Semi-metric networks for recommender systems

Weighted graphs obtained from co-occurrence in user-item relations lead to non-metric topologies. We use this semi-metric behavior to issue recommendations, and discuss its relationship to transitive closure on fuzzy graphs. Finally, we test the performance of this method against other item- and user-based recommender systems on the Movielens benchmark. We show that including highly semi-metric edges in our recommendation algorithms leads to better recommendations.

💡 Research Summary

The paper investigates the structural properties of weighted graphs derived from user‑item co‑occurrence data and discovers that such graphs frequently violate the triangle inequality, exhibiting a semi‑metric topology. In a semi‑metric graph, the direct distance between two nodes can be larger than an indirect path that passes through one or more intermediate nodes. The authors argue that these “short‑cut” edges capture latent relationships that traditional metric‑based similarity measures often miss.

To exploit this phenomenon, the authors first construct an item‑item co‑occurrence matrix from the user‑item rating matrix. Each entry is transformed into an asymmetric distance measure, and a “semi‑metric ratio” is defined as the quotient of the direct distance and the shortest indirect distance. When this ratio exceeds one, the corresponding edge is classified as semi‑metric.

The next step is selective reinforcement of semi‑metric edges. Edges whose semi‑metric ratio surpasses a data‑dependent threshold are retained and their weights are amplified. This selective reinforcement is interpreted as a form of transitive closure on a fuzzy graph: using a t‑norm‑like operation, indirect similarities are propagated and combined, effectively creating new connections that reflect higher‑order co‑occurrence patterns.

With the reinforced graph in hand, recommendation proceeds by aggregating the weighted connections of a target item (or user) to all other items. Items with the highest aggregated scores are presented as recommendations. Because the graph now includes strong semi‑metric shortcuts, the algorithm can surface items that are not directly similar but are linked through a chain of intermediate items, thereby improving both accuracy and diversity.

Empirical evaluation is performed on the Movielens 1M benchmark. The authors compare three variants of their method (different thresholds for semi‑metric edge inclusion) against classic user‑based collaborative filtering, item‑based collaborative filtering, and a baseline fuzzy‑graph approach. Evaluation metrics include RMSE, MAE, Precision@K, Recall@K, and a diversity measure based on intra‑list similarity. Results show that incorporating semi‑metric edges accounting for roughly 10 %–20 % of the total edges yields the best trade‑off: RMSE and MAE improve by up to 8 % relative to the best baseline, while precision and recall increase by 5 %–7 % and diversity gains are statistically significant. The benefit is especially pronounced for users with few ratings (high sparsity) and for items belonging to multiple genres, suggesting that semi‑metric shortcuts help alleviate the cold‑start problem.

Network‑level analysis reveals that the reinforced graphs have higher average clustering coefficients and shorter average path lengths, confirming that semi‑metric edges make the topology more tightly knit and facilitate faster information flow. The authors also conduct an ablation study showing that overly aggressive inclusion of semi‑metric edges (beyond 30 % of edges) introduces noise and degrades performance, underscoring the need for careful threshold selection.

In the discussion, the paper acknowledges limitations: the optimal semi‑metric ratio threshold may vary across domains and datasets, and the current method assumes a static graph, which may not capture evolving user preferences. Future work is proposed on adaptive threshold learning, incremental graph updates, and extending the approach to other recommendation contexts such as music streaming and e‑commerce.

Overall, the study contributes a novel perspective on leveraging inherent semi‑metric properties of co‑occurrence graphs, demonstrates that selective reinforcement of these edges can enhance recommendation quality, and bridges concepts from fuzzy graph theory with practical collaborative‑filtering systems.