Factorized Influence: A Cross-Domain Overview

• Factorized influence is a cross-domain concept that decomposes complex influence relations into structured latent factors using methods like Kronecker and bilinear decompositions.
• It enhances model interpretability and computational efficiency by replacing high-dimensional interactions with low-rank representations and separable formulations.
• Applications span dynamic recommendation, online influence maximization, and scalable influence-function computation, each leveraging tailored factorization techniques.

Factorized influence denotes a family of formulations in which “influence” is not modeled as an unfactored relation on observed variables, but through a structured decomposition such as latent factors, Kronecker factors, hidden variables, shared temporal patterns, or subsystem dispersion factors. The term is therefore not monosemous across the literature. In recommendation, social influence is embedded in latent user factors regularized by a graph Laplacian; in causal social analysis it is separated from homophily and item preferences by probabilistic factor models; in online influence maximization it appears as a bilinear factorization of edge activation probabilities; in scalable influence-function computation it refers to Kronecker-factored curvature together with factorized sketches; in representation learning it is realized as one latent flow per semantic factor; in temporal network analysis it is a rank-1 nonnegative factorization of influence trajectories; and in Bayesian networks and coupled physical systems it appears as exact multiplicative decompositions of deterministic dependence or dispersion relations (Aravkin et al., 2016, Sridhar et al., 2022, Wu et al., 2019, Hu et al., 11 Feb 2026, Song et al., 2023, Shaojie, 2023, Vomlel, 2012, Figotin, 24 Mar 2026).

1. Taxonomy of the concept

The recurring structural idea is that a high-dimensional or combinatorial influence object is replaced by a lower-dimensional or separable representation that preserves the operational quantity of interest. The object being factorized, however, differs sharply by field.

Domain	Factorized object	Representative form
Dynamic recommendation	Social influence on latent preferences	$\mathrm{tr}(U_t^\top L_t U_t)$
Causal social influence	Outcome intensity with substitute confounders	$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$
Online influence maximization	Edge activation probabilities	$p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$
Influence functions	Curvature and sketch	$F=A\otimes E,\; P=P_A\otimes P_E$
Representation learning	Semantic transformations as latent flows	$z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$
Temporal network analysis	Participant trajectories	$M^* \approx WH$
Bayesian-network inference	Deterministic CPTs via hidden variable	$\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$
Coupled physical systems	Dispersion relation	$$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$

Across these formulations, factorization may be exact or approximate. In some cases it is an algebraic identity, as with deterministic functional dependence and coupled-system dispersion. In others it is a modeling assumption or regularized estimator, as with low-rank recommendation, Poisson influence factorization, or participant-invariant NMF. A common misconception is that factorization always means “one scalar latent per factor.” The flow-based literature explicitly rejects that reduction: a factor may instead be represented by an entire latent vector field or probability flow (Song et al., 2023).

2. Latent social influence in dynamic recommendation

A particularly explicit use of the term appears in dynamic matrix factorization with social influence. The starting point is standard collaborative filtering with an unknown rating matrix $R\in\mathbb{R}^{m\times n}$, observed through a sampling operator $\mathcal A$, and approximated by a low-rank factorization $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$0 with $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$1 and $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$2. Static matrix factorization solves

$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$3

but the dynamic formulation introduces time-indexed user factors $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$4 and observations $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$5, with latent evolution

$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$6

Social influence is injected through a trust matrix $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$7, degree matrix $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$8, and graph Laplacian $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$9, yielding the regularizer

$p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$0

This regularizer smooths each latent dimension of the user factors over the social graph, so influence acts on $p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$1 rather than directly on observed ratings. The paper therefore realizes “influence in a factorized form” by embedding social influence into the same latent dimensions that explain ratings (Aravkin et al., 2016).

