Why relationship structure is worth testing
A known facility or business need does not automatically reveal which people matter, who knows whom, or how a particular seller can reach the right decision maker. A professional profile offers context. A relationship graph offers a different kind of evidence: how people connect across that context.
The working hypothesis is that richer, relevant relationship structure may improve the discovery of connection candidates beyond profile similarity alone. The desired outcome is a more credible next action for a user, not a larger list of names.
GraphSAGE is a published method introduced by William L. Hamilton, Rex Ying, and Jure Leskovec. It learns representations using node features and information aggregated from neighboring nodes.[1][2] ExecGraph applies that general approach to an Energy relationship-prediction task. GraphSAGE itself is not an ExecGraph invention.
The public contribution is a dated applied observation: how a fixed-configuration challenger and a profile-based comparator performed as the recorded graph changed. It is neither a new algorithm claim nor a controlled demonstration that graph density caused the gain.
People are more than isolated profiles
Two people with similar titles may sit in different professional networks. People with different titles may have relevant shared context. A graph-aware model can use those surroundings alongside profile information, rather than relying on profile resemblance alone.[1][2]
These are representative information categories, not a feature specification. A numerical representation does not establish buying authority, intent, or personal familiarity. Those claims require separate evidence.
Two snapshots. One within-snapshot comparison.
The first offline test used an Energy graph dated August 21, 2026. The September 4 replay reused the recorded model configuration and original model code against an expanded snapshot. The data population changed; the experiment was not retraining against an identical frozen dataset.
The task was link prediction: rank recorded relationships above sampled unrecorded pairs. Within each snapshot, GraphSAGE and the semantic comparator were evaluated on the same labeled test pairs. Positives represented recorded links. Negatives were sampled non-links, not verified absences of real-world relationships.
| Measurement | August 21 | September 4 |
|---|---|---|
| People represented | 47,326 | 57,083 |
| Recorded relationships | 21,348 | 54,267 |
| People with at least one link | 9,289 | 17,366 |
| Connected coverage | 19.63% | 30.42% |
| Average degree, including isolates | 0.902 | 1.901 |
Source: ExecGraph Energy offline experiment records, August 21 and September 4, 2026. These are experimental snapshot counts, not current platform totals or claims about a licensable dataset.
Recorded relationships increased by 154.2%, while the population increased by 20.6%. Connected coverage rose by 10.79 percentage points, but approximately 69.58% of September nodes still had no recorded edge. More links did not make the whole graph well connected.
Evaluation details needed to interpret the scores
The August test contained 2,077 positives and 2,077 negatives; September contained 5,291 of each. Validation and test positive edges were excluded from message passing. Known relationships were excluded from negative sampling, and the degree feature used training edges only. The negative-sampling universe and exclusion policy used the recorded snapshot, so this is not a strictly temporal forecasting test.
Checkpoint selection used validation average precision. Both experiment records report matching same-seed reruns, but those repeats are not independent-seed robustness evidence. The full cross-date runtime environment was not established as identical.
GraphSAGE also used structured context and supervised fitting. The semantic comparator therefore does not isolate topology, and it is not the complete production Warm Path V2 product. Feature-matched non-graph and controlled-topology comparisons remain necessary.
The challenger improved. The baseline still led on AUC.
| Snapshot | Method | ROC AUC | Average precision |
|---|---|---|---|
| GraphSAGE | 0.8302 | 0.8731 | |
| Semantic comparator | 0.8824 | 0.8740 | |
| GraphSAGE | 0.8802 | 0.9071 | |
| Semantic comparator | 0.9058 | 0.9061 |
Source: ExecGraph's own experiments. Scores rounded to four decimal places. External references explain methods and metrics; their authors did not evaluate or validate these results.
GraphSAGE gained 5.00 AUC points; the semantic comparator gained 2.34 points. The GraphSAGE AUC deficit narrowed by approximately 51.1%, calculated from unrounded experiment values. The semantic comparator nevertheless retained the higher September AUC: 0.9058 versus 0.8802.
Average precision moved from a small GraphSAGE deficit to a small numerical lead: 0.9071 versus 0.9061. That roughly 0.0010 difference has no estimated confidence interval here and does not establish superior top-five recommendations.
What AUC and average precision actually measure
ROC AUC describes how often a labeled positive ranks above a labeled negative, with tie handling. It is not the percentage of useful contacts or successful introductions. An AUC around 0.88 describes pairwise discrimination on the evaluated labels.[3]
Average precision summarizes precision as recall increases through a ranking. It is not precision at five and is sensitive to the evaluated class mixture. The balanced test does not establish how often a production recommendation will be useful.[4]
Interpretation: promising progress under the replay conditions, sufficient to motivate a product-level shadow comparison. Not evidence to replace the production system. Both recorded experiments specified zero GraphSAGE production traffic and unchanged Warm Path rankings.
