Sapling Similarity: A performing and interpretable memory-based tool for recommendation

Many bipartite networks describe systems where an edge represents a relation between user and item. Measuring the similarity either users or items is basis of memory-based collaborative filtering, widely used method to build recommender system with purpose proposing users. When edges network are unweighted, popular common neighbors-based approaches, allowing only positive values, neglect possibility effect two (or items) being very dissimilar. Moreover, they underperform respect model-based (machine learning) although providing higher interpretability. Inspired by functioning Decision Trees, we propose compute that allows also negative Sapling Similarity. The key idea look at how information connected item influences our prior estimation probability another same item: if it reduced, then will be negative, otherwise positive. We show that, when Similarity provides better recommendations than existing metrics. Then compare Collaborative Filtering (SSCF, hybrid item-based user-based) state-of-the-art models using standard datasets. Even SSCF depends on one straightforward hyperparameter, has comparable recommending accuracy, outperforms all other Amazon-Book dataset, while retaining high explainability approaches.