Machine Learning Over Wireless: An Application in Wireless Recommender Systems

Description

In the big data era, recommender systems become fundamental to filter out content by exploring users' preferences and tastes, rather than explicitly searching for pre-specified content. The recommendation applications range from multimedia recommendations, such as movies, music, news, and ads, to product recommendations in e-commerce. On the other hand, a growing amount of Internet traffic is transmitted in wireless networks and mobile devices become a major tool for content access. More computation is being pushed to the edge devices (i.e., edge computing), such as sensors, industrial machines, and Wi-Fi hotspots. New challenges emerge for these wireless applications, such as limited bandwidth, broadcast features, distributed operations, and dynamic environments. Thus, it is natural to ask how to design recommender systems featuring wireless environments, or even more broadly, how to design machine learning algorithms over wireless.The main goal of this proposal is to develop a new learning framework for recommender systems that fits well in the wireless environment with high dynamics and uncertainty, limited bandwidth, and distributed storage and computing resources. Note that this project is NOT about how we apply recommendation techniques to recommend wireless channels, bandwidth, etc. This project is to design a recommendation framework that fits into the wireless. To tackle these problems, we leverage the contextual multi-armed bandit framework for recommendations, where the contextual information is used to capture the situational or environmental knowledge when making a recommendation, e.g., time, location, user profiles, browsing history. To capture the aforementioned wireless features, we propose three new wireless recommendation problems in this new learning framework for wireless recommendations and we will design and analyze the learning algorithms. 1) Group recommendations under bandwidth constraints. A recommender can recommend and broadcast items to a group of users, and we aim to learn the preference of users and optimize the content delivery at the same time given the limited bandwidth. 2) Federated recommendations for distributed operations. Each computing server has local data for making recommendations, and they can communicate and incorporate their recommendation models to make better decisions. Our goal is to make communication-efficient and robust recommendations. 3) Contextual recommendations with latent structures. Latent structures of items and contexts (latent structures mainly refer to some unobserved underlying structures, such as the clustering of items or contexts, due to the high dynamics of the wireless environments) are to be explored in order to facilitate the learning efficiency.