Napovedovanje obsega komentiranja spletnih novic z modeli strojnega učenja

VIDONI, MARKO (author), Zupan, Blaž (mentor)

Svetovni splet nas spremlja na vsakem koraku in si življenja brez stalnega dostopa do spleta ne znamo več predstavljati. Na splet se je v skladu s tem preselila tudi objava ter branje dnevnih novic in drugih člankov. Pomembno lastnost, ki jo ima na spletu objavljena novica, in jo v preteklosti tiskani mediji niso imeli, je možnost izražanja mnenja bralcev s komentarji, spisanih na temo novice. Spletnim portalom je tako v interesu, da so objavljene novice čim bolj komentirane in s tem tudi bolj brane, to pa posledično pripelje do večjega obiska strani. V diplomski nalogi je bil razvit napovedni sistem, s katerim lahko na podlagi besedila novic in dodatnih meta podatkov napovemo število komentarjev, ki jih bo prejela spletna novica v slovenščini. V našem primeru se je najbolje odrezala uporaba odebeljenega dela besedila novice, dodatnih meta podatkov in nadgradnja modela gradientnega boostinga regresijskih dreves, ki je gradila ločene modele za vsako kategorijo novic. Z analizo smo dokazali, da je s pravo predpripravo podatkov in uporabo naprednejših tehnik strojnega učenja mogoče uspešno napovedati število komentarjev novic. Poleg modelov samih in njihovega testiranja je bil proučen tudi vpliv značilk na napoved količine komentarjev in lastnosti novic z dobrimi in slabimi napovedmi.

• Language: Slovenian
• Keywords: Tekstovno rudarjenje, regresijski model, spletne novice
• Work type: Bachelor thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2016
• PID: 20.500.12556/RUL-83749
• Publication date in RUL: 28.06.2016

Secondary language

• Language: English
• Title: Machine learning models for predicting the volume of online news comments

Abstract

World Wide Web accompanies us on every step of our lives and we cannot anymore imagine our lives without the constant access to the internet. Publishing industry has also moved online where news articles are now published and read. The important novelty of online articles is the ability of readers to express their opinions about articles’ topics in a form of comments. It is in the best interest of web portals that the published web news are frequently commented and read, driving up the web portal visitors traffic. In this thesis a prediction system has been developed to predict the number of comments a news article in Slovene language will generate based on news text content and metadata. We got the best prediction results from bold parts of the text, coupled with metadata attributes and the extended gradient boosted regression trees model which builds separate models for each article category. Our analysis has proven that with proper data preprocessing and use of machine learning techniques it is possible to successfully predict the number of comments an article gets. In addition we also studied an influence of features on the prediction and properties of articles with good and bad prediction results.

• Keywords: Text mining, regression model, web news articles