English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

CoreKG: Coreset-Guided Personalized Summarization of Knowledge Graphs for Small Devices

Forum topic · QianXun · 2026-05-18

Summary

A 2026 arXiv paper by researchers including teams from the Indian Institutes of Technology introduces CoreKG (CoreKG: Coreset-Guided Personalized Summarization of Knowledge Graphs), a method for compressing massive knowledge graphs like Wikidata into tiny, personalized summaries that run on resource-constrained devices. The system analyzes a user's historical query workload to build an 'interest profile,' then applies sensitivity sampling from coreset theory to score every triple by its usefulness—keeping facts that answer likely questions and discarding irrelevant ones. Unlike blind random sampling, CoreKG comes with provable mathematical error bounds: even when retaining a tiny fraction of the original graph, answers to personalized queries match those from the full graph with high probability. Reported experiments show compression from tens of gigabytes to just a few megabytes, near-identical accuracy, and in some cases over 100x query speedups because distracting data is removed. The work matters for on-device AI assistants, small robots, and personalized search, suggesting that effective intelligence lies in concentrating limited resources on the most valuable information rather than storing everything. The original Chinese forum post explains the technique with a Feynman-style analogy of 'folding an encyclopedia into a business card.'

This post introduces CoreKG: Coreset-Guided Personalized Summarization of Knowledge Graphs, a paper published on arXiv in May 2026 by a research team including institutions such as the Indian Institutes of Technology. The paper tackles the conflict between ever-larger knowledge graphs and small, resource-limited devices.

The Core Idea

The author opens with an analogy: if you could take only one book to a desert island containing everything you'd ever want to know, a typical AI would hand you the entire Encyclopedia Britannica—too heavy to carry (memory overload) and mostly irrelevant to you. The smart approach is an "extreme folding": based on your reading preferences, compress millions of pages down to the key knowledge that matters most to you, with every surviving line still mapping precisely back to the original. In AI, this "extreme folding" technique is called Coreset theory.

The insight: modern knowledge graphs (like the relation networks behind Wikipedia) contain hundreds of millions of triples (e.g., person–birthplace–city)—far too large for a personal phone or small robot. CoreKG's logic: rather than memorizing the whole encyclopedia, memorize only the pages most useful to you.

How the "Folding" Works

1. Build your interest radar (User Workload) — The system inspects your past queries. If you've been researching quantum mechanics and Heisenberg, it treats your sensitivity to physics as high and your sensitivity to celebrity gossip as near zero.

2. Score every fact (Sensitivity Scoring) — Instead of blind random sampling, each piece of knowledge is weighted:

  • "Heisenberg proposed the uncertainty principle": can answer ~10 of your potential questions — very high weight.
  • "Heisenberg likes steak": true but useless for your current tasks — discarded.
  • 3. Provable guarantees — Using coreset theory, the researchers prove a remarkable property: even keeping only ~0.01% (one ten-thousandth) of the graph, you get answers identical to querying the full graph with high probability, under mathematically provable error bounds.

    Experimental Results

  • Compression: graphs that would require tens of GB shrink to just a few megabytes.
  • Accuracy: on personalized queries, the compressed "business-card graph" is nearly as accurate as the full graph.
  • Speed: because distracting information is removed, some queries run over 100x faster.

Why It Matters

The post argues the paper reframes conventional wisdom that "bigger is better" for AI knowledge. The essence of intelligence is concentrating limited resources on the most valuable information: a truly personalized AI shouldn't move the whole world to you—it should build the shortest path to the truth. Future AI assistants could carry a mathematically reliable, pocket-sized "private logic graph" instead of a bloated bookshelf.

> Truth lies not in quantity, but in precision.

*Note: This is a translation/summary of a Chinese forum post; claims and experimental figures are as reported in the post about the paper.*

Tags

#knowledge-graphs#coreset#personalization#knowledge-compression#arxiv-paper#on-device-ai#sensitivity-sampling#database-systems

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/177620228