---
title: 'AGF-TI: Dual Roles in ML & Condensed Matter'
url: https://www.emergentmind.com/topics/agf-ti
type: topic
---

# AGF-TI: Dual Roles in ML & Condensed Matter

AGF-TI has two distinct meanings in recent research. In machine learning, it denotes “Adversarial Graph Fusion for Incomplete Multi-view Semi-supervised Learning with Tensorial Imputation,” a method for incomplete multi-view semi-supervised learning that combines adversarial graph fusion, tensorial recovery of missing graph structure, and anchor-based label propagation [2509.15955]. In condensed-matter physics, the same label denotes the dual-topological character of monolayer AgF$_2$, where a single ferroelastic/polar distortion simultaneously generates non-trivial electronic and magnonic topology in a relativistic altermagnet [2508.15944].

## 1. Terminological scope

The acronym is therefore domain-dependent rather than universally standardized. The two principal usages represented in the literature considered here are summarized below.

| Usage | Domain | Defining content |
|---|---|---|
| AGF-TI | Machine learning | Adversarial graph fusion with tensorial imputation for incomplete multi-view semi-supervised learning |
| AGF-TI | Condensed matter | Dual topology in monolayer AgF$_2$: fermionic high-Chern subbands plus a Chern magnon insulator |

In the machine-learning usage, AGF-TI is a graph-based semi-supervised framework designed for view-missing data, with explicit notation for $V$ views, $n$ samples, $m$ anchors, $\ell$ labeled samples, and $c$ classes. In the AgF$_2$ usage, “AGF-TI” denotes the dual topological character of monolayer AgF$_2$ (“AgF”), with “TI” used in a generalized “topological index/insulator” sense rather than as a time-reversal-invariant $Z_2$ topological insulator [2509.15955][2508.15944].

This dual usage is not merely lexical. The machine-learning construction addresses robustness under incomplete observations and distorted local geometry, whereas the AgF$_2$ construction concerns a symmetry-driven coexistence of bosonic and fermionic topological invariants in a single polar altermagnetic phase. The shared acronym therefore masks two unrelated but internally coherent technical programs.

## 2. AGF-TI in incomplete multi-view semi-supervised learning

In the machine-learning literature, AGF-TI is formulated for incomplete multi-view semi-supervised learning, where each view has features $X_v \in \mathbb{R}^{n \times d_v}$, the first $\ell$ samples are labeled, and the one-hot label matrix satisfies $Y \in \{0,1\}^{n \times c}$. The method is built around per-view sample-anchor bipartite graphs $Z_v \in \mathbb{R}^{n \times m}$, anchor-alignment matrices $T_v \in \mathbb{R}^{m \times m}$ with $T_v^\top T_v = I_m$, and a consensus fused bipartite graph $P \in \mathbb{R}^{n \times m}$ constrained to the row-simplex domain $\Delta_m^n$ [2509.15955].

Its central motivation is the “Sub-Cluster Problem” (SCP). SCP refers to the fragmentation of what should be a single manifold cluster into disconnected sub-clusters because missing samples create “vacuum regions” in a view. The stated consequence is that the classical smoothness assumption for label propagation is violated: labels cannot diffuse smoothly across the vacuum region, and graph fusion becomes distorted because per-view graphs omit missing nodes and misrepresent the underlying manifold geometry. This is the specific failure mode AGF-TI is designed to alleviate [2509.15955].

The fused sample-anchor graph is defined as
\[
S_P=
\begin{bmatrix}
0 & P\\
P^\top & 0
\end{bmatrix},
\]
with normalized Laplacian
\[
\tilde{L}_{S_P}=I_{n+m}-D_{S_P}^{-1/2}S_PD_{S_P}^{-1/2}.
\]
Within this graph, semi-supervised learning is expressed through a Laplacian smoothness objective. In the single-view form given in the source,
\[
\mathcal{L}(F,L_S)=\mathrm{Tr}(F^\top L_SF)+\mathrm{Tr}\big((F-Y)^\top B(F-Y)\big),
\]
where $F \in \mathbb{R}^{n \times c}$ contains soft labels and $B$ is a diagonal fitting-weight matrix. This smoothness-based formulation is then lifted to the fused sample-anchor graph used by AGF-TI [2509.15955].

