GGscore: Scoring in Neural, Bayesian & Earthquake Analysis

• GGscore is a polysemous term representing three distinct scoring methods in neural recommendation, Bayesian Gaussian networks, and earthquake prediction.
• In neural recommendation, GGscore measures per-sample gradient alignment to evaluate replay value and mitigate catastrophic forgetting.
• In Bayesian network and earthquake analyses, GGscore serves as a marginal likelihood and risk measure, each tailored to its domain's inferential goals.

GGscore is a polysemous term that appears in multiple technical literatures with substantially different meanings. In neural recommendation and incremental learning, it denotes the Gradient–Gradient score, a per-sample gradient-alignment quantity used to estimate replay value under distribution shift and catastrophic forgetting (Shi et al., 9 Sep 2025). In Gaussian Bayesian network structure learning, it is commonly used as a shorthand for the Bayesian Gaussian equivalent score (BGe), a marginal-likelihood score for Gaussian directed acyclic graphical models under a conjugate normal–Wishart prior (Kuipers et al., 2014). In earthquake prediction analysis, the term is associated with a family of gambling-score $R$-characteristics that weight prediction outcomes by the difficulty of guessing rare target events (Molchan et al., 2010). These usages share a concern with scoring, but they differ in mathematical object, inferential target, and operational setting.

1. Terminological disambiguation

The most immediate technical fact about GGscore is that it does not designate a single universally standardized construct across fields. In the recommender-systems literature represented by "MEGG: Replay via Maximally Extreme GGscore in Incremental Learning for Neural Recommendation Models," GGscore stands for Gradient–Gradient score and is defined as an inner product between a per-sample gradient and a reference gradient direction (Shi et al., 9 Sep 2025). In the Gaussian DAG literature represented by "Addendum on the scoring of Gaussian directed acyclic graphical models," GGscore is used in the sense of the Gaussian DAG score, often called the Bayesian Gaussian equivalent score (BGe), which assigns a marginal likelihood to a Gaussian Bayesian network (Kuipers et al., 2014). In the earthquake-prediction literature represented by "Gambling scores in earthquake prediction analysis," the label is attached to R-characteristics derived from gambling-score constructions, where alarm outcomes are weighted by event probabilities $p_i$ (Molchan et al., 2010).

This divergence matters because the three usages score different entities. The Gradient–Gradient score evaluates the influence of an individual training sample on an optimization trajectory. The BGe score evaluates the posterior support for a graph structure under Gaussian assumptions. The gambling score evaluates the skill of a prediction sequence over alarms. A plausible implication is that references to GGscore are uninterpretable without domain context, since the same label can refer to optimization geometry, Bayesian model evidence, or weighted forecast verification.

2. Gradient–Gradient score in incremental neural recommendation

In MEGG, GGscore is defined for model parameters $\theta \in \mathbb{R}^p$, a training sample $z=(x,y)$, and a reference vector $\mathbf{V}\in\mathbb{R}^p$ by

$$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$

Here $g(z,\theta):=\nabla_\theta L(z,\theta)$ is the per-sample gradient, $g(D,\theta):=\nabla_\theta L(D,\theta)$ is the full-dataset gradient, and $L(D,\theta)=\frac{1}{|D|}\sum_{z\in D}L(z,\theta)$ is the training loss on the reservoir $D$ (Shi et al., 9 Sep 2025). The sign of the score is explicitly interpretable: $p_i$0 indicates alignment between the sample gradient and the reference direction, whereas $p_i$1 indicates opposition.

The default reference choice in MEGG is

$p_i$2

with $p_i$3 the converged parameters. The framework also uses a time-staggered variant in which the per-sample gradient is evaluated at penultimate-epoch parameters $p_i$4, while the reference vector remains the dataset gradient at convergence, $p_i$5 (Shi et al., 9 Sep 2025). The paper additionally reports alternative epoch-index choices, including $p_i$6 and $p_i$7.

The formal motivation is a first-order approximation to One Step Loss Change under Mini-Batch Gradient Descent. If $p_i$8 is the batch at step $p_i$9, then

$\theta \in \mathbb{R}^p$0

and, if one sample $\theta \in \mathbb{R}^p$1 is removed,

$\theta \in \mathbb{R}^p$2

Using a first-order Taylor approximation, the one-step loss change is approximated by

$\theta \in \mathbb{R}^p$3

and, after removing mini-batch stochasticity through a fixed reference, the paper gives the proportional relation

$\theta \in \mathbb{R}^p$4

On this basis, GGscore is presented as a scalable proxy for per-sample influence and as a first-order surrogate relative to classic influence functions, which would require second-order curvature through inverse-Hessian terms (Shi et al., 9 Sep 2025).

