Expressed Emotions in Session Based News Article Recommender Systems

Abstract

Existing news recommenders that incorporate expressed emotions do not capture temporal and emotional changes over time. An active session approach, which repeatedly trains for five hours and evaluates on the next hour, allows for capturing temporal changes such as breaking news, trends, and changes in user interests. This thesis introduces a novel approach that incorporates expressed emotions into news article recommenders. Various scopes of historical clicks are studied, spanning from many historical clicks to only one historical click. The Mind and Adressa datasets are used for evaluation along with various ranking metrics. The results indicate that incorporating expressed emotion in an active session approach can significantly improve news recommender ranking metrics of models such as ST-NRMS, SR-GNN, and Item-KNN for user-level and session-level datasets.