Tema da Tese: Modern optimization of predictive models: a case study in mobile performance marketing

Autor: Pedro José Silva Pereira

Orientadores:  Paulo Alexandre Ribeiro Cortez, Rui Manuel Ribeiro de Castro Mendes

Data: 07/12/2021

Programa Doutoral: Programa Doutoral em Tecnologias e Sistemas de Informação

Abstract: 

The increase of mobile devices usage has leveraged digital business opportunities, particularly in the advertising sector. In particular, the Mobile Performance Marketing (MPM) industry has been a target of vast investments in the last years. This industry deals with advertising campaigns (owned by advertisers) presented on digital spaces (owned by publishers), such as mobile applications. In short, users are redirected to campaign pages in order to access publishers content. Between advertisers and publishers, intermediary companies perform matching between users and campaigns, by means of a Demand-Side Platform (DSP), aiming to lead the users to a purchase. However, the currently used mechanisms to perform this matching are rather rigid, using eventual profit or simple statistical rules to select the campaign to be displayed, instead of using other data attribute relationships. Overall, this DSP assignment method tends to produce a tiny conversion rate (e.g., 1%) between user redirects and purchases. Thus, improving the DSP performance with the application of a Machine Learning (ML) approach could potentially improve this business.

ML has proved to be extremely important in several domains, by being able to extract complex patterns from data and providing useful insights to decision makers. During this PhD, we particularly address the multi-objective Modern Optimization of predictive models, a much less researched topic when compared with traditional ML approaches. In particular, we use Evolutionary Algorithms (EAs) to design and evolve ML models, using MPM as the final (and target) use case. At an initial stage of this PhD thesis, when we did not had access to MPM data, we addressed a distinct ML use case. In effect, we proposed several neuroevolution models for multi-step ahead time series Prediction Intervals (PIs), based on a Pareto-based multi-objective EA that simultaneously considered PI coverage and width. This research included the proposal of a novel and robust evaluation method for multi-objective ML models, which vertically aggregates similar solutions using the Wilcoxon median and 95% confidence intervals. At a later stage, when MPM data was available, we developed two innovative approaches for the design and creation of Evolutionary Decision Trees (EDTs), using a multi-objective implementation of Grammatical Evolution (GE) that simultaneously considered EDTs predictive performance and complexity. Using a realistic and robust experimentation, with real-world data from the OLAmobile company, we have shown that the proposed GE is capable of evolving valuable prediction models for the MPM domain.