Multi-task learning in programmatic advertising

VREŠ, DOMEN (author), Robnik Šikonja, Marko (mentor), Jakomin, Martin (co-mentor)

With the popularity of the world wide web, online advertising became crucial for the advertising industry. A large part of online advertising is based on real-time bidding, where automated software is used to buy and sell ad space. In this ecosystem, demand-side platforms (DSP) work with advertisers to display their ads on websites by bidding for ad space. Three crucial problems for DSP companies are viewability prediction, click-through rate (CTR) prediction, and conversion rate (CVR) prediction. In this work, we model these three problems using multi-task learning, which was not done before. We take the deep & cross network architecture, a state-of-the-art for CTR and CVR prediction, and expand it into a multi-task model. Additionally, we expand the multi-task model with recent techniques of cross-stitch layer, soft attention mask, uncertainty-based loss weighing, and relationship layer. We compare the multi-task model with single-task baselines on a large proprietary real-world data set. We show that the multi-task models significantly outperform the CVR baseline, which is our main goal, as CVR prediction is the most difficult task to model.

• Language: English
• Keywords: artificial intelligence, machine learning, multi-task learning, real-time bidding, viewability prediction, click-through rate prediction, conversion rate prediction, deep & cross network, cross-stitch layer, soft attention, uncertainty based loss weighing
• 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-151061
• COBISS.SI-ID: 170173187
• Publication date in RUL: 28.09.2023

Secondary language

• Language: Slovenian
• Title: Večciljno učenje v programatičnem oglaševanju

Abstract

S priljubljenostjo svetovnega spleta je postalo spletno oglaševanje ključnega pomena za oglaševalsko industrijo. Velik del spletnega oglaševanja temelji na realno-časovnih dražbah, kjer avtomatizirani programi nakupujejo in prodajajo oglasni prostor. V tem ekosistemu platforme za povpraševanje pomagajo oglaševalcem pri prikazovanju njihovih oglasov na spletnih mestih s sodelovanjem na teh dražbah. Trije ključni problemi, s katerimi se soočajo te platforme so: napoved verjetnosti ogleda oglasa, napoved verjetnosti klika na oglas in napoved verjetnosti konverzije. V tem delu kot prvi modeliramo te tri probleme z uporabo večciljnega učenja. V delu uporabimo globoko prečno mrežo, ki je sodoben model za napovedovanje klikov in konverzij, ter jo razširimo v večciljni model. Osnovni večcliljni model nadgradimo z nedavnimi tehnikami kot so prečna povezovalna plast, blaga pozornost, uteževanje funkcije izgube na podlagi negotovosti in plast odnosov. Večciljni model primerjamo z osnovnimi pristopi za posamezne probleme na veliki realni lastniški podatkovni množici. Rezultati pokažejo, da večciljni model znatno prekaša izhodiščni model za napovedovanje verjetnosti konverzij, kar je tudi naš glavni cilj, saj so konverzije med temi tremi problemi najtežje za modeliranje.

• Keywords: umetna inteligenca, strojno učenje, realno-časovne dražbe, napoved razmerja med prikazi in ogledi, napoved razmerja med prikazi in kliki, napoved razmerja med prikazi in konverzijami, globoka prečna mreža, prečno povezovalna plast, blaga pozornost, uteževanje funkcij izgube na podlagi negotovosti