國立虎尾科技大學 |

Machine Learning for Data Analytics : = Profiling, Querying, and Beyond.

紀錄類型:	書目-語言資料,手稿 : Monograph/item
正題名/作者:	Machine Learning for Data Analytics :/
其他題名:	Profiling, Querying, and Beyond.
作者:	Zhang, Yunjia.
面頁冊數:	1 online resource (242 pages)
附註:	Source: Dissertations Abstracts International, Volume: 86-02, Section: B.
Contained By:	Dissertations Abstracts International86-02B.
標題:	Computer engineering. -
電子資源:	click for full text (PQDT)
ISBN:	9798383581827

Machine Learning for Data Analytics : = Profiling, Querying, and Beyond.
Zhang, Yunjia.

Machine Learning for Data Analytics :Profiling, Querying, and Beyond. - 1 online resource (242 pages)

Source: Dissertations Abstracts International, Volume: 86-02, Section: B.

Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.

Includes bibliographical references

Data analytics is crucial for identifying and interpreting meaningful patterns within data sets. The insights derived from this process can be effectively utilized to make informed decisions across various sectors that handle data. For instance, in healthcare, data analytics can be used to analyze trends from medical records, thereby predicting future healthcare needs. The importance of data analytics has further significantly increased alongside the exponential growth in data volume.With machine learning (ML) emerging as a transformative approach across various fields, this dissertation investigates how it can enhance the data analytics pipeline. Our research focuses on three primary questions: (1) How can ML enhance data profiling? (2) Can ML increase the efficiency of query execution in database management systems (DBMS)? and (3) How can ML-based systems improve question-answering capabilities within databases?For the first question, we initially address a fundamental challenge in data profiling: functional dependency discovery. We introduce FDX, an algorithm designed to efficiently uncover true functional dependencies from noisy data via structure learning. By employing a structured probabilistic model to capture dependencies, we demonstrate that discovering functional dependencies is equivalent to learning the structure of the model. Our experimental results indicate that FDX can effectively and efficiently identify functional dependencies in both synthetic and real datasets.Shifting focus to the industrial application of data profiling, we encounter scenarios where data analytics providers cannot access customer data instances directly. In these cases, we explore schema matching based solely on schema-level information to enhance data profiling quality. We developed a language model-based semi-automatic schema matching tool, namely LSM, that significantly reduces human effort in the matching process. LSM utilizes a language model-based featurizer that evaluates attribute pairs and outputs a similarity score. In real-world applications with Microsoft customers, LSM has reduced human effort in the schema matching task by up to 81%.Next, we explore the second question: Can ML enhance the efficiency of analytical query execution in DBMS? Our analysis of existing ML-based query optimizers shows that while they exhibit potential, their ability to generalize to unseen queries is often limited. Given that these ML-based optimizers utilize mechanisms similar to those in adaptive query execution techniques - specifically lookahead information passing (LIP) and adaptive join algorithm (AJA) - we integrate LIP+AJA into PostgreSQL for a comparative study. Surprisingly, we find that this simple adaptive query processing approach is not only competitive to the SOTA ML-based approaches but, in some cases, outperforms the ML-based approaches. Moreover, the adaptive approach is particularly attractive because it avoids the need for expensive training and offers full interpretability, unlike ML-based query optimization methods.Finally, for the third question, we investigate the development of ML-based table question answering systems, designed to provide data analysts with a more efficient and straightforward question-answering pipeline. Our initial focus is on accurately answering queries that involve only one table. We introduce a new large language model (LLM) prompting framework, ReAcTable, which is built upon the principles of existing React and Chain-of-Thought (CoT) frameworks. ReAcTable leverages out-of-the-box LLMs to process questions in a step-by-step manner. At each step, the LLM is prompted to either generate code or provide a direct answer to the question. If code is generated, its execution results are then fed back into the LLM to inform subsequent steps. Our evaluations show that ReAcTable achieves state-of-the-art performance across multiple open datasets, significantly enhancing the capabilities of table question answering systems.As data volumes expand, the challenge of identifying relevant datasets within vast data lakes becomes increasingly complex. Data analysts may not even know which tables are required to answer their queries. To tackle this issue, we introduce DataDisco, a system designed to identify and retrieve relevant tables from data lakes. DataDisco segments the extensive data lake into smaller, manageable sub-lakes and fine-tunes a smaller but specialized language model for each sub-lake. To assess the effectiveness of DataDisco, we create a novel dataset, BirdSpider, by adapting the existing Bird and Spider datasets specifically for the table discovery task. Our experimental results show that DataDisco outperforms the latest GPT-4 models on the BirdSpider dataset with significantly smaller amount of parameters.

