Generalized linear models with nonlinear feature transformations are widely used for large-scale regression and classification problems with sparse inputs. Memorization of feature interactions through a wide set of cross-product feature transformations are effective and interpretable, while generalization requires more feature engineering effort. With less feature engineering, deep neural networks can generalize better to unseen feature combinations through low-dimensional dense embeddings learned for the sparse features. However, deep neural networks with embeddings can over-generalize and recommend less relevant items when the user-item interactions are sparse and high-rank. In this paper, we present Wide & Deep learning-jointly trained wide linear models and deep neural networks-to combine the benefits of memorization and generalization for recommender systems. We productionized and evaluated the system on Google Play, a commercial mobile app store with over one billion active users and over one million apps. Online experiment results show that Wide & Deep significantly increased app acquisitions compared with wide-only and deep-only models. We have also open-sourced our implementation in TensorFlow.

In-network caching necessitates the transformation of centralised operations of traditional, overlay caching techniques to a decentralised and uncoordinated environment. Given that caching capacity in routers is relatively small in comparison to the amount of forwarded content, a key aspect is balanced distribution of content among the available caches. In this paper, we are concerned with decentralised, real-time distribution of content in router caches. Our goal is to reduce caching redundancy and in turn, make more efficient utilisation of available cache resources along a delivery path.Our in-network caching scheme, called ProbCache, approximates the caching capability of a path and caches contents probabilistically in order to: i) leave caching space for other flows sharing (part of) the same path, and ii) fairly multiplex contents of different flows among caches of a shared path.We compare our algorithm against universal caching and against schemes proposed in the past for Web-Caching architectures, such as Leave Copy Down (LCD). Our results show reduction of up to 20% in server hits, and up to 10% in the number of hops required to hit cached contents, but, most importantly, reduction of cache-evictions by an order of magnitude in comparison to universal caching.

Abstract. Ubiquitous in-network caching is one of the key aspects of information-centric networking (ICN) which has recently received widespread research interest. In one of the key relevant proposals known as Networking Named Content (NNC), the premise is that leveraging in-network caching to store content in every node it traverses along the delivery path can enhance content delivery. We question such indiscriminate universal caching strategy and investigate whether caching less can actually achieve more. Specifically, we investigate if caching only in a subset of node(s) along the content delivery path can achieve better performance in terms of cache and server hit rates. In this paper, we first study the behavior of NNC's ubiquitous caching and observe that even naïve random caching at one intermediate node within the delivery path can achieve similar and, under certain conditions, even better caching gain. We propose a centrality-based caching algorithm by exploiting the concept of (ego network) betweenness centrality to improve the caching gain and eliminate the uncertainty in the performance of the simplistic random caching strategy. Our results suggest that our solution can consistently achieve better gain across both synthetic and real network topologies that have different structural properties.Keywords: Information-centric networking, caching, betweenness centrality. IntroductionInformation-centric networking (ICN) has recently attracted significant attention, with various research initiatives (e.g.,[4] and COMET [5]) targetting this emerging research area. The main reasoning for advocating the departure from the current host-to-host communications paradigm to an information/content-centric one is that the Internet is currently mostly used for content access and delivery, with a high volume of digital content (e.g., 3D/HD movies, photos etc.) delivered to users who are only interested in the actual content rather than the source location. As such, we no longer need a natively supported content distribution framework. While the Internet was designed for and still focuses on host-to-host communication, ICN shifts the emphasis to content objects that can be cached and accessed from anywhere within the network rather than from the end hosts only. In ICN, content names are decoupled from host addresses, effectively separating the role of identifier and locator in distinct contrast to current IP addresses which are serving both purposes. Naming content directly enables the exploitation of in-network caching in order to improve delivery of popular content. Each content object can now be uniquely identified and authenticated without being associated to a specific host. This enables application-independent caching of content pieces that can be re-used by other end users requesting the same content. In fact, one of the salient ICN features is in-network caching, with potentially every network element (i.e., router) caching all content fragments 1 that traverse it; in this context, if a matching request is recei...

Networking Named Content (NNC) was recently proposed as a new networking paradigm to realise Content Centric Networks (CCNs). The new paradigm changes much about the current Internet, from security and content naming and resolution, to caching at routers, and new flow models. In this paper, we study the caching part of the proposed networking paradigm in isolation from the rest of the suggested features. In CCNs, every router caches packets of content and reuses those that are still in the cache, when subsequently requested. It is this caching feature of CCNs that we model and evaluate in this paper. Our modelling proceeds both analytically and by simulation. Initially, we develop a mathematical model for a single router, based on continuous time Markov-chains, which assesses the proportion of time a given piece of content is cached. This model is extended to multiple routers with some simple approximations. The mathematical model is complemented by simulations which look at the caching dynamics, at the packet-level, in isolation from the rest of the flow.

Abstract-We introduce the concept of resource management for in-network caching environments. We argue that in Information-Centric Networking environments, deterministically caching content messages at predefined places along the content delivery path results in unfair and inefficient content multiplexing between different content flows, as well as in significant caching redundancy. Instead, allocating resources along the path according to content flow characteristics results in better use of network resources and therefore, higher overall performance. The design principles of our proposed in-network caching scheme, which we call ProbCache, target these two outcomes, namely reduction of caching redundancy and fair content flow multiplexing along the delivery path. In particular, ProbCache approximates the caching capability of a path and caches contents probabilistically to: 1) leave caching space for other flows sharing (part of) the same path, and 2) fairly multiplex contents in caches along the path from the server to the client. We elaborate on the content multiplexing fairness of ProbCache and find that it sometimes behaves in favor of content flows connected far away from the source, that is, it gives higher priority to flows travelling longer paths, leaving little space to shorter -path flows. We introduce an enhanced version of the main algorithm that guarantees fair behavior to all participating content flows. We evaluate the proposed schemes in both homogeneous and heterogeneous cache size environments and formulate a framework for resource allocation in in-network caching environments. The proposed probabilistic approach to in-network caching exhibits ideal performance both in terms of network resource utilization and in terms of resource allocation fairness among competing content flows. Finally, and in contrast to the expected behavior, we find that the efficient design of ProbCache results in fast convergence to caching of popular content items.