The dynamic model is made more explicit through an augmented state with preference “position” and “velocity,”

$p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$2

and a constant-velocity transition. In the global smoothing formulation, the estimator minimizes a convex quadratic combining a data term, a dynamic prior, and a Laplacian penalty,

$p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$3

Given fixed item factors, the problem is convex in the user trajectory variables. Estimation uses matrix-free gradient-based methods such as L-BFGS, and gradient computation scales as

$p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$4

with $p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$5 time points, latent dimension $p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$6, $p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$7 users, average number of ratings $p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$8 per time, and average number of trust edges $p_{uv}=\boldsymbol\theta_u^\top \boldsymbol\Beta_v$9 per time.

Empirically, the model was evaluated on Epinions after filtering to $F=A\otimes E,\; P=P_A\otimes P_E$0 users with more than $F=A\otimes E,\; P=P_A\otimes P_E$1 ratings, $F=A\otimes E,\; P=P_A\otimes P_E$2 products, $F=A\otimes E,\; P=P_A\otimes P_E$3 ratings, and $F=A\otimes E,\; P=P_A\otimes P_E$4 undirected trust edges, with time quantized into $F=A\otimes E,\; P=P_A\otimes P_E$5 bins. Dynamic matrix factorization improved over static factorization for $F=A\otimes E,\; P=P_A\otimes P_E$6, and adding social influence further reduced test RMSE for intermediate $F=A\otimes E,\; P=P_A\otimes P_E$7, with the best reported RMSE equal to $F=A\otimes E,\; P=P_A\otimes P_E$8 at $F=A\otimes E,\; P=P_A\otimes P_E$9 and $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$0. Very large $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$1 degraded performance through oversmoothing. The formulation assumes low-rank structure, linear-Gaussian dynamics, fixed item factors during smoothing, an observed trust graph, and a single global influence weight $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$2.

3. Causal and online formulations on social networks

In causal observational analysis, factorized influence is used to separate peer effects from confounding by homophily and item preferences. Poisson Influence Factorization defines individual-level social influence for users $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$3 and item $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$4 as

$z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$5

and average influence of user $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$6 as

$z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$7

The key step is to construct substitute confounders from factor models for the network and for past behavior, then fit a Poisson outcome model

$z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$8

Under the paper’s assumptions, the causal estimand equals the influence parameter, $z_t=z_{t-1}+\nabla_z u^k(z_{t-1},t)$9. The substitutes $M^* \approx WH$0 and $M^* \approx WH$1 are obtained from Poisson factor models for adjacency and yesterday’s purchases, including a joint model

$M^* \approx WH$2

A central limitation is that factor models can recover only those latent traits that affect multiple edges or multiple purchases. In semi-synthetic experiments on Pokec, PIF-Joint was the best non-Oracle method across confounding regimes, and on Last.fm it achieved the best held-out log-likelihood, $M^* \approx WH$3, and best held-out AUC, $M^* \approx WH$4, while also estimating much smaller average influence than unadjusted methods (Sridhar et al., 2022).

A different usage appears in online influence maximization, where the unknown edge activation matrix is factorized into node-level broadcaster and receiver factors. In the Independent Cascade setting, the directed edge probability from giving node $M^* \approx WH$5 to receiving node $M^* \approx WH$6 is modeled as

$M^* \approx WH$7

This realizes network assortativity by reducing the parameter count from $M^* \approx WH$8 to $M^* \approx WH$9. The algorithm IMFB estimates $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$0 and $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$1 by regularized least squares with alternating closed-form updates, constructs UCBs for edge probabilities, and invokes an approximate influence-maximization oracle on the optimistic graph. Under standard assumptions, the scaled regret obeys

$\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$2

in the worst case, improving over dense-graph edge-level baselines by roughly a factor of $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$3. Empirically, IMFB outperformed CUCB, $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$4-greedy, IMLinUCB, and DILinUCB on NetHEPT and Flickr, with a controlled Flickr experiment reporting approximately $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$5 improvement over CUCB and $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$6-greedy (Wu et al., 2019).

These two lines of work attach very different meanings to factorization. PIF factorizes latent confounding structure to identify a causal peer-effect parameter, whereas IMFB factorizes edge propensities to improve sample efficiency in sequential learning. Both, however, replace edgewise independence by shared node-level structure.

