Representation Learning Methods for Sequential Information in Marketing and Customer Level Transactions

The rapid growth of data generated by businesses has surpassed human capabilities to produce actionable insights. Modern marketing applications depend on vast amounts of customer labelled data and supervised machine learning algorithms to predict customer behaviour and their potential next actions. However, this process requires significant effort in data pre-processing and the involvement of domain experts, which can be costly and time-consuming. This work reviews representation learning techniques as an alternative approach to feature engineering, aiming to eliminate the need for hand-crafted features and accelerate the process of extracting insights from data. Techniques such as Bayesian neural networks, general embeddings, and encoding-decoding architectures are explored to compress information obtained directly from raw input data into a dense probabilistic space. This thesis introduces the necessary technical aspects of neural networks and representation learning, from traditional methods like principal component analysis (PCA) and embeddings, to latent variable and generative methods that use deep neural networks, such as variational auto-encoders and Bayesian neural networks. It also explores the theoretical background of survival analysis and recommender systems, which serve as the foundation for the applications presented in this work to predict when individuals are likely to stop their relationship with businesses in a non-contractual settings or which items individuals are the most likely to interact with in their next purchase. Experiments conducted on real-world retail and benchmark datasets demonstrate comparable results in terms of predictive performance and superior computational efficiency when compared to existing methods.