Electronic reproduction.
Ann Arbor, Mich. :
ProQuest,
2024

Mode of access: World Wide Web

ISBN: 9798383581827Subjects--Topical Terms:

569006
Computer engineering.
Subjects--Index Terms:

Data analyticsIndex Terms--Genre/Form:

554714
Electronic books.

Machine Learning for Data Analytics : = Profiling, Querying, and Beyond.
LDR:06231ntm a22003977 4500 001 1150533
005 20241028051516.5
006 m o d
007 cr bn ---uuuuu
008 250605s2024 xx obm 000 0 eng d
020 $a 9798383581827
035 $a (MiAaPQ)AAI31486554
035 $a AAI31486554
040 $a MiAaPQ $b eng $c MiAaPQ $d NTU
100 1 $a Zhang, Yunjia. $3 1477026
245 1 0 $a Machine Learning for Data Analytics : $b Profiling, Querying, and Beyond.
264 0 $c 2024
300 $a 1 online resource (242 pages)
336 $a text $b txt $2 rdacontent
337 $a computer $b c $2 rdamedia
338 $a online resource $b cr $2 rdacarrier
500 $a Source: Dissertations Abstracts International, Volume: 86-02, Section: B.
500 $a Advisor: Patel, Jignesh M.;Rekatsinas, Theodoros.
502 $a Thesis (Ph.D.)--The University of Wisconsin - Madison, 2024.
504 $a Includes bibliographical references
520 $a Data analytics is crucial for identifying and interpreting meaningful patterns within data sets. The insights derived from this process can be effectively utilized to make informed decisions across various sectors that handle data. For instance, in healthcare, data analytics can be used to analyze trends from medical records, thereby predicting future healthcare needs. The importance of data analytics has further significantly increased alongside the exponential growth in data volume.With machine learning (ML) emerging as a transformative approach across various fields, this dissertation investigates how it can enhance the data analytics pipeline. Our research focuses on three primary questions: (1) How can ML enhance data profiling? (2) Can ML increase the efficiency of query execution in database management systems (DBMS)? and (3) How can ML-based systems improve question-answering capabilities within databases?For the first question, we initially address a fundamental challenge in data profiling: functional dependency discovery. We introduce FDX, an algorithm designed to efficiently uncover true functional dependencies from noisy data via structure learning. By employing a structured probabilistic model to capture dependencies, we demonstrate that discovering functional dependencies is equivalent to learning the structure of the model. Our experimental results indicate that FDX can effectively and efficiently identify functional dependencies in both synthetic and real datasets.Shifting focus to the industrial application of data profiling, we encounter scenarios where data analytics providers cannot access customer data instances directly. In these cases, we explore schema matching based solely on schema-level information to enhance data profiling quality. We developed a language model-based semi-automatic schema matching tool, namely LSM, that significantly reduces human effort in the matching process. LSM utilizes a language model-based featurizer that evaluates attribute pairs and outputs a similarity score. In real-world applications with Microsoft customers, LSM has reduced human effort in the schema matching task by up to 81%.Next, we explore the second question: Can ML enhance the efficiency of analytical query execution in DBMS? Our analysis of existing ML-based query optimizers shows that while they exhibit potential, their ability to generalize to unseen queries is often limited. Given that these ML-based optimizers utilize mechanisms similar to those in adaptive query execution techniques - specifically lookahead information passing (LIP) and adaptive join algorithm (AJA) - we integrate LIP+AJA into PostgreSQL for a comparative study. Surprisingly, we find that this simple adaptive query processing approach is not only competitive to the SOTA ML-based approaches but, in some cases, outperforms the ML-based approaches. Moreover, the adaptive approach is particularly attractive because it avoids the need for expensive training and offers full interpretability, unlike ML-based query optimization methods.Finally, for the third question, we investigate the development of ML-based table question answering systems, designed to provide data analysts with a more efficient and straightforward question-answering pipeline. Our initial focus is on accurately answering queries that involve only one table. We introduce a new large language model (LLM) prompting framework, ReAcTable, which is built upon the principles of existing React and Chain-of-Thought (CoT) frameworks. ReAcTable leverages out-of-the-box LLMs to process questions in a step-by-step manner. At each step, the LLM is prompted to either generate code or provide a direct answer to the question. If code is generated, its execution results are then fed back into the LLM to inform subsequent steps. Our evaluations show that ReAcTable achieves state-of-the-art performance across multiple open datasets, significantly enhancing the capabilities of table question answering systems.As data volumes expand, the challenge of identifying relevant datasets within vast data lakes becomes increasingly complex. Data analysts may not even know which tables are required to answer their queries. To tackle this issue, we introduce DataDisco, a system designed to identify and retrieve relevant tables from data lakes. DataDisco segments the extensive data lake into smaller, manageable sub-lakes and fine-tunes a smaller but specialized language model for each sub-lake. To assess the effectiveness of DataDisco, we create a novel dataset, BirdSpider, by adapting the existing Bird and Spider datasets specifically for the table discovery task. Our experimental results show that DataDisco outperforms the latest GPT-4 models on the BirdSpider dataset with significantly smaller amount of parameters.
533 $a Electronic reproduction. $b Ann Arbor, Mich. : $c ProQuest, $d 2024
538 $a Mode of access: World Wide Web
650 4 $a Computer engineering. $3 569006
650 4 $a Computer science. $3 573171
653 $a Data analytics
653 $a Database
653 $a Machine learning
653 $a Query optimization
653 $a Table question answering
655 7 $a Electronic books. $2 local $3 554714
690 $a 0984
690 $a 0800
690 $a 0464
710 2 $a The University of Wisconsin - Madison. $b Computer Sciences. $3 1179878
710 2 $a ProQuest Information and Learning Co. $3 1178819
773 0 $t Dissertations Abstracts International $g 86-02B.
856 4 0 $u http://pqdd.sinica.edu.tw/twdaoapp/servlet/advanced?query=31486554 $z click for full text (PQDT)