4. Factorized influence in second-order attribution

In influence-function analysis, the basic quantity is the inverse-sensitive bilinear form

$\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$7

where $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$8 is a curvature operator and $\phi=\sum_b h(Y,b)\prod_i g_i(X_i,b)$9 are gradients. Since forming or inverting $$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$0 is infeasible in modern models, the sketched approximation uses

$$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$1

A core point of the 2026 theory is that Johnson–Lindenstrauss arguments are insufficient because they preserve Euclidean geometry but do not control inversion. For $$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$2, unregularized projection preserves influence exactly if and only if $$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$3 is injective on $$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$4, equivalently $$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$5, which forces $$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$6. With ridge regularization, the barrier shifts from rank to effective dimension,

$$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$7

and an oblivious sketch with

$$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$8

suffices for uniform error control on the regularized influence form (Hu et al., 11 Feb 2026).

The paper’s use of “factorized influence” is specific to Kronecker-factored curvature,

$$G_1(\omega,\mathbf{k})G_2(\omega,\mathbf{k})=\gamma G_c(\omega,\mathbf{k})$$9

with a factorized sketch

$R\in\mathbb{R}^{m\times n}$0

Then the sketched curvature becomes

$R\in\mathbb{R}^{m\times n}$1

and sketched gradients can be computed as $R\in\mathbb{R}^{m\times n}$2 without forming $R\in\mathbb{R}^{m\times n}$3 explicitly. Exact unregularized preservation requires factorwise injectivity on $R\in\mathbb{R}^{m\times n}$4 and $R\in\mathbb{R}^{m\times n}$5, hence $R\in\mathbb{R}^{m\times n}$6 and $R\in\mathbb{R}^{m\times n}$7. For $R\in\mathbb{R}^{m\times n}$8, guarantees depend on the factor-level effective dimensions

$R\in\mathbb{R}^{m\times n}$9

with

$\mathcal A$0

This yields an explicit computational–statistical trade-off: factorized sketches are cheaper in memory and matrix multiplication, but their total sketch size scales more expensively with $\mathcal A$1 than unfactorized sketches. The same theory also quantifies a leakage term when test gradients have components in $\mathcal A$2.

5. Flows, trajectories, and participant-invariant patterns

In representation learning, factorized influence can mean that each semantic transformation is represented by its own latent flow rather than by a scalar latent coordinate. Flow Factorized Representation Learning models sequences through

$\mathcal A$3

where $\mathcal A$4 selects a transformation type and each posterior transition is driven by a learned potential $\mathcal A$5,

$\mathcal A$6

The resulting latent probability paths are regularized toward dynamic optimal transport by a generalized Hamilton–Jacobi PDE,

$\mathcal A$7

implemented with a PINN loss. Factorization here means one flow field per factor of variation, with compositionality obtained either by switching transformation types mid-trajectory or by superposing latent velocities,

$\mathcal A$8

On MNIST with Scaling/Rotation/Coloring, supervised FFRL achieved the lowest reported equivariance errors, approximately $\mathcal A$9 versus approximately $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$00 for PoFlow, and the best log-likelihood, around $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$01; similar trends held on 3D Shapes, Falcor3D, and Isaac3D. The paper is explicit that this is factorized dynamics rather than axis-aligned scalar disentanglement (Song et al., 2023).

A more restrictive notion appears in temporal network analysis, where influence trajectories are assumed participant-invariant up to scale. Given a temporal network $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$02 and a per-node influence increment such as

$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$03

the aligned influence matrix $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$04 is factorized by rank-1 NMF as

$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$05

with $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$06 and $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$07. The interpretation is that every participant’s influence dynamics are a scalar multiple of the same temporal pattern $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$08. Across $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$09 temporal networks, all preprocessed into $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$10 equal-edge snapshots, the authors report Frobenius reconstruction error tolerance below $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$11 and show that random participant subsets recover nearly identical $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$12 under normalized $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$13, normalized $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$14, and cosine similarity comparisons. Dynamic Time Warping between normalized $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$15 vectors further suggests that $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$16 is domain-related, with citation networks exhibiting rapid decay and social networks longer tails (Shaojie, 2023).