## 3. Adversarial graph fusion, tensorial imputation, and anchors

The adversarial graph-fusion component learns a robust consensus bipartite graph from incomplete per-view graphs through a min-max formulation. The operator is
\[
\min_{\alpha \in \Delta_V^1}\max_{P \in \Delta_m^n}
\sum_{v=1}^V \alpha_v^2\, \mathrm{Tr}\!\big(P^\top(Z_vT_v)\big)-\beta\|P\|_F^2,
\quad \text{s.t. } T_v^\top T_v=I_m,
\]
where $\alpha$ are per-view weights on the simplex and $\beta>0$ regularizes the connectivity of $P$ [2509.15955].

The full AGF-TI objective couples adversarial fusion, label propagation, and tensorial imputation:
\[
\max_{\{Z_{\omega_v},T_v\}_{v=1}^V,\hat{F}}
\min_{\alpha \in \Delta_V^1}
\max_{P \in \Delta_m^n}
\mathcal{R}(\hat{F},\tilde{L}_{S_P})
+\lambda\,\mathrm{AGF}(P,\{Z_v\}_{v=1}^V)
-\rho\|\mathcal{Z}\|_{\circledast},
\]
with $\mathcal{Z}=\Phi(Z_1,\ldots,Z_V)$. Here $\lambda$ balances adversarial fusion and label propagation, while $\rho$ controls tensorial imputation strength [2509.15955].

Tensorial imputation is the mechanism used to recover missing sample-anchor relations. The per-view bipartite graphs are stacked into a third-order tensor
\[
\mathcal{Z}=\Phi(Z_1,\ldots,Z_V)\in \mathbb{R}^{m \times V \times n},
\]
whose low-rank structure is enforced through the tensor nuclear norm based on t-SVD. The stated rationale is that “high-order consistency” across anchors, views, and samples enables recovery of incomplete graph structure that is invisible to viewwise methods. This makes imputation a joint component of the optimization rather than a preprocessing step [2509.15955].

Anchors provide the computational reduction. Instead of building dense sample-sample graphs, AGF-TI operates directly on sample-anchor bipartite graphs. The anchor construction problem is
\[
\min_{Z_v}\sum_{i=1}^n\sum_{j=1}^m
\big(\|x_i^{(v)}-a_j^{(v)}\|_2^2\, Z_{ij}^{(v)}+\gamma (Z_{ij}^{(v)})^2\big),
\quad
\text{s.t. } Z_v1_m=1_n,\; Z_v\ge 0.
\]
For missing samples, the rows $Z_{\omega_v}$ are initialized to equal probabilities $1/m$ and later updated within the alternating scheme. This anchor-based strategy reduces the stated time complexity from $O(n^3)$ to $O(nm^2)$ and the space complexity from $O(Vn^2)$ to $O(Vnm)$, with $m \ll n$ [2509.15955].

## 4. Optimization, empirical behavior, and limitations

The optimization is an alternating procedure with an inner min-max problem in $(\alpha,P)$ and outer ADMM-style updates for $Z_{\omega_v}$, $T_v$, $\hat{F}$, and the auxiliary tensor variable $G$. The inner objective is
\[
h(\alpha)\triangleq
\max_{P \in \Delta_m^n}
\left[
\lambda\sum_v \alpha_v^2 \mathrm{Tr}(P^\top(Z_vT_v))
-\beta_\lambda\|P\|_F^2
-\mathrm{Tr}(\hat{F}^\top \tilde{L}_{S_P}\hat{F})
\right].
\]
The paper states that $h(\alpha)$ is convex in $\alpha$, that the gradient is
\[
\frac{\partial h(\alpha)}{\partial \alpha_v}
=2\lambda \alpha_v \mathrm{Tr}\big(P^{\star\top}Z_vT_v\big),
\]
and that reduced gradient descent with Armijo line search converges to the global optimum of the inner minimization over $\alpha$ [2509.15955].

For label propagation, the block variable $\hat{F}=[F;Q]$ is updated by
\[
\hat{F}^\star=(\tilde{L}_{S_P}+\hat{B})^{-1}\hat{B}\hat{Y},
\]
with a Woodbury-based block inversion that lowers the principal cost of the update to $O(nm^2)$. The paper further states that, under a mild “sufficient inner optimization” assumption, the outer ADMM iterations yield bounded sequences and the overall algorithm converges to a stationary KKT point of the full objective [2509.15955].

