Reproducibility and Ethical Evaluation of a Session-Based Recommendation System in the Insurance Domain

This paper reproduces and critically assesses the research presented in "Learning Recommendations from User Actions in the Item-poor Insurance Domain." The original study introduces a cross-session-based recommendation model to address data sparsity in the insurance industry. Given the importance of reproducibility in machine learning, this work validates the findings, evaluates transparency, and examines the feasibility of implementing the proposed methods. The study attempts to replicate the dataset, implements the cross-session recurrent neural network (RNN) model, and compares it against baselines such as session-based k-nearest neighbors (SKNN), GRU4REC, and demographic-based models. The experiment investigates (1) the sufficiency of the original paper's details for reproduction, (2) the presence of statistically significant differences across experimental settings, and (3) whether the reproduced results align with the expected statistical distribution.