Struktureeritud andmetest duplikaatide eemaldamine Apache Spark'iga

Abstract

Üha rohkem avaldatakse veebis struktureeritud sisu, mis on loetav nii inimeste kui masinate poolt. Tänu otsimootorite loojatele, kes on defineerinud standardid struktureeritud sisu esitamiseks, teevad järjest rohkemad veebisaidid osa oma andmetest, nt toodete, isikute, organisatsioonide ja asukohtade kirjeldused, veebis avalikuks. Selleks kasutatakse RDFa, microdata jms vorminguid. Microdata on üks viimastest vormingutest ning saanud populaarseks suhteliselt lühikese aja jooksul. Sarnaselt on arenenud tehnoloogiad veebist struktureeritud sisu kättesaamiseks. Näiteks on Apache Any23, mis võimaldab veebilehtedest microdata andmeid eraldada ja linkandmetena kättesaadavaks teha. Samas pole struktureeritud andmete veebist kättesaamine enam suurim tehniline väljakutse. Nimelt on veebist saadud andmeid enne kasutamist vaja puhastada - eemaldada duplikaadid, lahendada ebakõlad ning hakkama tuleb saada ka ebamääraste andmetega.\n\rKäesoleva magistritöö peamiseks fookuseks on efektiivse lahenduse loomine veebis leiduvatest linkandmetest duplikaatide eemaldamine suurte andmekoguste jaoks. Kuigi deduplikeerimise algoritmid on saavutanud suhtelise küpsuse, tuleb neid konkreetsete andmekomplektide jaoks siiski peenhäälestada. Eelkõige tuleb tuvastada sobivaim võtme pikkus kirjete sortimiseks. Käesolevas töös tuvastatakse optimaalne võtme pikkus veebis leiduvate tooteandmete deduplikeerimise kontekstis. Suurte andmemahtude tõttu kasutatakse Apache Spark'i deduplikeerimist hajusalgoritmide realiseerimiseks.

The web is transforming from traditional web to web of data, where information is presented in such a way that it is readable by machines as well as human. As a part of this transformation, every day more and more websites implant structured data, e.g. product, person, organization, place etc., into the HTML pages. To implant the structured data different encoding vocabularies, such as RDFa, microdata, and microformats, are used. Microdata is the most recent addition to these structure data embedding standards, but it has gained more popularity over other formats in less time. Similarly, progress has been made in the extraction of the structured data from web pages, which has resulted in open source tools such as Apache Any23 and non-profit Common Crawl project. Any23 allows extraction of microdata from the web pages with less effort, whereas Common Crawl extracts data from websites and provides it publically for download. In fact, the microdata extraction tools only take care of parsing and data transformation steps of data cleansing. Although with the help of these state-of-the-art extraction tools microdata can be easily extracted, before the extracted data used in potential applications, duplicates should be removed and data unified. Since microdata origins from arbitrary web resources, it has arbitrary quality as well and should be treated correspondingly. \n\rThe main purpose of this thesis is to develop the effective mechanism for deduplication of microdata on the web scale. Although the deduplication algorithms have reached relative maturity, however, these algorithm needs to be executed on specific datasets for fine-tuning. In particular, the need to identify the most suitable length of sorting key in sorted-based deduplication approach. The present work identifies the optimum length of the sorting key in the context of extracted product microdata deduplication. Due to large volumes of data to be processed continuously, Apache Spark will be used for implementing the necessary procedures.