Uplift Modeling for Online Advertising. (Modélisation d'Uplift pour la Publicité en ligne)

Artem Betlei

Abstract: Uplift modeling is a machine learning-based technique for treatment effect prediction at the individual level, which has become one of the main trends in application areas where personalization is key, such as personalized medicine, performance marketing, social sciences, etc. This thesis is intended to expand the scope of uplift modeling for experimental data by developing new theory and solutions for several open challenges in the field, inspired by the online advertising applications perspective. Firstly we release a publicly available collection of 13.9 million samples collected from several randomized control trials, scaling up available datasets by a 210x factor. We formalize how uplift modeling can be performed with this data, along with relevant evaluation metrics. Then, we propose synthetic response surfaces and treatment assignment providing a general set-up for Conditional Average Treatment Effect (CATE) prediction and report experiments to validate key traits of the dataset. Secondly, we assume imbalanced treatment conditions and propose two new data representation-based methods inspired by cascade and multi-task learning paradigms. We provide then series of experimental results over several large-scale real-world collections to check the benefits of the proposed approaches. We then cover the problem of direct optimization of the Area Under the Uplift Curve (AUUC), a popular metric in the field. Using the relations between uplift modeling and bipartite ranking we provide a generalization bound for the AUUC and derive an algorithm optimizing this bound, usable with linear and deep models. We empirically study the tightness of the proposed bound, its efficacy for hyperparameters tuning, and investigate the performance of the method compared to a range of baselines on two real-world uplift modeling benchmarks. Finally, we consider the problem of learning uplift models from aggregated data. We propose a principled way to learn group-based uplift models from data aggregated according to a given set of groups that define a partition of the user space, using different unsupervised aggregation techniques, such as feature binning by value or by quantile. We proceed by introducing a bias-variance decomposition of the Precision when Estimating Heterogeneous Effect (PEHE) metric for models learned on a given grouping and show how this decomposition enables us to derive a theoretical optimal number of groups as a function of data size. Experimental results highlight the bias-variance trade-off and confirm theoretical insights concerning the optimal number of groups. In addition, we show that group-based uplift models can have comparable performance to baselines with full access to the data.