Model-Oriented Sub-population and Spectral Analysis

• MOSSA is a computational framework that uses spectral graph techniques and auxiliary covariate data to decompose complex datasets into meaningful sub-populations.
• It integrates low-frequency eigenvector constraints to enforce smooth, interpretable sub-cohort stratification, thereby enhancing downstream predictive accuracy.
• Empirical applications in neuroimaging and astrophysics demonstrate robust sub-cohort identification, reduced degeneracy, and improved model convergence.

Model-Oriented Sub-population and Spectral Analysis (MOSSA) encompasses a family of computational methodologies for revealing, characterizing, and leveraging subpopulation structure within complex datasets by integrating model-informed sample weighting and spectral graph techniques. The central principle is to use auxiliary covariate information and graph-theoretical spectral decomposition to construct interpretable subject-wise weights or compositions, ultimately improving downstream predictive modeling and enabling rigorous sub-cohort interpretation. MOSSA encompasses tools such as Spectral Graph Sample Weighting (SGSW) for neuroimaging population analysis (Paschali et al., 2024) and Fitting Analysis using Differential Evolution Optimization (FADO) for spectral population synthesis in extragalactic research (Gomes et al., 2017).

1. Population Graph Construction and Spectral Decomposition

A foundational step in MOSSA is to structure the dataset as a spectral population graph, where subjects (indexed $i = 1, \ldots, N$) are nodes $V = \{1, \ldots, N\}$, and their pairwise factor-similarity relations define the edge set $E$ via an affinity matrix $A \in \mathbb{R}^{N \times N}$. Auxiliary subject-level vectors $s_i \in \mathbb{R}^D$ (e.g., demographic, clinical, or genetic features) provide a factor space over which affinity is computed, typically using a $K$-nearest neighbor protocol:

$$A_{ij} = \begin{cases} \frac{1}{\|\mathbf{s}_i - \mathbf{s}_j\|^2 + 1} & \text{if } j \in \mathcal{N}_K(i) \text{ or } i \in \mathcal{N}_K(j) \ 0 & \text{otherwise} \end{cases}$$

Defining the degree matrix $D_{ii} = \sum_j A_{ij}$, the unnormalized Laplacian is $L = D - A$. Spectral decomposition solves $L u_{\ell} = \lambda_{\ell} u_{\ell}$, $\lambda_0 \leq \ldots \leq \lambda_{N-1}$, extracting $M$ low-frequency eigenvectors $U_M = [u_1 \ldots u_M]$ which serve as a graph-Fourier basis for representing smooth, population-level functions.

2. Sample Weighting and Model Parametrization

MOSSA applies spectral expansion to model subject-level weights $w = [w_1 \ldots w_N]^T$:

$w = c \cdot 1_N + U_M \alpha$

where $c \in \mathbb{R}$ is a constant shift and $\alpha \in \mathbb{R}^M$ parameterizes the weight projection in the eigenbasis. Limiting $w$ to low-frequency modes enforces smoothness with respect to the population graph, so similar subjects (by auxiliary factors) receive similar weights. This weighting scheme is integrated into the predictive model’s loss function, yielding a weighted training objective for parameters $\theta$ (model) and $\alpha$ (weights):

$$\mathcal{L}(\theta, \alpha) = \sum_{i=1}^{N'} w_i(\alpha) \, \ell(f(x_i; \theta), y_i) + \sum_{i=1}^{N'} \max(0, -w_i(\alpha)) + \frac{\lambda_\theta}{2} \|\theta\|_2^2 + \frac{\lambda_\alpha}{2} \|\alpha\|_2^2$$

where $\ell$ is the sample-wise loss (e.g., binary cross-entropy), and explicit penalties control nonnegativity and regularization.

3. Sub-cohort Identification and Interpretability

After model training, the learned weights $w$ provide a data-driven means to stratify the population into interpretable sub-cohorts. By thresholding $w$ (e.g., at the median), distinct sub-populations characterized by factor composition and model accuracy are identified. Visualization techniques, such as box-plots of $w$ against auxiliary factors (e.g., sex, age, SES, genotype status), elucidate which groups the predictive model is most reliant upon or most accurate for. Clustering $U_M \alpha$ in factor space can yield finer-grained sub-cohort partitions.

