Applying K-Means Clustering to Expertise Locator Knowledge Management Systems

Got it. Here’s a research-paper–style blogpost article that applies k-means clustering to expertise-locator knowledge management systems (KMS) using domain-specific web data, framed as a scholarly but accessible blogpost. I’ll format it with sections (Abstract, Introduction, Methodology, Discussion, Conclusion, References), APA-style citations, and plain-language flow so it reads like a hybrid between an academic paper and a thought-leadership article.

Applying K-Means Clustering to Expertise Locator Knowledge Management Systems

Abstract

Expertise locator knowledge management systems (KMS) help organizations identify and connect with internal experts. Early approaches, such as NASA’s Expert Seeker, relied on mining organizational documents and associating employees with keywords found in their work. This article explores how k-means clustering can extend such systems by grouping expert profiles into coherent domains. Drawing on the foundations of Expert Seeker (Becerra-Fernandez & Rodriguez, 2001), and integrating workflow-oriented architectures such as aicuban and wh-uzi, we show how domain-specific web data can be used to build expert vectors, reduce dimensionality, cluster profiles, and feed results into routing workflows and user interfaces.

Introduction

Knowledge-intensive organizations frequently struggle to answer the question: who knows what? Expertise locator systems were developed to address this challenge by mining available information sources and indexing employees by their expertise areas. NASA’s Expert Seeker exemplified this approach by building profiles from intranet documents and associating individuals with keywords, based on co-occurrence analysis (Becerra-Fernandez & Rodriguez, 2001).

While traditional systems emphasized keyword matching, modern knowledge environments face exponentially larger and more diverse data sources, often spanning structured HR databases and unstructured domain-specific web data. Clustering methods such as k-means offer an opportunity to automatically group expert profiles into coherent domains, surfacing communities of practice and supporting richer discovery.

Methodology

Feature Construction

Each expert profile is represented as a vector of features extracted from associated documents, publications, or domain-specific web data. Following the Expert Seeker model, profiles can be constructed by counting co-occurrences of employee names with technical keywords (Becerra-Fernandez & Rodriguez, 2001). Modern pipelines augment this with TF–IDF weighting, embeddings (e.g., SciBERT for scientific domains), and structured attributes such as project counts or departmental roles.

Dimensionality Reduction

High-dimensional text vectors (thousands of keyword features) create sparsity that can undermine clustering (Liu et al., 2013). Dimensionality reduction methods such as Principal Component Analysis (PCA) or Truncated SVD compress features while preserving the most significant variance. Embedding models inherently perform a similar compression (Reimers & Gurevych, 2019).

K-Means Clustering

The reduced feature vectors are clustered using k-means, which partitions experts into k groups based on similarity. The choice of k is determined using validation techniques such as the silhouette score, elbow method, or domain knowledge about the organization’s taxonomy (DataRobot, 2025).

Labeling and Validation

Clusters are labeled by examining top-weighted terms near cluster centroids and by reviewing representative expert profiles. Human validation remains essential: subject matter experts can confirm whether cluster assignments reflect meaningful domains of expertise.

Workflow and UI Integration

The integration of clustering output is twofold:

1. Workflow engines – Systems like wh-uzi can incorporate cluster IDs as routing rules, directing requests to the most relevant group of experts. The aicuban architecture can embed cluster assignments as decision points in its workflows, ensuring that domain-specific expertise clusters are actionable.
2. User interfaces – Clusters form a navigable taxonomy for browsing experts. Profiles can display cluster tags, and search results can be grouped or filtered by cluster, enabling more intuitive discovery of expertise domains.

Discussion

Clustering enhances expertise locator systems in several ways. First, it supports taxonomy generation: clusters can suggest emerging areas of expertise not captured by static ontologies. Second, it improves routing and decision automation: workflows can leverage cluster assignments to escalate queries to the right group. Third, it enables community discovery, surfacing hidden expert networks within organizations.

Challenges remain. Clustering results are sensitive to the quality of features; poor preprocessing may mix unrelated skills. Validation is resource-intensive and requires human oversight. Clusters may drift over time as experts update their profiles, necessitating periodic re-clustering and review.

Nonetheless, the integration of clustering into expertise locator KMS—grounded in early co-occurrence methods but extended with modern data mining and workflow engines—offers a scalable path to knowledge discovery.

Conclusion

By applying k-means clustering to domain-specific web data and expert profiles, organizations can move beyond keyword-based expertise search to cluster-driven knowledge discovery. Starting from the co-occurrence-based Expert Seeker approach, enriched with embeddings and dimensionality reduction, clusters reveal meaningful domains of expertise. When integrated into workflow engines (aicuban, wh-uzi) and user interfaces, these clusters enhance routing, discovery, and organizational learning. As knowledge management systems evolve, clustering provides both the structure and flexibility needed to navigate complex, dynamic expertise landscapes.

References

Becerra-Fernandez, I., & Rodriguez, D. (2001). Knowledge discovery in knowledge management: Methods, tools, and applications. In Proceedings of the Fourteenth International FLAIRS Conference (pp. 485–489). AAAI Press. https://cdn.aaai.org/FLAIRS/2001/FLAIRS01-054.pdf

DataRobot. (2025). Clustering in unsupervised learning. Retrieved from https://docs.datarobot.com

Liu, L., Tang, L., Dong, W., Yao, S., & Zhou, W. (2013). An overview of topic modeling and its current applications in bioinformatics. SpringerPlus, 2(1), 160. https://doi.org/10.1186/2193-1801-2-160

Reimers, N., & Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing (pp. 3982–3992). ACL. https://doi.org/10.18653/v1/D19-1410

Would you like me to also generate a sample clustering visualization (e.g., t-SNE scatterplot of expert clusters with labels) to go along with this blogpost article, so it looks more engaging for readers?