Customer behavior clustering with neural networks using mutual information bottleneck
Organization
Wise Ltd
Abstract
At Wise, we have built a Variational Autoencoder-inspired probabilistic neural model to predict customer behavior, directly using the sequences of earlier customer actions as input. An important additional outcome is an embedding of customer action sequences into a latent space, allowing for a natural clustering of similar customers. To our knowledge, this is the first time a latent space approach has been used in characterizing customer behavior, using the customer action sequences directly (without pre-aggregation).
In this project, the student will investigate if the model would be improved by replacing the KL regularization loss in the model by the Information Bottleneck approach as described eg in the "Mutual Information Neural Estimation" paper by Belghasi et al (Yoshua Bengio being one of the co-authors) or the Mutual Information Gradient Estimation paper (ICLR 2020) . The student will implement the information bottleneck version of the model in Pytorch Lightning (the framework used for the current model) and compare the quality of the clustering of customer behavior embeddings in latent space induced by it, on synthetic data and actual Wise customer data.
In this project, the student will investigate if the model would be improved by replacing the KL regularization loss in the model by the Information Bottleneck approach as described eg in the "Mutual Information Neural Estimation" paper by Belghasi et al (Yoshua Bengio being one of the co-authors) or the Mutual Information Gradient Estimation paper (ICLR 2020) . The student will implement the information bottleneck version of the model in Pytorch Lightning (the framework used for the current model) and compare the quality of the clustering of customer behavior embeddings in latent space induced by it, on synthetic data and actual Wise customer data.
Graduation Theses defence year
2021-2022
Supervisor
Dr. Egor Kraev
Spoken language (s)
English
Requirements for candidates
Familiarity with neural networks, especially variational autoencoders and/or Bayesian techniques; some background in (or interest in learning) information theory.
Level
Masters
Application of contact
Name
Egor Kraev
Phone
E-mail