Skip to main content

Citation Network: DBLP / AMiner V12

The DBLP / AMiner V12 snapshot contains 4,894,081 papers and 45,564,149 citation edges, dated April 2020.

An edge src → dst means paper src cites paper dst.

Do citation communities match the catalog's research fields?​

SQL selects papers from 2010 onward with one of ten AI primary_fos labels, keeps edges within that set and runs Louvain. It then counts the catalog labels within each returned community.

CREATE OR REPLACE VIEW ai_nodes AS
SELECT paper_id FROM papers
WHERE year >= 2010 AND primary_fos IN (
'Deep learning', 'Artificial neural network', 'Convolutional neural network',
'Recurrent neural network', 'Natural language processing',
'Reinforcement learning', 'Image segmentation', 'Feature extraction',
'Object detection', 'Speech recognition');

CREATE OR REPLACE VIEW ai_edges AS
SELECT e.src, e.dst
FROM citation_edges e
JOIN ai_nodes a ON a.paper_id = e.src
JOIN ai_nodes b ON b.paper_id = e.dst;

WITH community_fos AS (
SELECT c.community, p.primary_fos, COUNT(*) AS n
FROM cugraph_louvain(edges => (SELECT src, dst FROM ai_edges)) c
JOIN papers p ON p.paper_id = c.vertex
GROUP BY c.community, p.primary_fos),
ranked AS (
SELECT SUM(n) OVER (PARTITION BY community) AS members,
ROW_NUMBER() OVER (PARTITION BY community ORDER BY n DESC) AS rn,
primary_fos,
n
FROM community_fos)
SELECT members, rn, primary_fos, n
FROM ranked
WHERE members > 2500 AND rn <= 3
ORDER BY members DESC, rn;

These are the five captured groups larger than 2,500 papers. The listed fields cover only part of each community.

Papers in groupThree most common catalog fields, with paper counts
7,172Convolutional neural network (2,698); Deep learning (2,527); Artificial neural network (869)
4,477Deep learning (1,332); Recurrent neural network (1,280); Artificial neural network (624)
4,260Image segmentation (1,919); Convolutional neural network (921); Deep learning (799)
3,859Reinforcement learning (3,129); Artificial neural network (279); Deep learning (226)
3,518Object detection (2,193); Convolutional neural network (467); Deep learning (342)

The query took 0.17 s.

The 7,172-paper community combines convolutional neural network and deep learning labels in similar numbers. The 3,859-paper community is dominated by reinforcement learning. Citation structure and catalog fields agree to different degrees across these groups.

An editor can examine disagreements before updating taxonomy categories, treating community IDs as local cluster assignments.

The capture records the query and all returned field counts.

How are the papers around the Transformer connected?​

The selected graph contains Attention is all you need, its 22 references and its 110 most-cited citers in the April 2020 snapshot. The layout shows the links among those papers with their titles attached.

CREATE OR REPLACE VIEW attention_seed AS
SELECT paper_id FROM papers
WHERE title = 'Attention is all you need' AND year = 2017;

-- Include the seed, its references and its 110 most-cited citers.
-- Rank citers by the catalog's n_citation count.
CREATE OR REPLACE VIEW attention_ego_nodes AS
SELECT paper_id FROM (
SELECT e.dst AS paper_id
FROM citation_edges e JOIN attention_seed s ON s.paper_id = e.src
UNION ALL
SELECT src AS paper_id FROM (
SELECT e.src, p.n_citation
FROM citation_edges_by_dst e
JOIN attention_seed s ON s.paper_id = e.dst
JOIN papers p ON p.paper_id = e.src
ORDER BY p.n_citation DESC, e.src LIMIT 110) t
UNION ALL
SELECT paper_id FROM attention_seed
) u GROUP BY paper_id;

-- Keep only citations with both endpoints in the selected neighborhood.
CREATE OR REPLACE VIEW attention_ego_edges AS
SELECT e.src, e.dst
FROM citation_edges e
JOIN attention_ego_nodes a ON a.paper_id = e.src
JOIN attention_ego_nodes b ON b.paper_id = e.dst;

-- Use the same subgraph for positions (ForceAtlas2) and group IDs (Louvain).
WITH layout AS (
SELECT vertex, x, y
FROM cugraph_force_atlas2(
edges => (SELECT src, dst FROM attention_ego_edges), max_iter => 500, seed => 42)),
community AS (
SELECT vertex, community
FROM cugraph_louvain(edges => (SELECT src, dst FROM attention_ego_edges)))
SELECT l.vertex, l.x, l.y, c.community, p.title, p.year
FROM layout l
JOIN community c ON c.vertex = l.vertex
JOIN papers p ON p.paper_id = l.vertex;

The query took 0.93 s.

