Please use this identifier to cite or link to this item: http://artemis.cslab.ece.ntua.gr:8080/jspui/handle/123456789/17059
Title: Μοντέλα και πρόβλεψη συμπεριφοράς χρηστών σε δημοπρασίες επιχορηγούμενης αναζήτησης
Authors: Stamos, Filippos
Φωτάκης Δημήτριος
Keywords: Sponsored Search Auctions, Mechanism Design, Generalized Second-Price Auctions, Vickrey-Clarke-Groves Auctions, Cascade Model, Separable Model, Machine Learning, Hadoop
Issue Date: 27-Aug-2018
Abstract: This thesis focuses on the experimental evaluation of the user’s behaviour in sponsored search auctions. The majority of income of the biggest web companies like Google and Yahoo, is earned through auctions related to the advertisements of products. For example, in 2005, 98% of Google’s revenue derived from GSP (generalized second prize) auctions. In the first part of this thesis, we will study these auctions, their properties and we will compare GSP auctions with VCG (Vickrey-Clarke-Groves) auctions, another widely used type of auction. This thesis focuses on the sponsored search auctions problem happening on the web. Aiming to maximize the company’s revenue and to find the best possible allocation of ads, it becomes really important to be able to predict user’s behavior while browsing with the use of search engines. The two most widely used models, the Cascade Model and the Separable Model are studied and widely analyzed. The analysis of user’s behavior can be predicted through Machine Learning. That is why we continue with knowledge on the basics of Machine Learning. Although in the end Machine Learning methods weren’t used because of limited data, the understanding of metrics in order to evaluate the analyzed models was crucial during the analysis of the bibliographic models. In the last part of this thesis, models are evaluated experimentally. Yandex’s Person- alized Web Search dataset is used, analyzed with the use of Hadoop. As this data is the result of organic searches, we make the hypothesis that users behave the same way when being given organic and sponsored results. Firstly, the two main bibliography models are analyzed and compared extensively and in the end new models are tried. The model that we propose has similar accuracy with the other known models but is able to perform better at predicting when users click, while the rest of the models perform better at predicting when users don’t click.
URI: http://artemis.cslab.ece.ntua.gr:8080/jspui/handle/123456789/17059
Appears in Collections:Διπλωματικές Εργασίες - Theses

Files in This Item:
File Description SizeFormat 
Thesis_StamosFilippos.pdf1.5 MBAdobe PDFView/Open


Items in Artemis are protected by copyright, with all rights reserved, unless otherwise indicated.