to effectively place UGC and minimise the cost of serving UGC in their networks.

中文翻译：

---

用于在 Telco-CDN 中高效交付用户生成内容的在线学习和放置算法

用户生成的内容 (UGC) 占互联网流量的很大一部分。与其他内容相反，由于难以确定此类内容的优化放置，因此 UGC 迄今为止一直被排除在顶级提供网络运营商内容分发网络 (telco-CDN) 之外。这样做的副作用是 UGC 内容没有放置在靠近最终用户的位置，因此会占用不必要的网络资源。确定 UGC 最佳放置的困难源于内容生成器的不同地理和动态行为，更复杂的是，对于 UGC，需要实时放置内容，这会影响性能优化。尽管 CDN 在文献中得到了广泛的研究，但很少有人关注 UGC 放置这一具有挑战性的案例。在本文中，我们提出了一种在线放置算法，并将其性能与基于整数规划的离线对应算法进行比较，两者均假设内容的流行度是算法已知的。为了确定流行度，我们提出了一个在线学习模型来预测内容请求中的空间模式。此外，我们将模型与算法相结合，用于学习内容的早期流行度，即在内容已知后不久。我们共同表明，这些方法使服务提供商能够有效地放置 UGC，并将在其网络中提供 UGC 服务的成本降至最低。两者都假设内容的流行度是算法已知的。为了确定流行度，我们提出了一个在线学习模型来预测内容请求中的空间模式。此外，我们将模型与算法相结合，用于学习内容的早期流行度，即在内容已知后不久。我们共同表明，这些方法使服务提供商能够有效地放置 UGC，并将在其网络中提供 UGC 服务的成本降至最低。两者都假设内容的流行度是算法已知的。为了确定流行度，我们提出了一个在线学习模型来预测内容请求中的空间模式。此外，我们将模型与算法相结合，用于学习内容的早期流行度，即在内容已知后不久。我们共同表明，这些方法使服务提供商能够有效地放置 UGC，并将在其网络中提供 UGC 服务的成本降至最低。