Presentation meeting Display ads problem

Friday, Mar 24, 12:30h, in Pav. K16., Jason Rhuggenaath presents his recent work on the display ads allocation problem.

In this paper we study the the display ads allocation problem faced by online publishers. Many online (web) publishers such as news sites have multiple pages (homepage, sports, financial, etc) where they display different advertisements (ads). Ads can be displayed in different ways such as images, video or text ads and these ads are displayed on specific AD-slots that are on specific webpages. Each time a user visits a webpage that contains AD-slots a number of impressions are generated.

There are multiple ways to allocate the impressions and the resulting allocation needs to satisfy a number of constraints. In this paper we focus on the situation of online
publishers that are small and medium size enterprises (SMEs). Two options that these small online publishers usually have to allocate impressions are to (i) enter in guaranteed contracts with advertisers, or to (ii) sell an impression on the online AD auction market via supply side platforms (SSPs).

The online publisher offers the impression to the SSP and the SSP will try to sell it on the online AD market via an auction. Revenue from such a sale to an SSP is expected to vary due to the differences in SSPs such as the connected advertisers and the organization of the auctions. However the sale of the impression is not certain and there multiple possible  factors that could determine the success of an auction such as the relevant advertisers being out of budget or this particular impression being uninteresting for the advertisers connected to the SSP.

The allocation problem of the online publisher consists of finding a feasible allocation
of impressions between guaranteed contracts and SSPs that maximizes the expected
revenue.

We argue that the structure and properties of the problem, in particular the way that information is revealed over time, allows us to adopt a stochastic programming approach and formulate the allocation problem as an integer linear programming (ILP) model.

We compare the quality of the solutions obtained by the ILP model with the solutions of a greedy heuristic that is used in practice.