AI4RecNets

Content-centric network optimization schemes (e.g., caching, traffic routing, multicast, etc.) have traditionally been designed independently from application-level recommendation algorithms for such content. However, the ubiquitous recommendations engines of popular services are increasingly driving user requests, and thus can largely burden (as is the case today), but potentially also greatly facilitate (as we propose) network algorithms. For this reason, we argue that co-design of recommendation and network functions is a highly promising method to improve both network performance (reduced cost to serve traffic) and user experience (improved streaming QoS for recommended content). While some very recent articles on the topic have surfaced by our and a few other research teams, most key challenges remain largely open: (i) almost no data-driven studies exist to guide the few existing model-based optimization solutions; (ii) most co-design problems in this context turn out to be hard optimization problems in both discrete and continuous (relaxed) formulations; as a result, proposed solutions scale poorly even for mildly realistic content catalogue sizes; (iii) the majority of proposed algorithms are offline and cannot handle neither non-stationarity nor lack of knowledge about key environment variables; (iv) very little is known about the performance limits of such joint algorithms; (v) last but not least, adoption of such algorithms will require the cooperation of various market players such as network operators, content providers, content distribution networks, etc., and yet no network economics framework has been investigated to ensure such a cooperation will be, or can be made, profitable for all parties. To this end, project AI4RecNets will leverage and combine modern artificial intelligence, online optimization theory, complex network theory, and network economics to address these challenges in a unified and data-driven manner. To our best knowledge, this is the first project world-wide with a similarly ambitious scope and multi-disciplinary methodology, and we strongly believe that its success will pave the road for a radical shift in how application algorithm and network algorithm design is approached in future wireless (and wired) networks.