VaidhyaMegha/ai-research-kg

AI Research Open-Problems Knowledge Graph

29,707 nodes. 62,842 edges. A catalog of 1,000 open and hard research problems in AI and machine learning as a graph: each problem with its topic, the papers it cites and their authors and venues, the assumptions it rests on, the open questions it leaves, what blocks it, and the next experiment that would move it.

Built with Samyama Graph. Source catalog: samyama-ai/ai_research, commit a02e87b368016fd9568ba5ae8561ea601a8dee41 (source date 2026-09-10, retrieved 2026-10-07).

Licence: not yet chosen. The source repository states no licence. This card will be updated when one is chosen. Until then, do not assume any reuse rights.

What is in it

Every problem page in the source catalog follows a 10-section template. The graph keeps its structure:

• Problem (1,000) — title, statement, and status: open 401, empirically-open 295, partially-solved 189, methodologically-blocked 96, solved-but-impractical 19.
• Topic (35) — the catalog's taxonomy, from tokenization onward.
• Paper (4,932), Person (15,102), Venue (527) — the literature each problem cites, parsed from the pages' references.
• Assumption (3,263) — what a problem rests on, with whether it is violated.
• OpenQuestion (3,790) — classified as theoretically open (1,152), empirically open (1,509), methodologically blocked (1,088), or unclassified (41).
• Obstruction (6) — recurring reasons progress is blocked, e.g. confounded measurement.
• FutureDirection (1,000) — the concrete next experiment, one per problem, with its scale.
• Institution (51) — groups active on a problem.

Provenance — read this

• Problem pages are synthesized, not verified. The source template marks each page synthesized | verified; all 1,000 here are synthesized (catalog_provenance column). Treat statements, assumptions, open questions and proposed experiments as a structured draft, not as reviewed expert consensus.
• Citations are intended to be real. The source catalog's rules forbid fabricated identifiers and its audit checks titles against the arXiv API. Paper nodes carry arxiv_id or doi where present. They have not been independently re-verified for this release.
• Every node carries source_url pointing at the exact file and commit it was parsed from, and method saying how it was derived (parsed, name-matched, keyword-classified).
• Person nodes are author names as printed in citations, not resolved identities. Two people with the same name may share a node, and one person cited two ways may have two.

Why a graph

// Which papers underpin the most open problems?
MATCH (p:Problem)-[:CITES]->(paper:Paper)
WHERE p.status IN ['open', 'empirically-open']
RETURN paper.title, count(p) AS problems ORDER BY problems DESC LIMIT 10

// Problems that are empirically open — the experiment is runnable — and what it would cost
MATCH (p:Problem)-[:HAS_GAP]->(q:OpenQuestion {kind: 'empirically_open'}),
      (p)-[:HAS_NEXT_EXPERIMENT]->(f:FutureDirection)
RETURN p.title, f.scale LIMIT 20

// Shared hidden assumptions across topics
MATCH (t1:Topic)<-[:IN_TOPIC]-(:Problem)-[:RESTS_ON]->(a:Assumption)<-[:RESTS_ON]-(:Problem)-[:IN_TOPIC]->(t2:Topic)
WHERE t1 <> t2
RETURN a.name, collect(DISTINCT t1.name) AS topics LIMIT 20

Files

Nodes carry an id; edges reference those ids as src and tgt. A node's own key from the source (e.g. person:rico_sennrich) is in the key column. List-valued properties are JSON strings.

nodes/

File	Rows
person.csv	15,102
paper.csv	4,932
openquestion.csv	3,790
assumption.csv	3,263
problem.csv	1,000
futuredirection.csv	1,000
venue.csv	527
institution.csv	51
topic.csv	35
obstruction.csv	6
datasource.csv	1

edges/

File	Rows	Connects
authored_by.csv	28,757	Paper → Person
cites.csv	12,615	Problem → Paper
appeared_at.csv	4,365	Paper → Venue
has_gap.csv	3,790	Problem → OpenQuestion
rests_on.csv	3,263	Problem → Assumption
related_to.csv	3,128	Problem → Problem
active_on.csv	2,476	Institution → Problem
blocked_by.csv	1,448	Problem → Obstruction
has_next_experiment.csv	1,000	Problem → FutureDirection
from_source.csv	1,000	Problem → DataSource
in_topic.csv	1,000	Problem → Topic

Usage

from datasets import load_dataset
problems = load_dataset("VaidhyaMegha/ai-research-kg", "problem")

docker run --rm -p 8080:8080 -p 6379:6379 public.ecr.aws/f9f6l5u4/samyama-graph:1.1.0
curl -X POST localhost:8080/api/snapshot/import -F "file=@ai-research.sgsnap"

Verification

• Totals reconcile against the snapshot header: 29,707 nodes, 62,842 edges, 11 labels, 11 edge types.
• 0 dangling edges. 0 orphan nodes.
• CSVs were exported from the same snapshot that is published here.

Limitations

• Problem content is synthesized and unreviewed (above).
• Edge creation timestamps are not in the snapshot; node timestamps are.
• Author names are not disambiguated.
• A point-in-time export of one commit; the source catalog keeps changing.

Citation

AI Research Open-Problems Knowledge Graph, v1.0 (29,707 nodes, 62,842 edges).
Built with Samyama Graph. https://huggingface.co/datasets/VaidhyaMegha/ai-research-kg
Source: https://github.com/samyama-ai/ai-research-kg.git