CoPo: Self-supervised Contrastive Learning for Popularity Prediction in MEC Networks

CoPo: Self-supervised Contrastive Learning for Popularity Prediction in MEC Networks

연구 분야: Artificial Intelligence

학회: 2023 24th International Conference on Digital Signal Processing (DSP)

Mobile Edge Caching (MEC) technology aims to provide high-quality multimedia content to mobile users by bringing storage and computation resources closer to the edge of the network. MEC networks, however, face several challenges such as limited storage capacity, dynamic network conditions, and the need for low-latency content delivery. To address these challenges, recent research has focused on integrating MEC networks with Deep Neural Networks (DNNs), in particular, supervised learning models. One significant limitation of supervised popularity prediction models is the requirement for manual labeling of contents as popular or unpopular by investigating users’ past behavior, which can be a time-intensive task. This paper proposes a self-supervised learning algorithm called Contrastive learning Popularity (CoPo) prediction framework to predict the dynamic content popularity in a MEC network. The framework utilizes the distinguishing aspect of the Contrastive Learning (CL) paradigm to recognize differences among input samples, including users’ contextual information and is based on the Long Short Term Memory (LSTM) model to capture temporal information. Simulation results illustrate that the proposed CoPo framework outperforms the self-supervised/unsupervised state-of-the-art methods.