Athens University Economics and Business (AUEB)
Declarative specification of data analytics tasks
Abstract
dc:descriptionThe thesis studies the application of the declarative paradigm on the area of data analytics. Declarative programming is usually summarized in the phrase "describe what the program computes rather than how to compute it". As the adoption of data science grows rapidly, a need for democratizing data analysis tasks by making their development more approachable and less tedious through high-level languages has emerged. Inspired by the success of the declarative paradigm in relational database systems, researchers have recently started exploring whether the use of declarative languages in the domain of data analytics with a focus on machine learning (ML) algorithms can provide a productivity leap for developers. The first part of the thesis conducts a detailed survey of current systems in the area of data analytics and machine learning describing themselves as declarative. The main characteristics of their programming models, as well as the optimization techniques used in these systems, are analyzed through specific code examples. Based on these characteristics the thesis provides a classification of a wide range of systems in the aforementioned area and proposes a set of desired properties for a data analytics/ML system, in order to follow the objectives of the declarative paradigm closer. The second part of the thesis focuses on the expression of ML algorithms with database query languages and on their efficient compilation and execution on specialized mathematical optimization solvers and ML frameworks. The proposed approaches are based on the "model + solver" paradigm, where there is a description of the objective function (a.k.a. loss or cost function) of a machine learning model and a solver that provides the optimal solution for it. First, the thesis explores the use of Datalog as a declarative language to write the components of a linear machine learning algorithm, which is then transformed to a lower-level data representation/format consumed by a linear programming solver. The second approach proposes a method for translating objective functions of ML models written in SQL to linear algebra operators as provided by an ML specialized system, such as TensorFlow. This approach extends on the previous one in that it covers non-linear ML models and it automatically generates code using the primitives of an ML framework, rather than a data format. In both cases the computed solution is stored back to the database, which allows for generating future predictions on test data inside the database. In the context of these approaches, techniques for automating and optimizing data movement between the structures of the database (relations) and the ones of the ML framework (tensors) are proposed. The thesis demonstrates the proposed approaches through prototype systems implementing an end-to-end workflow that starts in the database, pushes the training of the model to an ML library/solver, and stores the computed weights back in the database. Experiments regarding the translation process and data movement optimization techniques on popular machine learning algorithms, including Linear Regression, Factorization Machines and Logistic Regression, and different datasets are presented.
Degree
thesis:*- Grantor dc:publisher
- Athens University Economics and Business (AUEB)
- Year dc:date
- 2020
Author and committee
dc:creator, dc:contributor.*- Authors dc:creator
-
- Makrynioti, Konstantina
- Μακρυνιώτη, Κωνσταντίνα
Subjects
dc:subject × 14- Δηλωτικός προγραμματισμός
- Δηλωτικές γλώσσες προγραμματισμού
- Ανάλυση δεδομένων
- Επιστήμη δεδομένων
- Μηχανική μάθηση
- Declarative programming
- Declarative programming languages
- Data analysis
- Data science
- Machine learning
- Φυσικές Επιστήμες
- Επιστήμη Ηλεκτρονικών Υπολογιστών και Πληροφορική
- Natural Sciences
- Computer and Information Sciences
Rights
- Language dc:language
- eng
Identifiers
dc:identifier.*- Identifier
- 10.12681/eadd/47340