On the joint optimization of content caching and recommendations

Abstract

Content caching has been experiencing revived interest within the context of current and next generation wireless networks. It brings content closer to users, decreasing the aggregate delivery costs and service delays for network providers. Recommender systems have become integral components of content provision sites. They offer personalized recommendations, increasing the individual user satisfaction and engagement with the content provider’s platform. Traditionally, caching and recommendation decisions are taken separately. However, there is a recent persistent trend where both network and content providers tend to deploy their own content delivery solutions. In light of this, we explore how the phenomenally conflicting objectives of content caching and recommendation can be jointly addressed. In this chapter we approach recommender systems as network traffic engineering tools that actively shape content demand to serve both user- and network-centric performance objectives. We introduce a model that captures the coupling between caching decisions and issued recommendations. Based on experimental evidence, we describe the impact of recommendations on user content requests and present a systematic way of engineering the user recommendations. Our viewpoint to recommender systems raises some concerns, not least ethical ones. Hence, we introduce a measure called preference distortion tolerance to quantify how much the engineered recommendations distort the original user content preferences. We describe the “recommendation window” as a way to controllably bound the distortion that recommendations will undergo and discuss specific properties that the recommender system should have in order to ensure that the “Quality of Recommendations (QoR)" is kept high by the issued recommendations. We formulate a joint optimization problem for the content to cache and recommend to each user, aiming to maximize the cache hit ratio. The preference distortion tolerance is embedded as a constraint to this problem formulation and marks an equally important dimension for assessing possible solutions. We prove that the problem lacks those properties that would guarantee its approximability, and devise a low-complexity practical algorithm that solves it efficiently. The algorithm is essentially a form of lightweight control over user recommendations, so that the recommended content is both appealing to the end-user and more friendly to the caching system and the network resources. We thoroughly evaluate it, by establishing its main properties and analytical performance bounds. We conduct an extensive sensitivity analysis over various system parameters, both analytically and through simulations with real and synthetic dataset. Next we extend our model and approach for a system in which users can access content from multiple caches, and analyse the intermediate joint caching and user association problem. We then discuss open directions and provide a detailed taxonomy of the related work. Due to space limitations, we do not provide proofs for our results. However, the interested reader may refer to [1-3] for more propositions and their detailed proofs.