Optimizing Click-Through Rate in Online Advertising Using Cost-Sensitive Learning

Petek, Bernarda (author), Bosnić, Zoran (mentor), Jakomin, Martin (co-mentor)

Real-time bidding is a type of online advertising, which displays personalized advertisements online to users based on their interests in real time. Demand-side platforms participate in such bidding for ad spaces. The bidding price, usually computed using a predicted click-through rate and target cost-per-click, reflects the value of an advertisement to the bidder. The primary focus of research in online advertising revolves around improving the prediction of click-through-rate. We focus on improving click-through-rate prediction for higher-cost advertisements, as they are typically less represented in the dataset, while keeping performance of lower-cost advertisements unaffected. Our approach involves the implementation of cost-sensitive machine learning by weighting the loss function. We explore various mappings of target cost-per-click values as weights. We train our model on the real-world like dataset, using weighted loss functions, resulting in different machine learning models. We evaluate the results with log-loss and calibration metrics. Our results reveal promising outcomes, indicating that some weights improve click-through-rate prediction for higher-cost advertisements while maintaining the quality for lower-cost advertisements.

• Language: English
• Keywords: real-time bidding, click-through rate prediction, cost-sensitive learning, custom loss function
• Work type: Master's thesis/paper
• Organization: FMF - Faculty of Mathematics and Physics
• Year: 2023
• PID: 20.500.12556/RUL-150023
• UDC: 004.4
• COBISS.SI-ID: 163981315
• Publication date in RUL: 13.09.2023

Secondary language

• Language: Slovenian
• Title: Optimizacija razmerja med prikazi in kliki pri oglaševanju s cenovno občutljivim učenjem

Abstract

Spletno oglaševanje omogoča oglaševalcem oglaševanje preko spleta. Dražbe v stvarnem času so en izmed načinov spletnega oglaševanja, ki omogočajo, da so uporabniku prikazani personalizirani oglasi. Na dražbah v stvarnem času v imenu oglaševalcev sodelujejo platforme za samodejno povpraševanje. Oglasi, s katerimi sodelujejo na dražbah, se delijo po dražbeni ceni, ki je odvisna od ciljne cene na klik in napovedane verjetnost za klik na oglas. Napoved verjetnosti klika za posamezen oglas je en izmed temeljnih izzivov v spletnem oglaševanju. Modeli, ki jo napovedujejo, obravnavajo vse oglase enakovredno, ne glede na njihovo ceno. To lahko privede do finančnih izgub. Poraja se vprašanje, ali bi bilo mogoče izboljšati napoved verjetnosti klika za dražje oglase in hkrati ohraniti kakovost napovedi za cenejše oglase. Problema se lotimo s cenovno občutljivim učenjem, tako da iščemo primerne uteži za funkcijo izgube. Z različnimi uteženimi funkcijami izgube učimo model za napovedovanje verjetnosti klika in rezultate evalviramo. Naši rezultati nakazujejo, da lahko uporaba ustreznih uteži izboljša razmerje med prikazi in kliki za dražje oglase in ohrani razmerje za cenejše oglase.

• Keywords: dražbe v stvarnem času, optimizacija razmerja med prikazi in kliki, cenovno občutljivo učenje, utežena funkcija izgube