Computation-friendly compression of matrices and tries

In this thesis, we continue the research on repetitive data compression by investigating novel general compression schemes that are data-independent. Although we specifically focus on machine learning and key-value systems, we believe that our methods provide insights applicable to a wider range of application domains. Our proposed methods adapt one-dimensional general-purpose compression tools to handle complex data structures such as matrices, graphs and tries. These schemes effectively capture redundancies and interdependencies among the data, enabling compression that surpasses what can be achieved through sparsity alone, and without compromising the quality metrics such as precision or recall of the resulting models. Following the “computation-friendly” paradigm, our compressed representations allow for direct operations on the compressed data, with time comparable to operations on uncompressed data.