EntroGD: Efficient Compression and Accurate Direct Analytics on Compressed Data
Abstract
Generalized Deduplication (GD) enables lossless compression with direct analytics on compressed data by dividing data into \emph{bases} and \emph{deviations} and performing dictionary encoding on the former. However, GD algorithms face scalability challenges for high-dimensional data. For example, the GreedyGD algorithm relies on an iterative bit-selection process across $d$-dimensional data resulting in $O(nd^2)$ complexity for $n$ data rows to select bits to be used as bases and deviations. Although the $n$ data rows can be reduced during training at the expense of performance, highly dimensional data still experiences a marked loss in performance. This paper introduces EntroGD, an entropy-guided GD framework that reduces complexity of the bit-selection algorithm to $O(nd)$. EntroGD operates considers a two-step process. First, it generates condensed samples to preserve analytic fidelity. Second, it applies entropy-guided bit selection to maximize compression efficiency. Across 18 datasets of varying types and dimensionalities, EntroGD achieves compression performance comparable to GD-based and universal compressors, while reducing configuration time by up to 53.5$\times$ over GreedyGD and accelerating clustering by up to 31.6$\times$ over the original data with negligible accuracy loss by performing analytics on the condensed samples, which are much fewer than original samples. Thus, EntroGD provides an efficient and scalable solution to performing analytics directly on compressed data.