How to Surprisingly Consider Recommendations? A Knowledge-Graph-based Approach Relying on Complex Network Metrics

TL;DR

The paper tackles the tendency of traditional recommender systems to overemphasize mainstream items by introducing a knowledge-graph–based layer that enables a user-controlled degree of surprise. It proposes a two-step retrieval-and-ranking framework where a base RS provides candidates and a KG-informed re-ranking evaluates each candidate's impact on network metrics computed within a user-profile subgraph, using $m$ and, when appropriate, the normalized $HHI^*$ to aggregate distributions. The authors identify betweenness centrality as a key metric correlating with novelty and diversity, and demonstrate through experiments on LastFM (LFM-1b) and synthetic Netflix data that reranking by network metrics can produce more unexpected and diverse recommendation lists, albeit with trade-offs in rank stability (nDCG). The work contributes a general, metric-driven reranking layer that can augment any RS, highlights practical alternatives to betweenness for scalability, and suggests future work on cross-domain KG extensions and the inclusion of textual item features. Overall, the study shows that graph-structure-aware reranking can meaningfully diversify recommendations beyond accuracy, supporting serendipitous discovery in large catalogs.

Abstract

Traditional recommendation proposals, including content-based and collaborative filtering, usually focus on similarity between items or users. Existing approaches lack ways of introducing unexpectedness into recommendations, prioritizing globally popular items over exposing users to unforeseen items. This investigation aims to design and evaluate a novel layer on top of recommender systems suited to incorporate relational information and suggest items with a user-defined degree of surprise. We propose a Knowledge Graph (KG) based recommender system by encoding user interactions on item catalogs. Our study explores whether network-level metrics on KGs can influence the degree of surprise in recommendations. We hypothesize that surprisingness correlates with certain network metrics, treating user profiles as subgraphs within a larger catalog KG. The achieved solution reranks recommendations based on their impact on structural graph metrics. Our research contributes to optimizing recommendations to reflect the metrics. We experimentally evaluate our approach on two datasets of LastFM listening histories and synthetic Netflix viewing profiles. We find that reranking items based on complex network metrics leads to a more unexpected and surprising composition of recommendation lists.

Figures (6)

• Figure 1: Overview of the proposed knowledge-graph informed recommendations. KGs are constructed for the item catalog and all user profiles. The latter serve as input to an arbitrary state-of-the-art RS, whose results are re-ranked according to the impact the items would have were they included in the original user profile.
• Figure 2: KG-Informed Recommendation. The sub-KG represents a user profile. A node representing a recommendation is obtained from the catalog KG and included in the sub-KG along with applicable edges. We then compute network metrics from the sub-KG.
• Figure 3: Measuring surprise on feature-level for recommendations re-ranked by metric. For Unexpectedness, the highlighted bars denote the comparison to the SOTA recommendations. For Diversity, the highlighted bars provide the intra-list values of SOTA and original user profiles (base and profile, resp.).
• Figure 4: nDCG@10 for LastFM. Panels show nDCG scores obtained on recommendation lists ranked by metric in ascending and descending order.
• Figure 5: nDCG@10 for Netflix. Panels show nDCG scores obtained on recommendation lists ranked by metric in ascending and descending order.