In neuroimaging contexts, this approach reveals that predictability and learned weight assignments align with established clinical and demographic heterogeneities; for instance, higher balanced-accuracy and weights for younger subjects or females in alcohol use initiation prediction, and genotype-based stratification in dementia risk modeling (Paschali et al., 2024).

4. Spectral Population Synthesis via Evolutionary Optimization

An alternative MOSSA realization, as exemplified by FADO, addresses astrophysical spectral population synthesis (PSS) by inferring sub-population compositions and nebular continuum fractions from observed galaxy spectra. FADO casts the inverse PSS task as:

$$M_\lambda(\vec{p}) = \biggl( \sum_{j=1}^{N_\star} M_j L_{j,\lambda} 10^{-0.4 A_V q_\lambda} \otimes S(v_\star, \sigma_\star) \biggr) + \biggl( \Gamma_\lambda(n_e, T_e) 10^{-0.4 A_V^{\rm neb} q_\lambda} \otimes N(v_{\rm neb}, \sigma_{\rm neb}) \biggr)$$

with sub-population weights represented as normalized light fractions $x_j$, nebular fraction $y$, and subject to self-consistency constraints from spectral physics (e.g., LyC photon rate $Q_H$, predicted Balmer line luminosities $L_{H\alpha}^{\rm mod}$, Case B recombination boundary conditions) (Gomes et al., 2017).

FADO employs a Differential Evolution Optimizer (DEO), where each chromosome encodes trial vectors $$\vec{x}_i = (x_{1,i}, ..., x_{N_\star,i}, y_i, ...)$$, and the search is performed under feasibility constraints and multi-objective criteria (continuum fit, line matches, parameter bounds). Artificial intelligence methods are used for spectral library pruning via clustering, accelerating convergence and preserving spectral coverage.

5. Algorithmic Features: Optimization, Parallelization, and Convergence

Both SGSW and FADO frameworks integrate advanced computational strategies:

• Stochastic optimization (Adam for SGSW; DEO for FADO) enables simultaneous learning of model and weight parameters.
• FADO utilizes quasi-parallelization with Fortran 2008 coarrays or OpenMP pragmas for population-wise computation, achieving run times of 1–5 minutes per galaxy spectrum with competitive speed to classical codes.
• Convergence diagnostics rely on variance ratio tests (Gelman–Rubin style) for evolutionary approaches, halting when between-generation and within-generation variances equilibrate and progress stalls.

6. Empirical Applications and Impact

MOSSA methodologies have demonstrated empirical value in diverse domains:

Application Domain	Dataset	Population Size	Auxiliary Factors	Key Findings
Neuroimaging symptom prediction	NCANDA	N=399	sex, SES, alcohol history	SGSW weights highlight sub-cohorts (e.g. females, low-SES) with higher BACC (66.5% vs. 60.5%) (Paschali et al., 2024)
Dementia/MCI stratification	ADNI	N=1191	sex, age, APOE ε4	Young age/high weight group achieves BACC≈73.5%, genotype effect gap ≈8.5%
Galactic spectral synthesis	SDSS	varies	stellar population, nebular continuum	FADO accurately recovers star-forming history and line EWs with <5% error, outperforming purely stellar models (Gomes et al., 2017)

In each case, sub-cohort interpretability is directly enhanced: learned weights and population decompositions meaningfully correspond to known scientific categories, and predictive accuracy is stratified by these subpopulations.

7. Degeneracy Reduction and Unique Solution Guarantees

A defining feature of MOSSA-based approaches such as FADO is the rigorous imposition of physical or demographic self-consistency constraints. In FADO, nebular emission fractions and line luminosities are tied to the stellar population vector via physically motivated equations, and candidate solutions that mismatch observed emission features are penalized or rejected. This strategy effectively reduces the degeneracy endemic to classical spectral fitting, yielding unique, astrophysically consistent fossil record solutions (Gomes et al., 2017).

Similarly, in the SGSW framework, the restriction of sample weights to low-frequency graph spectra and the smoothness prior (no large-eigenvalue modes) ensures that sub-cohort definitions are stable and interpretable, preventing overfitting to noise or isolated outliers (Paschali et al., 2024).

---

Model-Oriented Sub-population and Spectral Analysis synthesizes spectral graph theory, evolutionary optimization, and domain self-consistency principles to deliver interpretable, robust sub-cohort quantification in both astrophysical and biomedical research, with direct empirical improvements in predictive accuracy, interpretability, and solution uniqueness.