Recommend what to cache: a simple self-supervised graph-based recommendation framework for edge caching networks

Deep Learning-based edge caching networks can accurately infer what to cache based on a user's historical content requests, thereby significantly relieving the burden of the backbone networks. However, the cold-start problem inherent in deep learning may limit the performance of history-based caching strategies. Due to the mobile and dynamic nature of wireless networks, base stations often lack sufficient data to accurately estimate the user's demands and cache the possible requested data. In this context, we adopt self-supervised learning (SSL) into the caching strategies and propose a Simple Self-supervised Graph-based Recommendation framework for edge caching networks (SimSGR). Specifically, we propose two new network layers: the Mixing layer and the Conversion layer. The former replaces the data augmentation of the SSL paradigm to avoid destroying the semantic loss, while the latter greatly simplifies the loss function, which helps to lighten the model structure and facilitates deployment on edge caching networks. Simulation results show that our model outperforms baseline algorithms that are sensitive to augmentation hyper-parameters, particularly when trained in a cold-start environment.