From Nov 17th-19th, I got to present our paper “How To Surprisingly Consider Recommendations?” at KEOD 2024. The conference is over, the proceedings are out, so grab it from SCITEPRESS while it’s hot! (mirror)

Read on to learn more!

Why should I read it?

You’ve probably found yourself in a situation where some odd online recommender system has produced expectable recommendations. Think <music streaming service> - how many of the recommended items are actually items you don’t know? How many carry the potential of expanding your musical knowledge?

In many cases, the surprising items are few and far between, which is fine for some use cases, and detrimental to others. Consider the “browsing to explore”-case above: if the recos don’t allow you to leave your bubble, this will impact your satisfaction with the system. In other cases, diverse and heterogeneous recommendations of, e.g., research papers, can foster trans- and interdisciplinary research and understanding.

In a nutshell, we approach the task of determining surprising items through complex network metrics on knowledge graphs. Knowledge graphs exhibit certain properties we can measure - since they’re just graphs, after all - and that allow us to identify the surprising ones.

For fans of…

• recommender systems
• knowledge graphs
• network metrics and their application on knowledge graphs