The saved illustration contains 133 papers and 393 citation links. Of those links, 261 connect papers other than the Transformer seed. The Sigma view supports hover, selection, dragging and zoom. Its browser layout also moves nodes, so displayed positions are not a fixed plot of a single SQL result.

Every selected paper links to the Transformer by construction. The other links, including those involving LSTM and Adam, point to related references inside the neighborhood. The saved group names are display annotations.

A fixed seed controls initial placement, but parallel force accumulation can produce different coordinates across runs. Save coordinates when later comparisons need the same positions. The ForceAtlas2 reference includes an example that reuses a layout as the initial positions for another pass.

Which papers share the Transformer's references?​

The seed is Attention is all you need, and related reading is ranked by shared references. Candidates are papers from 2017 to 2020 with at least ten references and one of four catalog labels: natural language processing, language model, machine translation or question answering. Papers with the seed's title are excluded.

Jaccard divides shared references by the union of the reference lists. The query gives cited papers negative reference-role IDs and citing papers positive IDs, allowing an undirected Jaccard call to compare outgoing lists. SQL filters the candidates and joins the returned scores to paper metadata.

CREATE OR REPLACE VIEW similarity_seed AS
SELECT paper_id, title, n_references
FROM papers
WHERE title = 'Attention is all you need' AND year = 2017;

CREATE OR REPLACE VIEW similarity_candidates AS
SELECT p.paper_id, p.title, p.n_references, p.year
FROM papers p
WHERE p.year BETWEEN 2017 AND 2020
AND p.n_references >= 10
AND p.primary_fos IN (
'Natural language processing', 'Language model',
'Machine translation', 'Question answering');

CREATE OR REPLACE VIEW bibliography_edges AS
SELECT e.src, -e.dst AS dst
FROM citation_edges e
JOIN (
SELECT paper_id FROM similarity_seed
UNION
SELECT paper_id FROM similarity_candidates
) selected ON selected.paper_id = e.src;

SELECT p.year, p.title, p.n_references,
ROUND(r.similarity, 3) AS jaccard
FROM cugraph_jaccard(
edges => (SELECT src, dst FROM bibliography_edges),
vertex_pairs => (
SELECT s.paper_id AS first, c.paper_id AS second
FROM similarity_seed s
CROSS JOIN similarity_candidates c
WHERE LOWER(c.title) <> LOWER(s.title)),
directed => false) r
JOIN papers p ON p.paper_id = r.second
ORDER BY r.similarity DESC, p.paper_id
LIMIT 5;

The first three of the saved result rows are:

PaperYearJaccard
Self-Attention with Relative Position Representations20180.296
Future-Prediction-Based Model for Neural Machine Translation.20180.273
Weighted Transformer Network for Machine Translation20180.262

The query took 0.24 s.

Follow-up: inspect the shared references. The leading pair has 22 and 13 references. This SQL returns their eight shared records, including Adam, Sequence to Sequence Learning with Neural Networks, and Neural Machine Translation by Jointly Learning to Align and Translate.

WITH seed AS (
SELECT paper_id FROM papers
WHERE title = 'Attention is all you need' AND year = 2017
), candidate AS (
SELECT paper_id FROM papers
WHERE title = 'Self-Attention with Relative Position Representations' AND year = 2018
)
SELECT p.paper_id, p.year, p.title
FROM citation_edges a
JOIN seed s ON s.paper_id = a.src
JOIN citation_edges b ON b.dst = a.dst
JOIN candidate c ON c.paper_id = b.src
JOIN papers p ON p.paper_id = a.dst
ORDER BY p.year, p.paper_id;
  • Adam: A Method for Stochastic Optimization (2014)
  • Neural Machine Translation by Jointly Learning to Align and Translate (2014)
  • Sequence to Sequence Learning with Neural Networks (2014)
  • Effective Approaches to Attention-based Neural Machine Translation (2015)
  • End-To-End Memory Networks (2015)
  • A Decomposable Attention Model for Natural Language Inference (2016)
  • Google's Neural Machine Translation System: Bridging the Gap between Human and Machine Translation (2016)
  • Neural Machine Translation in Linear Time (2016)

The query took 0.33 s.

The score is 8 / (22 + 13 − 8) = 0.296 after rounding. The shared works explain what produced the ranking and form a bibliography to inspect.

The capture records a native_preferred run with one host execution boundary for the similarity query. It demonstrates composition with SQL but not a fully native execution path. The shared-reference result was independently checked.