These two uses of factorization are almost opposite in granularity. FFRL assigns one dynamical operator per factor of variation, whereas participant-invariant NMF collapses all participant heterogeneity in trajectory shape into a single shared temporal profile plus scalar amplitudes.

6. Exact algebraic factorizations beyond social data

In Bayesian-network inference, the relevant object is not social influence but functional dependence. If a child variable $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$17 is a deterministic function of parents $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$18, then its conditional probability table can be written as a hidden-variable factorization

$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$19

The hidden states correspond to hyperrectangles in the parent configuration space, and the central combinatorial problem is to find a minimal base of hyperrectangles from which each fiber $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$20 can be generated via proper difference and disjunctive union. Belief updating in the original and factorized models yields equivalent marginals. The computational value is that deterministic CPTs are replaced by two-dimensional potentials, which can reduce clique sizes in junction-tree propagation. In a computerized adaptive test on fraction operations, this factorization produced significantly smaller total clique sizes than both direct encoding and parent divorcing (Vomlel, 2012).

In coupled physical systems, factorized influence becomes an exact statement about dispersion. For any system composed of two coupled subsystems governed by a space-time homogeneous quadratic Lagrangian, the coupled dispersion relation admits the form

$$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$21

where $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$22 and $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$23 are the uncoupled subsystem dispersion functions and $$\lambda_{ik}=\gamma_k^\top \hat c_i+\alpha_i^\top \hat w_k+\sum_j a_{ij}x_{jk}\beta_j$$24 is a coupling function. The determinant expansion theorem underlying this identity is applied to examples including a traveling wave tube, an airplane wing with bending–torsion coupling, and the Mindlin–Reissner plate. In the latter case, the factorized form is used to analyze mode hybridization and avoided crossing: for nonzero coupling, all branches carry the imprint of both subsystem factors, but asymptotically recover pure uncoupled identities at large frequencies and wavenumbers. Near an uncoupled cross-point, linearization yields a universal hyperbolic geometry for the coupled branches (Figotin, 24 Mar 2026).

Taken together, these exact formulations show that factorized influence is not confined to latent-variable statistics. It can also denote a strict algebraic decomposition that exposes how local deterministic constraints or inter-subsystem coupling enter inference and spectral structure.

7. Common themes, assumptions, and points of divergence

Across the literature, factorized influence serves three recurrent purposes. First, it reduces dimensionality or parameter count: node-level factors replace edgewise probabilities in IMFB, low-rank latent states replace full rating matrices in dynamic recommendation, and Kronecker factors replace unfactored curvature blocks in scalable influence functions. Second, it separates mechanisms that would otherwise be confounded: PIF separates peer influence from homophily and item preference, while latent-flow models separate semantic transformations by assigning each one its own vector field. Third, it often improves computational tractability by converting large objects into products of smaller ones: this is explicit in matrix-free optimization for recommendation, factorized sketches for influence functions, and hidden-variable decompositions for deterministic CPTs.

The assumptions differ substantially. Dynamic recommendation assumes low-rank ratings, linear-Gaussian state dynamics, observed trust graphs, and fixed item factors during smoothing. PIF requires that substitute confounders capture the relevant latent traits affecting multiple interactions, and that the Poisson outcome model be correctly specified. IMFB assumes a bilinear edge model under Independent Cascade with edge-level feedback. The influence-function theory assumes access to a PSD curvature operator or a Kronecker-factored approximation thereof, and distinguishes sharply between unregularized and ridge-regularized regimes. Participant-invariant NMF assumes that all aligned trajectories are scalar multiples of one common pattern. In Bayesian networks and coupled physical systems, by contrast, the central factorizations are exact once the deterministic or Lagrangian structure is fixed.

This suggests that “factorized influence” is best understood as a cross-domain design pattern rather than a single method. The invariant feature is the relocation of influence from an unfactored observable relation to a structured intermediate representation—latent factors, substitute confounders, broadcaster/susceptibility embeddings, Kronecker factors, hyperrectangle states, shared temporal signatures, or subsystem dispersion factors. The substantive meaning of that relocation depends entirely on the surrounding theory.