CollegeMsg temporal closure¶

Static clustering cannot tell whether a triangle preceded or followed its third edge. CollegeMsg supplies the missing order, but its edges record communication rather than friendship.

Source boundary¶

The SNAP CollegeMsg source contains directed private-message events from an online social network at the University of California, Irvine. Each row is source target unix_timestamp. The source page reports no redistribution license, so tinymesh does not commit the artifact. The public adapter accepts only the 345,339-byte gzip with SHA-256 50ae2d98ed3bad9ddb18dbd495a89e5e10cfb8f7e86932827db29fc41b41f9fa.

directed timestamped messages              source truth
              |
              v
groups with equal timestamps               no arbitrary within-time order
              |
              v
undirected first contact per pair           explicit measurement projection
              |
       +------+------+
       |             |
       v             v
prior open wedge   prior non-wedge          mutually exclusive risk sets
       |             |
       +------+------+
              v
       formations / pair-time exposure

The parser preserves direction, identity, duplicate messages, and timestamps. Only the closure measurement projects a directed message to an undirected observed contact. A pair enters the risk set after both users have appeared and leaves after its first contact. Contacts that introduce either user are reported separately because the source cannot establish their prior exposure.

At each distinct timestamp, the experiment first measures elapsed exposure and classifies every new contact against the graph strictly before that timestamp. It then adds the whole timestamp group. A non-edge is wedge-exposed when its endpoints have at least one prior common neighbor. The incremental wedge set costs O(degree(source) + degree(target)) per new contact and stores no dense adjacency or node-pair-by-time table.

Reference boundary¶

The pinned graph libraries answer adjacent but different questions:

PyG Temporal's TwitterTennisDatasetLoader returns popularity-filtered mention snapshots and a future-mention target. That selection changes the degree and closure population.
PyG's BitcoinOTC bins timestamped signed trust ratings and gives edges a fixed ten-window lifetime. Trust, sign, bin width, and edge lifetime are different contracts.
PyG's JODIEDataset preserves temporal interactions but offsets destination identities into a bipartite user-item graph, where ordinary user-user triadic closure does not apply.

Tinymesh therefore uses no framework container or snapshot policy here. TemporalEdges retains only aligned event truth, while college_msg() owns source validation and identity lowering. The executable evidence consumes only that public boundary, so the catalog records no reference gitlink.

Observation¶

Revision 3297990 produced this witness:

uv run --locked python -m experiments.run college_msg_closure

Source property	Observation
Messages / users	59,835 / 1,899
Directed message pairs	20,296
Undirected contacted pairs	13,838
Repeated messages after first contact	45,997
First contacts introducing a user	1,826

The remaining 12,012 first contacts have both endpoints in the prior observed population:

Prior relation	Formations	Pair-seconds exposed	Formations per million pair-days
Open wedge	5,609	4,501,578,548,508	107.655
Non-wedge	6,403	17,487,499,269,136	31.635

The wedge incidence rate is 3.403 times the non-wedge rate. This is a temporal association consistent with triadic closure. It is not a causal estimate: homophily, user activity, exposure outside the platform, left censoring, and the undirected projection can all explain part of the difference. A message is also not evidence of friendship.

Decision¶

Promote the reusable source truth and retain the interpretation as research:

src/tinymesh.TemporalEdges          ordered event identity + strict prefix
tinymesh.datasets.college_msg       checksum + source identity lowering
experiments.college_msg_closure     first-contact projection + measurement

CollegeMsg is the live public caller for aligned temporal edges; the closure experiment proves that the carrier retains every count from the earlier experiment-owned parser. No Graph snapshot, undirected projection, durable tie, feature tensor, or model follows automatically from an interaction event.

The dataset is now available for studying temporal closure, repeated interaction, embeddedness, and local bridges. The separate tie-structure record owns final-graph strength, overlap, and fragmentation policies. Any predictive stage must first freeze controls for activity, degree, recency, temporal splits, and negative sampling. MBTA's event-as-node mesh remains a different ontology and does not need to share this carrier.