Citation paths​

How does BERT connect to the original LSTM paper?​

Citation edges run from the citing paper to its reference. The query selects BERT and LSTM by title and year, follows those edges with BFS, and joins the path to titles.

SELECT b.path_index, b.distance, p.year, p.title
FROM cugraph_bfs(
edges => (SELECT src, dst FROM citation_edges),
sources => (SELECT paper_id AS vertex FROM papers
WHERE title = 'BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding'
AND year = 2018),
output_mode => 'path',
target_vertices => (SELECT paper_id AS vertex FROM papers
WHERE title = 'Long short-term memory' AND year = 1997)) b
JOIN papers p ON p.paper_id = b.vertex
ORDER BY b.path_index;
  1. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding (2018)
  2. Character-Level Language Modeling with Deeper Self-Attention (2018)
  3. Long short-term memory (1997)

The query took 0.19 s.

Follow-up: compare the other shortest routes. A relational query checks which papers BERT cites that also cite LSTM.

SELECT p.paper_id, p.title, p.year
FROM citation_edges first_step
JOIN papers source ON source.paper_id = first_step.src
JOIN citation_edges second_step ON second_step.src = first_step.dst
JOIN papers target ON target.paper_id = second_step.dst
JOIN papers p ON p.paper_id = first_step.dst
WHERE source.title = 'BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding'
AND source.year = 2018
AND target.title = 'Long short-term memory' AND target.year = 1997
ORDER BY p.year, p.paper_id;

The query took 0.52 s.

It returns ten intermediate papers, including Attention is all you need. The BERT → Transformer → LSTM route is therefore another reading option.

The capture includes all ten records and an independent check of these edges.

How many papers cite AlexNet within three steps?​

The query counts papers at each distance from AlexNet without counting one paper repeatedly. Reversing the edges follows papers that cite AlexNet and then papers that cite those papers.

SELECT b.distance, COUNT(*) AS papers, ROUND(AVG(p.year), 1) AS avg_year
FROM cugraph_bfs(
edges => (SELECT dst AS src, src AS dst FROM citation_edges_by_dst),
sources => (SELECT paper_id AS vertex FROM papers
WHERE title = 'ImageNet Classification with Deep Convolutional Neural Networks'
AND year = 2012),
depth_limit => 3, output_mode => 'vertices') b
JOIN papers p ON p.paper_id = b.vertex
WHERE b.reachable
GROUP BY b.distance
ORDER BY b.distance;
Shortest distance from AlexNetPapersMean publication year
1 step12,1852017.5
2 steps67,5172017.9
3 steps71,5412017.8

The query took 0.17 s.

The three layers contain 151,243 papers, excluding AlexNet itself. Each paper appears at its shortest distance.

The count describes reach within the April 2020 snapshot.

Data scope​

papers includes title, year, reported citation count, reference count, and primary_fos, the source's primary field-of-study label.

TableRowsContents
citation_edges45,564,149src, dst, weight = 1.0; sorted by src
citation_edges_by_dst45,564,149The same citations sorted by dst for reverse traversal
papers4,894,081paper_id, title, year, venue, n_citation, n_references, primary_fos, …
paper_authors14.9MPaper-author links
paper_fos45.0MPaper-field-of-study links

papers.n_citation is AMiner's reported global count. In-graph degree counts only edges present in this snapshot. The in-degree example compares them.

Seed paper records. Queries use paper_id as the graph vertex. Titles can repeat across records. The examples select title and year together when identifying a seed.

Paperpaper_idExact title in the corpusyearExtra predicate
Transformer2963403868Attention is all you need2017
BERT2896457183BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding2018
AlexNet2163605009ImageNet Classification with Deep Convolutional Neural Networks2012
VGG1686810756VERY DEEP CONVOLUTIONAL NETWORKS FOR LARGE-SCALE IMAGE RECOGNITION2014
ResNet2194775991Deep Residual Learning for Image Recognition2016
LSTM2064675550Long short-term memory1997
PageRank2066636486The anatomy of a large-scale hypertextual Web search engine1998venue = 'The Web Conference'

Attention Is All You Need with a capitalized Is matches a separate arXiv record. Copy exact titles from this table or query papers for the intended record.

The investigation capture contains the SQL, returned rows, query IDs and source counts. A separate SQL check enumerates the ten BERT-to-LSTM intermediates and another returns the eight shared references. Independent relational checks matched both results. The bibliography query used mixed execution; the other captured queries ran natively. Execution details are recorded with each query.

The query timings include planning, execution and result receipt.