Tensor Completion Algorithm and Its Applications to Wireless Edge Caching and Hyper-Spectral Imaging

This paper presents a lightweight tensor completion algorithm with applications in wireless edge caching and hyper-spectral imaging. In wireless edge caching, the dynamic and selective nature of content popularity inevitably leads to missing observations. We address this challenge by employing tensor-based models, which are notably practical for imputing missing data. We approach cooperative caching as a problem of third-order tensor completion and prediction, acknowledging the correlations across time, files, and base stations. Given the potentially large content libraries, we modify the latent norm-based Frank-Wolfe (FW) algorithm. This modification, which incorporates multi-rank updates instead of the traditional rank-1 updates, significantly reduces computational time complexity. This advancement facilitates the development of an efficient online caching algorithm. Through simulations using Movielens datasets and comparing them with the conventional FW algorithm, our proposed algorithm demonstrates a reduced computational overhead reduction and improved normalized cache hit rates within linear prediction models. Furthermore, since hyper-spectral images often have missing or corrupted data due to various noise sources, we conduct simulations on hyper-spectral images. From the Pavia University and Pavia City datasets with noisy data, our proposed tensor completion algorithm can effectively recover the original images from heavily interfered with communication noise. The effectiveness of our proposed tensor completion algorithm across various applications, including wireless edge caching and hyper-spectral imaging, accentuates its versatility and significant utility. Its ability to adeptly handle diverse challenges underscores its value as a comprehensive solution for solving intricate data recovery problems.