Sequence classification applied to user log data An approach to identify characteristics of user sessions in a music streaming service

Abstract

Applying machine learning techniques to sequential user log data can provide insights about users that can guide companies towards making decisions that improve user experience. Recurrent neural networks have been proven to work well in combination with sequential data and recent research suggests that incorporating residual connections in recurrent structures outperforms standard recurrent structures. In this thesis, we show that residual recurrent neural networks can be applied to user log data from a complex domain in order to identify regularities in user behavior. To our knowledge, no research have been conducted with these model structures in domains other than text and image classification. A proof of concept is implemented in collaboration with Spotify where this approach is used to identify how users behave when they save music in the music streaming service. By conducting experiments with different models, we show that models with increased input complexity slightly outperform models with lower input complexity in the artificial classification task defined in this thesis. We also show that results from a more complex model can be analyzed and provide valuable insights. However, we conclude that the approach is ineffective and needs more developement in order to become sufficient.