Forgetting on Evolving Graphs for Accurate and Diverse Stream-Based Recommendation
Murilo Falleiros Lemos Schmitt, Eduardo Jaques Spinosa · 2022
Stream-based recommender systems are an active research field, relying on incremental algorithms to update models by incorporating new data on a single pass, discarding such data after processing. A limitation of solely including new data is the accumulation of obsolete concepts, which eventually raises accuracy and scalability concerns. In this work, we propose a gradual forgetting technique for incremental neighborhood-based methods that locally forgets items based on recency and popularity, by decreasing importance of neighborhood of items for every incoming observation to emphasize more recent and reinforced ones. The technique includes parameters to increase diversity, by retaining less popular yet relevant items, and scalability, by pruning obsolete connections not reinforced by new data. Experiments conducted by extending a recent incremental graph-based approach highlight the effectiveness of the proposed technique, as its application improved scalability and diversity, outperforming baselines.