BlogCatalog (single-membership subset)
Task: Blogger interest-group classification
Size band: medium · Label type: semantic
Label column: group · Converter: converters/convert_blogcatalog.py
Blogger friendship network whose original labels are multi-label (39 overlapping interest groups; a blogger can belong to up to 11). To fit the single-label contract, the converted graph keeps only bloggers with exactly one group membership and the induced subgraph on them.
Labeling methodology (decision approved 2026-07-23)
The source group matrix is multi-label: of 10,312 bloggers, 7,460 (72.3%) belong to exactly one group, 2,852 to two or more. Three reductions were considered — (A) keep single-membership bloggers only, (B) assign multi-member bloggers a primary group, (C) one-vs-rest on a single group. Option A was chosen: it is the only reduction that invents no assignment rule and keeps labels exactly as authored.
- Kept: 7,460 single-membership bloggers; induced subgraph has 131,034 edges.
- Dropped: 2,852 multi-membership bloggers and all their edges (recorded in
metadata.jsonunderconversion.labeling_rule). - 38 of the 39 groups survive (one group has no single-membership member).
- Original blogger matrix indices preserved in
neext/id_mapping.csv.
Consequence: results are NOT directly comparable to multi-label BlogCatalog numbers in the embedding literature (deepwalk/node2vec macro-F1), which score all 10,312 bloggers with one-vs-rest classifiers. Implemented in converters/convert_blogcatalog.py.
Converted graphs (neext/)
| graph | nodes | edges | classes | feature cols | isolated | class counts |
|---|---|---|---|---|---|---|
| default | 7,460 | 131,034 | 38 | 0 | 170 | 7: 970, 4: 597, 23: 514, 5: 504, 1: 481, 18: 432, … |
Conversion notes: Multi-label source (39 overlapping groups). Kept only bloggers with exactly one group membership (7,460 of 10,312; 72.3%) and the induced subgraph on them; multi-member bloggers and their edges are excluded rather than force-assigned a primary group. Original blogger indices in id_mapping.csv.
Source
- blogcatalog.mat — 1,255,783 bytes, sha256
d4f4fb89ce1ccd4b…, fetched 2026-07-23
License: deepwalk repo GPL-3.0; data from ASU social computing repository
Citation: Tang, Liu. Relational Learning via Latent Social Dimensions. KDD 2009.
Caveats
- Subset graph — 28% of bloggers (the multi-membership ones) are excluded, so published multi-label baselines don't apply.
- 38 imbalanced classes; use macro-averaged metrics.
Generated by converters/make_cards.py; stats from metadata.json. Raw files: source/. NEExT tables: neext/.