# Signed Theta Geometric Embedding Loss

- ID: 1686
- Canonical URL: https://synthcore.org/idea/1686/signed-theta-geometric-embedding-loss
- API JSON: https://synthcore.org/api/idea/1686.json
- API Markdown: https://synthcore.org/api/idea/1686.md
- Verification status: unverified
- Source: [arXiv:2608.01716](https://arxiv.org/abs/2608.01716)
- Category: regularization
- Solves: accuracy, generalization, stability
- ML areas: graph-nn, embedding, regularization
- Math tags: geometry, linear-algebra, convex-analysis
- Ratings: usefulness 5/10; difficulty 3/10; novelty 6/10

## Idea description

Use the Lovász-style prescribed inner product as a differentiable regularizer on node embeddings. Positive and negative signed relations are compared through the identity or the involution respectively, encouraging a representation whose geometry respects signed colouring constraints and remains invariant to switching gauges.

## Mathematical statement

Let z_i=u_i+v_i be the embedding of node i, with u_i in the fixed subspace E_+ and v_i in the anti-fixed subspace E_-. Let J(u_i,v_i)=(u_i,-v_i), so J^2=I. For a target continuous colouring parameter q, the paper gives the prescribed adjacent-pair inner product c(q)=1/(1-q). Define the normalized embedding r_i=z_i/||z_i||_2. For edge sign sigma_ij, use r_j unchanged when sigma_ij=+1 and Jr_j when sigma_ij=-1. The resulting soft penalty compares the signed relation inner product with c(q). Under switching by s_i, sigma_ij becomes s_i sigma_ij s_j and the anti-fixed embedding becomes s_i v_i, so the relation geometry is gauge-consistent.

## Key formulas

- $$\langle z_i,z_j\rangle=\frac{1}{1-q}$$
- $$J(u,v)=(u,-v),\qquad J^2=I$$
- $$r_i=\frac{z_i}{\|z_i\|_2},\qquad c(q)=\frac{1}{1-q}$$
- $$\mathcal L_{\mathrm{signed\text{-}theta}}=\frac{1}{|E|}\sum_{(i,j)\in E}\left(\left\langle r_i,J^{(1-\sigma_{ij})/2}r_j\right\rangle-c(q)\right)^2$$

## Implementation notes

Attach this regularizer to the penultimate node embeddings of a signed GNN, graph transformer, or relation-aware encoder. Split each embedding z_i into equal-width fixed and anti-fixed parts, z_i=(u_i,v_i), and normalize with r_i=z_i/(||z_i||_2+epsilon). Define J(u_i,v_i)=(u_i,-v_i). For each edge (i,j), set k_ij=(1-sigma_ij)/2, so k_ij=0 for a positive edge and 1 for a negative edge. Compute t_ij=dot(r_i,J^{k_ij}r_j), then add lambda times the mean squared penalty (t_ij-c(q))^2 to the task loss, with c(q)=1/(1-q). Initially fix q, for example q=0 or q=0.25; alternatively learn q through q=q_min+(q_max-q_min)sigmoid(a), using a scalar parameter a and safe bounds that keep the target geometrically feasible. Sample edges in minibatches for large graphs. Generate switched copies using random vertex signs s_i, set sigma'_ij=s_i sigma_ij s_j, and multiply each anti-fixed input feature by s_i; verify that the regularizer and graph output are unchanged. Compare cross-entropy alone and cross-entropy plus this loss on signed stochastic block models and a small signed link-prediction dataset, sweeping lambda in {0.01,0.1,1}. Track task accuracy, link AUC, performance after gauge switching, embedding norms, and gradient variance. A successful result is higher accuracy or AUC at equal parameter count, negligible degradation under switching, and fewer unstable training runs.

## Disclaimer

AI-generated research hypothesis, automatically tested. Not peer-reviewed.