Empirically, AGF-TI is evaluated on CUB, UCI-Digit, Caltech101-20, OutScene, MNIST-USPS, and AwA, using View Missing Ratio (VMR) in $\{30\%,50\%,70\%\}$, Label Annotation Ratio (LAR) in $\{1\%,2\%,4\%,8\%\}$, and the metrics ACC, Precision, and F1. Each experiment is repeated 10 times. Under LAR$=5\%$, the reported highlights include UCI-Digit ACC values of $95.98\%$, $95.23\%$, and $95.16\%$ at VMR $=30$, $50$, and $70$, respectively; CUB ACC values of $78.33\%$, $80.23\%$, and $74.25\%$; OutScene ACC values of $70.34\%$, $69.24\%$, and $64.52\%$; and MNIST-USPS ACC values of $95.09\%$, $95.58\%$, and $95.62\%$ [2509.15955].

The ablation results identify three components as consequential: anchor alignment $T_v$, view weights $\alpha_v$, and tensorial imputation. On Caltech101-20 and OutScene, removing $T_v$ causes large drops; a cited example is ACC decreasing from $81.88$ to $42.71$ on Caltech101-20 at VMR $=30\%$. Tensorial imputation becomes especially important at high missing rates, where removing TI causes ACC to drop by up to approximately $17.8\%$ at VMR $=70\%$ versus approximately $6\%$ at VMR $=30\%$. The reported robust parameter ranges are $\log_2(\beta_\lambda)\in[0,2]$ and $\log_{10}(\rho)\in[2,5]$, with $\lambda$ fixed to $V^2$ in the experiments. Runtime examples at VMR$=50\%$, LAR$=5\%$ are $1.06$ s for CUB, $15.35$ s for UCI-Digit, $45.03$ s for Caltech101-20, $35.84$ s for OutScene, $31.06$ s for MNIST-USPS, and $354.54$ s for AwA [2509.15955].

The stated limitations are equally specific. Anchors are pre-selected rather than jointly optimized with fusion and imputation; the convergence proof assumes sufficiently accurate inner optimization; gains from tensorial imputation may diminish when cross-view correlations are weak; and the tensor nuclear norm step can be heavy for extremely large $m$ or $V$. These caveats place AGF-TI within the broader class of structured but computationally nontrivial graph-based SSL methods [2509.15955].

## 5. AGF-TI as dual topology in monolayer AgF$_2$

In condensed-matter usage, AGF-TI refers to monolayer AgF$_2$ as a “dual-topological” altermagnet, meaning that both electronic and magnonic excitations acquire non-trivial Chern topology through the same ferroelastic/polar distortion [2508.15944]. The high-symmetry reference structure is a flat centrosymmetric monolayer with space group $P4/mmm$ (No. 123), whereas the ground state is a tilted polar structure with space group $P2_1$ (No. 4). The order parameter $Q_{\mathrm{FE}}\sim (\theta,\phi,h)$ is quantified by out-of-plane tilt $|\theta|=24.1^\circ$, in-plane rotation $|\phi_{21}|=14.4^\circ$, and vertical buckling $h=1.49$ Å, with in-plane anisotropy $a/b \approx 0.87$. The flat reference is metastable and higher in energy by approximately $0.2$ eV/f.u. [2508.15944].

The magnetic state is a collinear, compensated altermagnet in the nonrelativistic limit, with zero net moment, while spin-orbit coupling induces a tiny weak-ferromagnetic canting of approximately $0.1\,\mu_B$ per formula unit. The distortion breaks inversion and lowers rotational symmetry while retaining a nonsymmorphic $2_1$ screw along the crystallographic $b$ axis. Without SOC, the system exhibits even-parity, momentum-symmetric spin splitting $\Delta E(k)=\Delta E(-k)$, with anisotropy of approximately $100$ meV along $M'$–$\Gamma$–$M$ and vanishing splitting along $\Gamma$–$Y$. With SOC, a multipolar analysis gives an overall balance $f_d \approx 0.53$ and $f_p \approx 0.47$, with $p$-wave strongest in the $y$ and $z$ channels and $d$-wave dominant along $x$ [2508.15944].

