2017/03/14 by Mahmoud Abo Khamis, Khamis, Mahmoud Abo, Hung Q. Ngo +7 · 2 citations
Computer Science · Decision Sciences · #Data Mining Algorithms and Applications #Data Quality and Management #Databases (cs.DB) #FOS: Computer and information sciences #H.2.4 #I.2.6 #Machine Learning and Data Classification
paper · pdf · doi:10.48550/arxiv.1703.04780
openalex publication_date 2017/03/14 · openalex created_date 2022/10/02 · openalex updated_date 2026/07/28
Integrated solutions for analytics over relational databases are of great\npractical importance as they avoid the costly repeated loop data scientists\nhave to deal with on a daily basis: select features from data residing in\nrelational databases using feature extraction queries involving joins,\nprojections, and aggregations; export the training dataset defined by such\nqueries; convert this dataset into the format of an external learning tool; and\ntrain the desired model using this tool. These integrated solutions are also a\nfertile ground of theoretically fundamental and challenging problems at the\nintersection of relational and statistical data models.\n This article introduces a unified framework for training and evaluating a\nclass of statistical learning models over relational databases. This class\nincludes ridge linear regression, polynomial regression, factorization\nmachines, and principal component analysis. We show that, by synergizing key\ntools from database theory such as schema information, query structure,\nfunctional dependencies, recent advances in query evaluation algorithms, and\nfrom linear algebra such as tensor and matrix operations, one can formulate\nrelational analytics problems and design efficient (query and data)\nstructure-aware algorithms to solve them.\n This theoretical development informed the design and implementation of the\nAC/DC system for structure-aware learning. We benchmark the performance of\nAC/DC against R, MADlib, libFM, and TensorFlow. For typical retail forecasting\nand advertisement planning applications, AC/DC can learn polynomial regression\nmodels and factorization machines with at least the same accuracy as its\ncompetitors and up to three orders of magnitude faster than its competitors\nwhenever they do not run out of memory, exceed 24-hour timeout, or encounter\ninternal design limitations.\n