Optimizing online advertising space bidding

V procesu nakupovanja spletnega oglasnega prostora dandanes prevladuje tako imenovano programatično oglaševanje. Gre za trženje oglasnega prostora preko avtomatiziranih dražb v realnem času. Za uspešno sodelovanje na dražbah je potrebno usklajeno delovanje različnih procesov, v katerih ključno vlogo igrajo sodobne metode strojnega učenja. Eden izmed ključnih procesov je modeliranje tržne cene prikaza, saj lahko le-ta pomaga pri določitvi ponudbe za določen oglasni prostor. Na ta aspekt programatičnega oglaševanja se osredotoča naša magistrska naloga. Algoritme za modeliranje tržne cene prikaza lahko razdelimo v dve skupini. V prvo spadajo algoritmi, ki napovedo celotno gostoto verjetnosti zmage v odvisnosti od višine ponudbe. V drugo skupino spadajo algoritmi, ki napovedo verjetnost zmage samo pri dani vrednosti ponudbe. V nalogi smo implementirali in primerno ovrednotili več algoritmov iz obeh skupin. Za potrebe temeljitega ovrednotenja smo razvili novo metodo, ki algoritme primerja na podlagi zgrajenih referenčnih gostot verjetnosti in Kullback-Lieblerjeve divergence. Rezultati kažejo, da algoritmi, ki napovedujejo celotno gostoto verjetnosti zmage v odvisnosti od višine ponudb, dosegajo občutno boljše rezultate. Poleg tega ti algoritmi za podajanje napovedi potrebujejo manj časa.Programmatic advertising is the automated process of selling and buying online advertising space in real time, commonly referred to as the real-time bidding. Successful collaboration in real time bidding requires coordinated work of several processes, in which modern machine learning approaches play the crucial role. One of these is the modeling of the market price, which can in later stages help with identifying the optimal bid for a given advertising space. We divide the algorithms for modeling the market price into two groups, the algorithms that model the entire probability distribution of the market price and the pointwise algorithms that predict the probability of winning only at a given bid value. In this work we have implemented and experimentally evaluated several algorithms from both groups. We have also proposed a new method for evaluating the predicted probability distributions that compares algorithms based on the generated reference probability distribution and the Kullback-Liebler divergence measure. Our experiments show that algorithms that predict the entire probability distribution preform much better. Moreover, this type of algorithms require less time for the inference process than the pointwise algorithms