Keep the definitions close.

Notation, primary sources and a reproducible arithmetic companion.

Notation at a glance

Graph and adjacency

$G=(V,E)$ names nodes and connections. Our graph is undirected, without edge features.

Feature row

$x_i\in\mathbb R^d$ describes node $i$. Here $d=2$, with dimensionless synthetic channels.

Mask set

$M$ selects training targets. Corruption and loss selection are separate uses of this set.

Representation

$h_i$ is an encoder output; it need not have the input dimension.

Re-masking

Zero the projected representation at the same selected nodes before the graph decoder.

SCE

Average $(1-c_i)^\gamma$ over selected nodes, with an explicit zero-vector policy.

Prescribed reconstruction

A chosen vector for inspecting the objective. No learning algorithm produced the lab's vectors.

Reproduce the arithmetic

The canonical numerical module (local copy) powers both the live inspector and static comparisons. It uses native JavaScript numbers, fixed arrays and no randomness. The small floating-point clamp keeps cosine within its mathematical range; zero vectors are rejected. This differs from the author's epsilon-normalized tensor code.

node lab.mjs
node --test model.test.mjs

Run these from this guide's site directory with Node.js. lab.mjs (local copy) prints inputs, baseline, all control states and transfer checks as deterministic JSON. It does not fetch data or train weights. The numerical checks (local copy) include all 36 UI configurations and invalid boundaries.

Claim-level evidence ledger (local copy) · Concept blueprint (local copy) · Lab provenance (local copy)

Sources behind the explanation

Primary sources checked 4 September 2026. The code is pinned to a specific public revision. Source-backed facts, local derivations and proposed evaluations are labeled separately; no benchmark or learner gains are reported.

1. Hou et al. (2022). GraphMAE: Self-Supervised Masked Graph Autoencoders.

   Original paper, arXiv v3, §3. Benchmarks have their own datasets and protocols.

2. THUDM/GraphMAE: core encoder–decoder implementation.

   Pinned commit b14f080c919257b495e3cb6474286d5252d6a635. Input token, corruption, projection, re-mask, decoding and embed.

3. THUDM/GraphMAE: SCE implementation.

   Normalize, dot, residual power, mean. Standalone alpha default differs from configured arguments.

4. THUDM/GraphMAE: arguments.

   CLI defaults mask_rate .5, replace_rate 0, alpha_l 2; dataset settings may override them.

5. THUDM/GraphMAE: dataset configuration.

   Dataset-specific settings, not a universal recommended mask ratio.

6. THUDM/GraphMAE: inductive training and evaluation.

   Inspect train/validation/test loaders and frozen representation evaluation.

7. THUDM/GraphMAE: chemistry pretraining.

   Separate optional mask_edge branch; do not generalize one path to the whole repository.

8. Gilmer et al. (2017). Neural Message Passing for Quantum Chemistry.

   §2, shared message/update formulation. Our mean aggregation is illustrative.

9. Veličković et al. (2018). Graph Attention Networks.

   §2, attention-weighted neighborhood aggregation.

10. Hou et al. (2023). GraphMAE2: A Decoding-Enhanced Masked Self-Supervised Graph Learner.

    §2.2 and Figure 2, multi-view random re-masking and latent representation prediction.

11. He et al. (2022). Masked Autoencoders Are Scalable Vision Learners.

    §3, image patch masking and asymmetric architecture.

12. PyTorch: torch.nn.functional.normalize.

    Epsilon-protected normalization. Checked 4 September 2026; lab instead rejects zero norm.