iSummary: Workload-based Selective Summaries for Knowledge Graph Exploration

Submitted by Haridimos Kondylakis on 04/26/2026 - 01:38

Tracking #: 4072-5286

Authors: 
Giannis Vassiliou
Nikolaos Papadakis
Haridimos Kondylakis

Responsible editor: 
Katja Hose

Submission type: 
Full Paper
Abstract: 
The rapid growth in size and complexity of knowledge graphs (KGs) available on the web has created a pressing need for efficient and adaptive methods to facilitate their understanding and exploration. Recently, semantic summaries have emerged as a means to quickly comprehend and explore large KGs. However, most existing approaches are static, failing to adapt to user needs and often struggling to scale. In this paper, we introduce iSummary, a workload-based and scalable approach for constructing selective summaries tailored to specific user requests. Unlike prior methods that process the entire KG, iSummary leverages query logs, exploiting the collective knowledge embedded in past user queries to identify relevant resources and relationships. The summarization process operates linearly with respect to the number of queries, enabling incremental and scalable summary generation even for large workloads. We formally define the Selective Summary problem, provide an approximate and efficient algorithm with theoretical guarantees, and evaluate it on two real-world datasets. Experimental results demonstrate that iSummary consistently outperforms existing techniques in both coverage and efficiency, producing high-coverage summaries up to 40× faster than state-of-the-art approaches.
Full PDF Version: 
PDF icon swj4072.pdf
Previous Version: 
iSummary: Workload-based Selective Summaries for Knowledge Graph Exploration
Tags: 
Reviewed
Long-term Stable Link to Resources: 
https://github.com/giannisvassiliou/iSummary

Decision/Status: 
Accept

Solicited Reviews:
Click to Expand/Collapse
Review #1
Anonymous submitted on 25/May/2026
Suggestion:
Accept
Review Comment:

RDF modeling. Section 3's algorithmic graph GQ as a predicate-labeled multigraph with rdf:type treated as an edge label resolves the abstraction mismatch. The new limitations paragraph on rdf:type, including the schema/instance presentation-layer proposal, is adequate.

SPARQL handling. The property-path clarification in Section 2.1 is fine. Section 4.1 reports preprocessing survival rates (≈53% for DBpedia, ≈47% for WikiData), as requested.

Variable resolution. Section 3.3's three alternative rules (LM1, DET, TopK) and the evaluation in Section 4.5.6 address the endpoint ordering concern. The acknowledgement that the below-10% figure is snapshot-specific is appropriate.

Evaluation. PPR-QI is a substantive addition. The decomposition into scope and algorithmic components, with iSummary outperforming PPR-QI in every configuration, isolates the contribution of workload-derived frequencies. The GLIMPSE input is now specified. The seed-held-out split was not run, but the limitation is flagged in Section 4.3 and the coverage results reframed as generalization to unseen queries about already-seen targets, which is acceptable given the targeted use case.

Review #2
Anonymous submitted on 21/Jun/2026
Suggestion:
Accept
Review Comment:

I would like to thank the authors for their effort and additions. I believe that this paper is ready to be published.

Review #3
By Blerina Spahiu submitted on 03/Aug/2026
Suggestion:
Accept
Review Comment:

The authors have carefully addressed my comments. In particular, they clarified the theoretical formulation, improved the fairness of the experimental comparison by introducing a new query induced baseline, explicitly discussed the limitations of coverage as an evaluation metric, and repositioned the qualitative demo as preliminary evidence rather than a user study. The manuscript is significantly clearer and methodologically stronger.