What the measurements leave open
The effect of network structure is not isolated
Node populations, relationships, features, and evaluation pairs changed between dates. Metadata and supervision also distinguish the challenger from the comparator. A fixed-population experiment that changes only training topology would make the causal question more interpretable.
Unrecorded does not mean unrelated
Some sampled non-links may be real relationships absent from the records. Plausible same-context alternatives can also be harder than random pairs. These concerns align with published work on link-prediction evaluation, which emphasizes strong baselines, consistent splits, and realistic negative examples.[5]
Repeatability is not statistical confidence
Graph edges share people and contexts, so they are not necessarily independent observations. No independent multi-seed study or cluster-aware uncertainty estimate accompanies these point estimates. The tiny AP lead should not be described as significant.
A link score is not a usable introduction
The test does not label relationship strength, willingness to introduce, buying authority, or customer-specific access. Better aggregate scores may fail to improve the first few recommendations. No meeting, revenue, sales-cycle, or customer-accuracy lift was measured.
The public release is not a replication package
Only aggregate findings are released. Person-level records, topology, encodings, model settings, and checkpoints remain private. The retained summaries also do not provide a complete historical input-and-prediction package, so full retraining and retrospective significance analysis cannot be established from the public material.
The next test is whether a useful option was missed.
The proposed product comparison freezes real Warm Path targets, user context, eligibility rules, and baseline outputs. It then separates two experiments: reranking the same eligible candidate set, and expanding the candidate set to test discovery. New retrieval should not be mistaken for better ordering.
Agreement at the first position, top-three and top-five overlap, rank displacement, valid-path coverage, and judged precision at five would characterize the difference. The central endpoint is validated added value: an additional relevant, evidence-supported option the baseline shortlist missed. A different name is not automatically an improvement.
Further technical work should include feature-matched non-graph baselines, controlled edge ablations, harder negatives, independent seeds, and temporal or node-disjoint holdouts. These are proposed tests, not completed results.
For external research teams, the substantive question is whether professional context and relationship structure support better evaluation, retrieval, or representation learning. A general-purpose model-training benefit has not been demonstrated by this link-prediction benchmark.
Sources and research lineage
ExecGraph Energy conducted the two applied experiments. The references below establish the published method, metric definitions, and evaluation context. Citation does not imply an affiliation, endorsement, or partnership.
- [1] Inductive Representation Learning on Large GraphsWilliam L. Hamilton, Rex Ying, and Jure Leskovec. NeurIPS 2017; arXiv:1706.02216.Original GraphSAGE method. Not a source for ExecGraph's experiment results.
- [2] GraphSAGE projectStanford Network Analysis Project. Original project documentation.Node features, neighborhood aggregation, and inductive representations.
- [3] Classification: ROC and AUCGoogle for Developers. Machine Learning Crash Course.ROC AUC definition and interpretation.
- [4] average_precision_scorescikit-learn. Official metric documentation.Recall-weighted, non-interpolated AP definition; not a claim about the experiment's software implementation.
- [5] Evaluating Graph Neural Networks for Link Prediction: Current Pitfalls and New BenchmarkingJuanhui Li, Harry Shomer, Haitao Mao, Shenglai Zeng, Yao Ma, Neil Shah, Jiliang Tang, and Dawei Yin. NeurIPS 2023, Datasets and Benchmarks.Motivation for shared evaluation settings, stronger baselines, and more realistic negative examples. Not an independent validation of ExecGraph.
Citation for this research note
ExecGraph Energy (2026). A richer relationship graph, a smaller performance gap. EXG-RN-001, version 1.0. Published September 5, 2026. Offline research note, not peer reviewed. Use the canonical URL below to reference these aggregate observations.
@techreport{execgraph2026energygraph,
author = {{ExecGraph Energy}},
title = {A richer relationship graph, a smaller performance gap: GraphSAGE and a semantic baseline across two Energy graph snapshots},
institution = {ExecGraph Energy},
year = {2026},
number = {EXG-RN-001},
version = {1.0},
url = {https://execgraphenergy.com/research/energy-graphsage-replay-2026-09},
note = {Offline research note; not peer reviewed. Experiments: August 21 and September 4, 2026.}
}
Revision history: version 1.0, September 5, 2026, first public release of the August 21 and September 4 aggregate comparison. Corrections or later experiments will receive a dated update rather than silently replacing this result. Report a correction.