Popular Content Distribution in Public Transportation Using Artificial Intelligence Techniques

Abstract

Outdoor wireless networks suffer nowadays from an increasing data traffic demand which comes at the time where almost no vacant frequency spectrum has been left. A vast majority of this demand comes from popular content generated by video streaming and social media sites. In the future, other sources will generate even more demand with emerging applications such as virtual reality, connected cars and environmental sensing. While a significant progress has been made to address this network saturation in indoor environments, current outdoor solutions, based on fixed network deployments, are expensive to build and maintain. They tend to be immobile and therefore are inflexible in coping with the dynamics of outdoor data demand. On the other hand, Vehicular Ad-hoc NETworks (VANETs) are in nature more scalable, dynamic, flexible, and therefore more promising in terms of addressing such demand. This is especially feasible if we take advantage of public transportation vehicles and stops. These vehicles and stops are often owned and operated by the same administrative entity which overcomes the routing selfishness issue. Moreover, the mobility patterns of these vehicles are highly predictable given their regular schedules; their locations are publicly-sharable and their location distribution is uniform throughout space and time. Given these factors, a system that utilizes public transportation vehicles and stops to build a reliable, scalable and dynamic VANET for wireless network offloading in outdoor environments is proposed. This is done by exploiting the predictability demonstrated by such vehicles using an Artificial-Intelligence (AI) based system for wireless network offloading via popular content distribution. The AI techniques used are the Upper Popularity Bound (UPB) collaborative and group-based recommender based on multi-armed bandits for content recommendation and bayesian optimization based on batch-based Random Forest (RF) regression for content routing. They are used after analyzing the mobility data of public transportation vehicles and stops. This analysis includes both preprocessing and processing the data in order to select the optimal set of stops and clustering vehicles and stops based on cumulative contact duration thresholds. The final system has shown the promising networking potential of public transportation. It incorporates a recommender that has shown a versatile performance under different consumer interest and network capacity scenarios. It has also demonstrated a superior performance using a bayesian optimization technique that offloads as high as 95% of the wireless network load in an interference and collision free manner.