A Machine Learning Approach to Compare Causal Inference Modelling Strategies in the Digital Advertising Industry

Cause and effect relationships are not extensively explored in machine learning but prior to creating a system that understands causal relationships, there needs to be a general ability to grasp and inculcate cause-and-effect statistically. The way of choice for the proposed model is causal inference where the causes are inferred from the data. The study of causality, unlike correlation, falls in the domain of prescriptive analytics, that is, the likelihood of an occurrence given that another occurrence has happened. As a result, the study of causality and predictions based on causal inference are highly important when it comes to the domain of marketing and sales, largely because many practices in the domain of advertising try to fundamentally mimic causal inference such as A/B testing, primarily used in e-commerce and the digital advertising industry. For the purposes of the study, the treatment will be exposure to multiple advertisements and the uplift modelling will be used to direct marketing efforts to the most agreeable customers. Uplift modelling is a causal learning approach that estimates the dataset's individual treatment effect and using the same, calculates the incremental impact of a direct treatment on the individual behavior of the customer, which is why the advertising use-case felt particularly suitable. The usual way causal inference is inculcated in machine learning models is that one attribute is taken as a treatment, that is, the intervention or policy that has been done, while the target label is the desired outcome while accounting for the confounders which are the rest of the attributes. However, for the crux of this study, the proposed model will be taking two treatment attributes so as to see how the causality of two outcomes improves or worsens the impact of a rudimentary machine learning model. At the same time, this comparative study will be in tandem with two other models featuring one causal relationship as well as no causal relationship to truly understand the use of causality in this particular use-case. Furthermore, the modelling strategies will also include basic machine learning algorithms along with Base Meta Learners that are standard use in causal machine learning and compare results of those as well.