Design and Evaluation of Recommendation Systems for Newsletters

Sammanfattning

This study explores the feasibility of using recommendation systems to deliver personalized newsletter suggestions to users on a digital platform. Three types of recommendation systems were implemented and evaluated: content-based, collaborative and hybrid. Due to the lack of data available for newsletter content, the content-based system was implemented using a neural network. Two approaches were tried, one where the network was trained on user behavior and another where the user’s reading behavior was incorporated using BERT. The collaborative model approach used the SVD Funk algorithm, trained on user interactions with articles to predict newsletter preferences. Finally a feature augmenting hybrid model was created by combining the content-based and collaborative models. All models were evaluated using four metrics: two ranking (MRR and Precision@5) and two for prediction error (RMSE and MAE). Results showed that the content-based neural network outperformed others on error metrics, while the collaborative model achieved the highest ranking performance with Precision@5. The hybrid model performed similarly to the content-based system across all metrics.