Identifying Leaders Through Graph-Based HR Analytics
Everyone's performance is rated as "similar," and everyone is working "hard."
But when asked, "Who is the real leader?", it’s not easy to answer. Traditional indicators like sales figures or supervisor reviews seem insufficient.
Leadership in an organization requires more than technical skills—it demands communication ability, conflict resolution, and the capacity to influence others.
This post introduces a graph-based HR study that identifies potential managerial candidates by analyzing behavioral networks among employees[1].
Compared to conventional attribute-based models,
the graph-based approach demonstrated 5–10% higher accuracy in identifying managerial potential.
Background
The Limitations of Traditional HR Evaluation
Conventional performance evaluations are often subjective and hard to account for communication patterns or contextual behavior. There has been limited use of quantitative analysis of social network structures—in other words, how employees interact and exchange information within the organization. Traditional models primarily rely on attributes such as department, position, or rank, and neglect the form and significance of communication and collaboration.
Core Idea
Identifying Managers Based on Communication Centrality
The study builds an interaction graph using data from emails, meetings, and collaboration logs. Each node represents an employee, and the intensity and direction of interactions are mapped as edges. For instance, the number of emails sent, meeting invitations, and recipient counts are aggregated to calculate the centrality of each employee in collaboration. These metrics are used as relationship-based features in the machine learning model to classify “managerial potential.”
Solution
Step 1. Constructing the Organizational Collaboration Network
The study is based on a real-world dataset from a mid-sized IT company in Europe (~200 employees).
Node (Employee): Includes attributes such as employee ID, language, department, division, and rank.
Edge (Interaction): Includes five types of relationships. And each relationship is directional and weighted based on frequency:
(a) Email CC
(b) Email recived
(c) Sent Email with Attachment
(d) Received Email with Attachment
(e) Arranged Meeting
<Five types of collaboration networks>
Step 2. Calculating Network Centrality
To evaluate the importance of each employee within the collaboration network, PageRank centrality is calculated. PageRank is based on the idea that "my centrality is influenced by the centrality of my neighbors." It is similar to eigenvector centrality but more suitable for directional networks, such as email flows. Nodes connected to high-ranking nodes receive higher scores themselves. These PageRank scores are added as features for each employee.
<Example Training Dataset>
Step 3. Identifying Potential Managers
Using the training data, various classification models were built to predict managerial potential:
The Graph Convolutional Network (GCN) is a type of Graph Neural Network (GNN) that learns not only the node’s own features, but also aggregates information from connected neighbors
—i.e., it considers relational context.
<GCN Cencept>
Result
HR Analytics Enhanced by Organizational Network Structure
The Graph based model achieved 5–10% higher accuracy than traditional machine learning models in identifying managers. In particular, combining PageRank scores with word2vec embeddings of department and job titles significantly improved prediction performance, with GCN reaching a maximum accuracy of 80.02%. The study demonstrates that incorporating organizational network structure allows for more precise talent analytics compared to using attribute data alone.
<Model accuracy comparison – with and without graph features>
<GCN accuracy by hidden layer size>
As workplace communication grows more complex, traditional evaluations and performance-based filtering alone are no longer sufficient to identify future leaders. By modeling employee interactions (emails, meetings) as a network and applying graph analytics (PageRank, GCN), organizations can quantitatively identify central influencers and leadership potential. This approach offers HR professionals a data-driven alternative to subjective evaluation, helping to strategically shape leadership pipelines and succession planning.
netminer@cyram.com
References
[1] Chan, J. Y. K., Wang, Z., Xie, Y., Meisel, C. A., Meisel, J. D., Solano, P., & Murillo, H. (2021).
Identifying Potential Managerial Personnel Using PageRank and Social Network Analysis: The Case Study of a European IT Company. Applied Sciences, 11(15), 6985. https://doi.org/10.3390/app11156985