Implicit User Network Analysis of Communication Platform Open Data for Channel Recommendation

Abstract

Recommender systems play a pivotal role in various human-centered online systems by filtering out relevant information from large databases. However, most recommender systems consume explicit private user information such as exchanged messages and information between users and items such as likes and shares without exploring other latent factors. Past events have shown that this can have decremental consequences on users' privacy. One type of application where alternative solutions have not yet been investigated are messaging platforms in larger corporate environments. These applications would benefit from recommender systems that consume only anonymized implicit data to enable employees to discover new communities and people. As a first step in developing such a recommender system, this paper describes the construction and analysis of implicit social network data from the messaging platform Mattermost at CERN and the extraction of measures for indicating similarity between users and channels. Additionally, it describes the use of these measures to evaluate multiple existing collaborative filter-based recommender systems, where their performances are compared and evaluated against simple measures. The evaluation results indicate that combining clustering approaches and custom features extracted through our data analysis outperforms standard collaborative filtering techniques. These results will be used in the future to create a new custom recommender system for messaging at CERN that only uses anonymized and implicit data.