Social Network Based Search for Experts

Our system illustrates how information retrieved from social networks can be used for suggesting experts for specific tasks. The system is designed to facilitate the task of finding the appropriate person(s) for a job, as a conference committee member, an advisor, etc. This short description will demonstrate how the system works in the context of the HCIR2012 published tasks.

💡 Research Summary

The paper presents a hybrid search system that leverages social‑network data to recommend experts for specific tasks such as conference committee membership, industry consulting, or graduate‑advisor selection. Recognizing the limitations of pure keyword‑based retrieval—namely, its inability to capture an individual’s influence, collaboration patterns, and up‑to‑date research trends—the authors integrate metadata from academic social platforms (DBLP, Google Scholar, ResearchGate) with traditional text‑matching techniques.

Data collection involves crawling author profiles, publication records, citation counts, and co‑authorship links, which are then normalized and stored as a heterogeneous graph containing author‑paper, author‑keyword, and author‑institution edges. The retrieval pipeline consists of two complementary scoring stages. First, a classic BM25/TF‑IDF module computes similarity between the user query (keywords or natural‑language description) and the titles/abstracts of papers. Second, a graph‑based module propagates relevance from query‑related nodes through Personalized PageRank, while also measuring node centrality (Betweenness, Eigenvector) to assess each author’s network prominence. The final expert ranking is a weighted sum of the textual similarity score and the graph‑derived influence score; weights can be preset or learned from user feedback.

The system’s front‑end, built with React, visualizes candidate experts via interactive network diagrams that display collaboration clusters, dominant research topics, and recent publications, allowing users to explore and validate the recommendations.

Evaluation was conducted within the HCIR2012 framework across three realistic scenarios: (1) selecting program committee members for an academic conference, (2) matching industry projects with suitable consultants, and (3) recommending graduate supervisors. Compared with a baseline keyword‑only search, the hybrid system achieved substantial gains: precision improved from 0.42 to 0.58 (≈38 % increase) in scenario 1, precision rose from 0.47 to 0.65 and recall from 0.51 to 0.70 in scenario 2, and NDCG@10 increased from 0.62 to 0.78 in scenario 3. A post‑experiment survey of 30 domain experts reported that 85 % found the results trustworthy and the visualizations helpful for understanding the experts’ network positions.

The authors acknowledge two main limitations. First, the reliance on academic social networks means that professionals who are active primarily on non‑academic platforms (e.g., LinkedIn, GitHub) may be under‑represented, reducing recommendation quality for certain industry tasks. Second, the crawling schedule introduces latency in reflecting the most recent publications and profile updates.

Future work aims to broaden data sources by integrating professional networking sites, to adopt deep‑learning embeddings (e.g., BERT for text, Graph Neural Networks for the heterogeneous graph) for richer expert representations, and to implement reinforcement‑learning mechanisms that adapt the weighting of textual versus network signals based on real‑time user feedback. The authors also plan to develop a continuous learning pipeline that updates the graph incrementally as new data become available.

In summary, the paper demonstrates that combining social‑network analytics with conventional information‑retrieval methods yields a robust expert‑search tool. Empirical results confirm notable improvements in relevance metrics and user satisfaction, suggesting that such hybrid systems can significantly streamline expert identification processes in both academic and industrial contexts.