From Cloud to Device: Transforming Recommender Systems for On-Device Deployment

The School of EECS is hosting the following PhD progress review 3 seminar

From Cloud to Device: Transforming Recommender Systems for On-Device Deployment

Speaker: Xin Xia
Host: Assoc Prof Hongzhi Yin

Abstract: Deep learning-powered recommender systems leverage massive neural networks and models to extract hidden user preferences, enabling unrivaled capabilities. Nevertheless, these systems are constructed entirely on the server-side, relying on substantial storage, memory, and computational resources from the cloud. The frequent training and updating of recommendation models, along with the high-speed processing of millions of concurrent user requests, entail the usage of energy-consuming CPUs/GPUs, necessitating an enormous amount of carbon cost. Moreover, these cloud-based models are trained using vast amounts of user behavior data and even require immediate contextual data for real-time inference, intensifying public concerns regarding data privacy.

Recently, there has been a rising interest in on-device machine learning, which aims to enable models to operate on resource-constrained devices. Unlike the cloud-based paradigm,  models are initially trained on the cloud and then downloaded and deployed on local energy-efficient devices, such as smartphones. In this paradigm, users no longer need to upload sensitive data to servers, thereby enjoying low-latency services that are significantly faster than server-side models. Since the concept of on-device machine learning aligns well with the need for low-cost and privacy-preserving recommender systems, the trend is moving towards on-device recommendations. However, transitioning deep learning-powered recommender systems from their traditional cloud-based infrastructures to a more localized, on-device paradigm presents unique challenges. In this thesis, we present three major challenges in transitioning deep learning-powered recommender systems from their traditional cloud-based infrastructures to a more localized, on-device paradigm and propose innovative solutions.

Bio: Miss Xin Xia obtained her B.E. in Software Engineering from Jilin University in 2019. Then she started her Ph.D. in computer science under the supervision of A/Prof. Hongzhi Yin and Dr. Miao Xu. Her research interests include self-supervised learning, on-device machine learning and session-based recommendation.