This structural and magnetic setting is essential because the same distortion activates both topological sectors. The paper describes the mechanism as follows: inversion breaking and SOC open avoided electronic crossings and enable strong Dzyaloshinskii–Moriya interaction on the distorted bonds, so the single structural order parameter $Q_{\mathrm{FE}}$ simultaneously topologizes electrons and magnons. In this usage, AGF-TI is therefore a material classification rather than an algorithmic framework [2508.15944].

## 6. Electronic and magnonic topology, control, and broader AgF$_2$ context

The electronic sector of monolayer AgF$_2$ is not classified as a time-reversal-invariant $Z_2$ topological insulator, because time-reversal symmetry is broken by magnetic order, and it is not yet an intrinsic quantum anomalous Hall insulator in its pristine ground state because the net occupied Chern number cancels. Instead, the top two valence bands carry opposite high Chern numbers,
\[
C_1^E=+3,\qquad C_2^E=-3,
\]
so the total occupied valence-manifold Chern number remains $C_{\mathrm{tot}}^E=0$. SOC opens avoided crossings up to approximately $20$ meV near $M$ and $M'$, where the Berry curvature concentrates. The paper states that if a perturbation such as uniaxial strain, an out-of-plane electric field, a staggered substrate potential, or appropriate heterostructure coupling isolates one of these bands and places $E_F$ in the Chern gap, a QAH state with
\[
\sigma_{xy}=Ce^2/h,\qquad |C|=3
\]
should be obtainable, with three co-propagating electronic edge channels per boundary [2508.15944].

The magnon sector is already topological in the distorted ground state. In the flat $P4/mmm$ reference, DMI is symmetry-forbidden and the magnon branches are degenerate and topologically trivial. In the polar $P2_1$ phase with SOC, representative exchange parameters are $J_1 \approx -10.66$ meV, $J_2 \approx -10.69$ meV, $J_4 \approx -10.61$ meV, with $|D| \approx 1.55$–$1.99$ meV on dominant bonds and symmetric anisotropy up to approximately $1.7$. Linear spin-wave theory and bosonic BdG analysis yield two non-degenerate magnon branches with a full topological gap and Chern numbers
\[
C_1^M=+1,\qquad C_2^M=-1.
\]
Bulk-boundary correspondence then implies one chiral magnon edge mode per boundary traversing the gap, together with a finite transverse thermal Hall conductivity $\kappa_{xy}^{\mathrm{mag}}(T)$ below the estimated Néel temperature $T_N \approx 100$–$150$ K [2508.15944].

The unifying control parameter is ferroelastic/polar domain switching. Reversing the polar axis, $Q_{\mathrm{FE}}\to -Q_{\mathrm{FE}}$, reverses the relevant symmetry-allowed SOC and DMI vectors, flips the signs of both the electronic and magnonic Chern indices, and reverses the anomalous charge and heat Hall responses. The paper identifies the transverse thermal Hall effect as the clearest macroscopic fingerprint of the magnon Chern bands, while ribbon spectroscopy or multi-terminal transport would be natural probes of a strain- or field-induced $|C|=3$ QAH state. Chiral magnon edge modes are proposed to be accessible by Brillouin light scattering or nonlocal magnon transport, and RIXS is identified as a probe of both multi-magnon dynamics and electronic band splittings [2508.15944].

Broader AgF$_2$ chemistry provides context for why this topological proposal is notable. Independent spectroscopy and cluster/DFT analysis identify AgF$_2$ as a charge-transfer insulator closely analogous to undoped La$_2$CuO$_4$, with optical interband absorption onset at approximately $1.75$ eV, a pronounced charge-transfer peak centered at about $3.4$ eV, an estimated optical fundamental gap $E_{\mathrm{gap}} \approx 2.2 \pm 0.3$ eV, and F K-edge RIXS $dd$ excitations at $1.59$, $1.86$, and $2.33$ eV [2105.08862]. Related theoretical work on Ag$_2$OF$_2$ explicitly describes AgF$_2$ as an antiferromagnetic charge-transfer positive-$U$ insulator with Ag(II) $d^9$ character [2103.04733]. This suggests that the dual-topological monolayer proposal is embedded in a broader silver-fluoride materials landscape already known for strong covalency, charge-transfer physics, and unusual competing phases.

Source: https://www.emergentmind.com/topics/agf-ti