izpis_h1_title_alt

A comparison of click-through rate prediction models in real-time-bidding
ID Alič, Anže (Author), ID Demšar, Jure (Mentor) More about this mentor... This link opens in a new window, ID Hartman, Jan (Comentor)

.pdfPDF - Presentation file, Download (1,51 MB)
MD5: 678BAE280BDF5A3E4F1032A426878194

Abstract
Accurately predicting user clicks is crucial for the success of online advertising in the real-time bidding industry. In this thesis, our goal is to conduct a thorough comparison and evaluation of click-through rate prediction models commonly used in practice. Our advantage lies in our unified implementation approach, which allows for a fair and comprehensive comparison of the models, with the most important contribution being an online A/B test. We evaluated several click-through rate prediction models. Our offline results showed that the DCN-V2 model outperforms the other models in terms of log loss, with DeepFM and DeepFwFM following behind. However, due to domain constraints, the models were subject to underfitting. We demonstrated that in this regard, the DCN-V2 model is again the best as it was the least affected by it. We conducted our evaluation on both a public and a private dataset owned by Outbrain and obtained consistent results. Additionally, we conducted an online A/B test on the production traffic of Outbrain, a demand-side platform in the real-time bidding ecosystem. Our results showed that the DCN-V2 model outperformed the production DeepFM-based model, resulting in a revenue increase of 5.8\%.

Language:English
Keywords:artificial intelligence, machine learning, click-through rate prediction, real-time bidding, incremental learning, big data, demand-side platform
Work type:Master's thesis/paper
Typology:2.09 - Master's Thesis
Organization:FRI - Faculty of Computer and Information Science
Year:2023
PID:20.500.12556/RUL-146737 This link opens in a new window
COBISS.SI-ID:158062595 This link opens in a new window
Publication date in RUL:09.06.2023
Views:737
Downloads:128
Metadata:XML DC-XML DC-RDF
:
Copy citation
Share:Bookmark and Share

Secondary language

Language:Slovenian
Title:Primerjava modelov za napovedovanje verjetnosti klika v realno-časovnih dražbah
Abstract:
Natančno napovedovanje klikov je ključnega pomena za uspeh spletnega oglaševanja v industriji ralno-časovnih dražb. V tem delu je naš cilj izvesti temeljito primerjavo in vrednotenje modelov za napovedovanje verjetnosti klika, ki so pogosto uporabljeni v industriji. Naša prednost je v enotni implementaciji modelov, ki omogoča pošteno in celovito primerjavo modelov, pri čemer je najpomembnejši prispevek te naloge produkcijski A/B test. Evalvirali smo več modelov za napovedovanje verjetnosti klika. Naši rezultati v testnem okolju so pokazali, da model DCN-V2 dosega boljše rezultate kot druga dva modela v smislu logaritmične funkcije izgube. Vendar pa so bili modeli zaradi domenskih omejitev podvrženi podprilagajanju. Pokazali smo, da je model DCN-V2 znova najboljši, saj je nanj podprilaganjanje imelo manjši učinek. Analizo smo opravili tako na prosto dostopnih kot zasebnih podatkih podjetja Outbrain in dosegli konsistentne rezultate. Dodatno smo opravili A/B test na produkcijskem prometu podjetja Outbrain, ki je platforma za povpraševanje v ekosistemu realno-časovnih dražb. Pokazali smo, da je model DCN-V2 dosegel boljše rezultate kot model, ki temelji na metodi DeepFM, in da je povečal prihodke za 5,8\%.

Keywords:umetna inteligenca, strojno učenje, napovedovanje klikov, realno-časovne dražbe, inkrementalno učenje, velepodatki, platforma za povpraševanje

Similar documents

Similar works from RUL:
Similar works from other Slovenian collections:

Back