Sparse Variational GP-GAMs

• Sparse Variational GP-GAMs are a scalable Bayesian framework that combines Gaussian Processes with Generalized Additive Models to enable flexible, interpretable function decomposition.
• They employ inducing variables and sparse variational inference to significantly reduce computational complexity while accurately quantifying uncertainty.
• The framework uses stochastic optimization and structured posterior coupling to efficiently handle diverse likelihoods in regression, classification, and count data applications.

Sparse Variational Gaussian Process Generalized Additive Models (Sparse Variational GP-GAMs) provide a scalable Bayesian framework for learning flexible, interpretable additive function decompositions while rigorously quantifying uncertainty. These models merge the representational power of Gaussian Processes (GPs) with Generalized Additive Models (GAMs) and employ advanced variational inference to make computation tractable for large datasets through sparsity and structured posteriors (Adam, 2017, Adam et al., 2018).

1. Model Formulation and Additive Structure

A core attribute of Sparse Variational GP-GAMs is their additive structure. For inputs $x=(x_1,\ldots,x_D)$ and outputs $y$, the latent predictor is

$\eta(x)=\sum_{c=1}^C f_c(x_c)$

where each $f_c$ is a real-valued function with independent GP priors, $f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$ (Adam et al., 2018).

Observed responses may be modeled via a (possibly non-Gaussian) factorizing likelihood:

$$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$

For generalized additive modeling, link functions $g$ (identity, logit, log, etc.) are incorporated: $g(\mathbb{E}[y_n]) = \sum_{c=1}^C f_c(x_{n,c})$. This formulation accommodates regression and a broad family of exponential-family likelihoods (Adam et al., 2018).

2. Inducing Variable Framework and Sparse Approximation

To reduce the $O(N^3)$ complexity of GP inference, Sparse Variational GP-GAMs introduce $M$ inducing points per component:

$y$0

Stacking all components yields

$y$1

The prior on $y$2 is Gaussian with block-diagonal covariance, and the conditional process for all $y$3 given $y$4 is analytic (Adam et al., 2018, Adam, 2017).

This sparse approach enables scalable computation, as $y$5, providing near-linear scaling in $y$6 and cubic scaling in $y$7 per component (Adam, 2017).

3. Variational Posterior Parameterization and Posterior Coupling

Variational inference approximates the true posterior $y$8 via:

$y$9

Structured covariance $\eta(x)=\sum_{c=1}^C f_c(x_c)$0 can capture posterior dependencies among components, exceeding the expressive power of mean-field approximations. Marginalization of $\eta(x)=\sum_{c=1}^C f_c(x_c)$1 and calculation of predictive means/free variational parameters $\eta(x)=\sum_{c=1}^C f_c(x_c)$2 respect both inter- and intra-component dependencies (Adam, 2017).

The use of a single multivariate Gaussian for $\eta(x)=\sum_{c=1}^C f_c(x_c)$3 enables the representation of cross-component posterior coupling—critical for calibrated posterior variances and uncertainty quantification, especially when the posterior is not well-approximated by independent marginals (Adam, 2017, Adam et al., 2018). Imposing block-low-rank or diagonal structure in $\eta(x)=\sum_{c=1}^C f_c(x_c)$4 can reduce computational costs.

4. Evidence Lower Bound (ELBO) and Computational Complexity

The central variational objective is the Evidence Lower Bound (ELBO):

$\eta(x)=\sum_{c=1}^C f_c(x_c)$5

with terms that factorize efficiently owing to the additive model structure. The expectation is typically estimated by Monte Carlo sampling $\eta(x)=\sum_{c=1}^C f_c(x_c)$6, $\eta(x)=\sum_{c=1}^C f_c(x_c)$7.

The cost for an ELBO evaluation depends on the structure of $\eta(x)=\sum_{c=1}^C f_c(x_c)$8:

• Mean-field $\eta(x)=\sum_{c=1}^C f_c(x_c)$9: $f_c$0 per iteration.
• Full posterior coupling (full $f_c$1): $f_c$2 per iteration (Adam et al., 2018).
• KL calculations and conditional predictions further benefit from analytic tractability between Gaussians (Adam, 2017).

Complexity for prediction and sampling at batch size $f_c$3 is $f_c$4, with further improvements possible by exploiting low-rank or block structures.

5. Inference Algorithms and Optimization Procedures

Stochastic variational inference is used to optimize variational parameters $f_c$5, inducing inputs $f_c$6, and kernel hyperparameters $f_c$7. The canonical procedure consists of:

1. Sampling mini-batches of data $f_c$8.
2. Sampling $f_c$9, setting $f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$0.
3. Computing predictive means/variances for all $f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$1 for the batch.
4. Monte-Carlo or analytic evaluation of the expected log-likelihood.
5. Closed-form evaluation of the KL divergence.
6. Gradient estimation via automatic differentiation and parameter update (e.g., Adam, natural gradient) (Adam, 2017, Adam et al., 2018).

The following table summarizes major computational steps:

Step	Complexity (mean-field)	Complexity (full coupling)
Factorize/factor matrices	$f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$2	$f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$3
Predictive means/variances	$f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$4	$f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$5
KL divergence	$f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$6	$f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$7

6. Practical Implementation and Calibration

Efficient implementation recommendations include using automatic-differentiation frameworks (e.g., TensorFlow with GPflow, PyTorch with GPyTorch), placing inducing points $f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$8 via domain-covering heuristics (grids or k-means), and monitoring the ELBO and held-out log-likelihood for convergence (Adam et al., 2018). The sum of marginal variances across additive components provides the variance of $f_c\sim\mathcal{GP}(0, k_c(\cdot,\cdot))$9, forming credible intervals for the predictor and facilitating Bayesian model calibration.

Posterior variance and credible intervals are derived from the GP-marginal perspective, yielding well-calibrated uncertainty quantification that accounts for both model and approximation error, as the posterior approximation $$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$0 remains a valid Gaussian process (Adam et al., 2018).

7. Applications and Extensions

The sparse variational GP-GAM framework generalizes across regression (Gaussian noise) and non-Gaussian likelihoods (Binomial, Poisson, etc.), with appropriate link function integration:

• For regression: identity link.
• For count/Bernoulli data: logit/log link.

Each $$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$1 operates on the subspace $$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$2, so the kernel $$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$3 and inducing inputs $$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$4 are defined over $$p(y_{1:N}\mid f_{1:C}) = \prod_{n=1}^N p(y_n\mid\eta(x_n)).$$5. The same inference procedure extends directly to these tasks by changing the pointwise log-likelihood in the ELBO (Adam, 2017, Adam et al., 2018).

This framework is applicable in any scenario requiring interpretable, uncertainty-aware additive modeling for large datasets, including biostatistics, spatio-temporal modeling, and automatic feature-effect discovery.

---

References:

• (Adam, 2017) "Structured Variational Inference for Coupled Gaussian Processes"
• (Adam et al., 2018) "Scalable GAM using sparse variational Gaussian processes"