Optimizing Real-Time Bidding Systems in Programmatic Advertising

Možina, Marko (author), Demšar, Jure (mentor), Jakomin, Martin (comentor)

A demand-side platform has only tens of milliseconds per request to choose the advertisement it will bid with, so the selection runs in stages. The first stage carries much of that weight. A computationally cheap two-tower model scores the large set of eligible candidates and forwards about a hundred of them to heavier processes. Such a model has to tell good candidates from bad ones, yet its training data holds only the advertisements that were chosen in the past. We therefore sample the bad examples from the training batch, and that sampling introduces several kinds of bias into the model. In this thesis we reworked the negative sampling of preselection in a production real-time bidder. We tuned how many negative examples a batch supplies, corrected the bias with a streaming estimate of sampling frequencies, and widened the negatives with uniform draws from the whole set of advertisements. Drawing from two sources calls for a correction that the literature only hints at, so we derived it exactly as a single mixture distribution and showed that it works. We checked every method in a controlled setting on static data and tested it in the live production system. Our sampling methods improved the ranking of candidates decisively, but that improvement did not translate into clear gains in the financial metrics. We attribute the mismatch to the oversized depth of the forwarded shortlist, which masks better ordering before it can reach a business outcome.

• Language: English
• Keywords: programmatic advertising, real-time bidding, candidate retrieval, two-tower models, sampled softmax, mixed negative sampling, sampling bias correction, online A/B testing
• Work type: Master's thesis/paper
• Organization: FRI - Faculty of Computer and Information Science
• Year: 2026
• Publication date in RUL: 28.09.2026

Secondary language

• Language: Slovenian
• Title: Optimizacija realnočasovnih sistemov za dražbe v programatičnem oglaševanju

Abstract

Platforma na strani povpraševanja ima za posamezno zahtevo le nekaj deset milisekund za izbiro oglasa, s katerim bo licitirala, zato izbor poteka v več stopnjah. Posebej pomembna je prva, kjer računsko nezahteven dvostolpni model oceni veliko število primernih kandidatov, in jih nato približno sto posreduje zahtevnejšim procesom. Tak model mora zato znati ločiti med dobrimi in slabimi primeri, a ker njegova učna množica vsebuje le primere preteklih uspešno izbranih oglasov, slabe primere vzorčimo iz učnega paketa. Vendar pa tako vzorčenje v model vpelje več vrst pristranskosti. V tem magistrskem delu smo prenovili vzorčenje predizbora v produkcijskem sistemu za licitiranje v realnem času. Uravnali smo število negativnih primerov v paketu, pristranskost popravili s sprotnim ocenjevanjem pogostosti vzorčenja, negativne primere pa razširili z enakomernimi vzorci iz nabora vseh oglasov. Vzorčenje iz dveh virov zahteva popravek, ki je v literaturi le nakazan, zato smo ga natančno izpeljali in dokazali njegovo učinkovitost. Vsako metodo smo preverili v kontroliranem okolju na statičnih podatkih in testirali v pravem produkcijskem okolju. Čeprav so naše metode vzorčenja občutno izboljšale razvrščanje kandidatov, se to v finančnih metrikah ni odražalo. To neskladje pripisujemo prekomerni globini posredovanega izbora, ki zakrije izboljšano razvrstitev in prepreči, da bi pripomogla k boljšemu poslovnemu izidu.

• Keywords: programatično oglaševanje, licitiranje v realnem času, pridobivanje kandidatov, dvostolpni modeli, vzorčena funkcija softmax, mešano vzorčenje negativnih primerov, popravek pristranskosti vzorčenja, A/B testiranje v produkciji