master's thesis
Jan Hartman (Author), Jure Demšar (Mentor), Erik Štrumbelj (Co-mentor)

Abstract

Real-time bidding is a fast-growing part of online advertising in which ad space on websites is sold in real-time while the page is still loading. The ad space is sold in auctions where several bidders compete. One of the central problems in RTB is click-through rate prediction, which has to deal with censored data -- since the bidders do not receive data about the auctions they lose, the predictive models cannot learn from them. To tackle this problem, we propose two strategies that explore by buying more ad impressions on unknown parts of the market. The proposed strategies use either hand-crafted insights or model uncertainty to guide the exploration. To test the strategies in the real world, we conducted A/B tests on the production traffic of Zemanta, a DSP in the RTB ecosystem. We also compared the obtained models' performances offline. Our results show that exploring the market through publishers did not bring significant improvements to the business or the model metrics. On the other hand, exploring with the uncertainty of the predictions showed increases in revenue and CTR as well as improvements in model performance metrics, indicating that using the uncertainty of the CTR model for exploration can be beneficial.

Keywords

censored data;click-through rate prediction;real-time bidding;incremental learning;big data;demand-side platform;computer science;computer and information science;master's thesis;

Data

Language: English
Year of publishing:
Typology: 2.09 - Master's Thesis
Organization: UL FRI - Faculty of Computer and Information Science
Publisher: [J. Hartman]
UDC: 004(043.2)
COBISS: 79729411 Link will open in a new window
Views: 300
Downloads: 174
Average score: 0 (0 votes)
Metadata: JSON JSON-RDF JSON-LD TURTLE N-TRIPLES XML RDFA MICRODATA DC-XML DC-RDF RDF

Other data

Secondary language: Slovenian
Secondary title: Uporaba raziskovanja trga za reševanje problema cenzuriranih podatkov v realnočasovnih avkcijah
Secondary abstract: Realnočasovne avkcije (RTB) so hitro rastoči del spletnega oglaševanja, v katerem se oglasni prostor na spletnih mestih prodaja v realnem času, medtem ko se stran še nalaga. Oglasni prostor se prodaja na dražbah, na katerih tekmuje več dražiteljev. Ena osrednjih težav v RTB-ju je napovedovanje klikov, ki ga ovira cenzuriranost podatkov -- ker dražitelji ne prejmejo podatkov o izgubljenih dražbah, se napovedni modeli iz njih ne morejo učiti. Za reševanje tega problema predlagamo dve strategiji, ki raziskujeta z nakupovanjem več prikazov oglasov na neznanih delih trga. Predlagani strategiji za vodenje raziskovanja uporabljata bodisi ročno izdelane vpoglede bodisi negotovost modelov. Za preizkušanje strategij v resničnem svetu smo izvedli A/B teste na produkcijskem prometu podjetja Zemanta, ki je DSP v ekosistemu RTB. Primerjali smo tudi uspešnosti pridobljenih modelov. Naši rezultati kažejo, da raziskovanje trga preko spletnih založnikov ni prineslo bistvenih izboljšav v poslu ali metrikah uspešnosti modelov. Po drugi strani je raziskovanje z negotovostjo napovedi pokazalo povečanje prihodkov in CTR-ja ter izboljšanje metrik uspešnosti modelov, kar kaže, da je lahko uporaba negotovosti napovednega modela za raziskovanje koristna.
Secondary keywords: cenzurirani podatki;napovedovanje klikov;realnočasovne avkcije;inkrementalno učenje;velepodatki;platforma za povprašanje;računalništvo in informatika;magisteriji;Računalništvo;Univerzitetna in visokošolska dela;
Type (COBISS): Master's thesis/paper
Study programme: 1000471
Thesis comment: Univ. v Ljubljani, Fak. za računalništvo in informatiko
Pages: VI, 69 str.
ID: 13595013