The score is described as model- and loss-agnostic. Typical losses listed are squared error for ratings,

$\theta \in \mathbb{R}^p$5

binary cross-entropy for implicit feedback,

$\theta \in \mathbb{R}^p$6

and pairwise BPR-style loss,

$\theta \in \mathbb{R}^p$7

Regardless of loss form, GGscore remains the same gradient dot product (Shi et al., 9 Sep 2025).

3. MEGG: bilateral extreme selection and replay

MEGG uses GGscore for experience replay based incremental learning in a domain-incremental recommendation setting with streaming interaction blocks $\theta \in \mathbb{R}^p$8, reservoir capacity $\theta \in \mathbb{R}^p$9, and incoming block size $z=(x,y)$0 (Shi et al., 9 Sep 2025). The central selection principle is maximally extreme GGscore: the method keeps both highly positive and highly negative samples rather than only one tail.

Given reservoir $z=(x,y)$1 and scores $z=(x,y)$2 for $z=(x,y)$3, the number of retained samples is $z=(x,y)$4. After sorting the scores non-decreasingly,

$z=(x,y)$5

the method keeps

$z=(x,y)$6

with left-tail subset

$z=(x,y)$7

right-tail subset

$z=(x,y)$8

and combined retained set

$z=(x,y)$9

If many samples share the same boundary score, ties are broken uniformly at random for stability (Shi et al., 9 Sep 2025).

The per-stage loop is specified as follows. The model $\mathbf{V}\in\mathbb{R}^p$0 is trained from scratch on $\mathbf{V}\in\mathbb{R}^p$1 for $\mathbf{V}\in\mathbb{R}^p$2 epochs; $\mathbf{V}\in\mathbb{R}^p$3 denotes the final parameters and $\mathbf{V}\in\mathbb{R}^p$4 the penultimate-epoch parameters. The reference vector is computed as $\mathbf{V}\in\mathbb{R}^p$5. For each $\mathbf{V}\in\mathbb{R}^p$6, the method computes $\mathbf{V}\in\mathbb{R}^p$7, selects bilateral extremes, and updates the reservoir by $\mathbf{V}\in\mathbb{R}^p$8 (Shi et al., 9 Sep 2025).

For efficiency, GGscore is not computed over all parameters. The paper restricts per-sample gradients to the user embedding $\mathbf{V}\in\mathbb{R}^p$9, item embedding $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$0, and final fully connected layer parameters $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$1, with all other gradients set to zero. The resulting score is

$$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$2

where the inner product is taken only over the selected parameter subset (Shi et al., 9 Sep 2025). The paper states that this exploits parameter locality in NCF-like recommenders, where a single interaction primarily affects $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$3, $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$4, and the output layer.

The rationale for bilateral replay is explicit. Samples in the right tail are described as “prototypical” for future training direction and help stabilize the learned signal; samples in the left tail are those most likely to be interfered with by upcoming updates, so replaying them counteracts forgetting and preserves decision boundaries and long-term preferences (Shi et al., 9 Sep 2025). This suggests that MEGG interprets support and interference as equally important retention signals in streaming recommendation.

The reported experimental setting covers MovieLens-1M, Douban Movie, LastFM-1k, and Taobao2014; model classes WDL, DCN, and NFM; and evaluation by AVG RMSE or AVG AUC across stages (Shi et al., 9 Sep 2025). The paper states that MEGG consistently outperforms iCaRL, MIR, GDumb, IncCTR, and SML across datasets and models, that it matches or surpasses Full-Batch in several settings, and that efficiency degrades only slightly with embedding dimension, with modest overhead for $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$5. Hyperparameter guidance includes $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$6 of the total dataset, $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$7, and four epoch-selection schemes A–D, with Scheme A as default: $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$8 The paper also lists limitations: the score is first-order, depends on parameter locality, and remains subject to stochastic batch effects, although the latter are reduced by using a global reference vector $$GG(z, \theta \mid \mathbf{V}) \;=\; \mathbf{V}^{\top} \nabla_{\theta} L(z, \theta).$$9 (Shi et al., 9 Sep 2025).

4. Bayesian Gaussian equivalent score for Gaussian DAGs

In Gaussian graphical-model structure learning, GGscore refers to the Gaussian DAG score, often called the Bayesian Gaussian equivalent score (BGe) (Kuipers et al., 2014). It assigns a marginal likelihood to a Gaussian Bayesian network under a conjugate normal–Wishart prior with global parameter independence and modularity. For a complete data sample $g(z,\theta):=\nabla_\theta L(z,\theta)$0 and a complete DAG $g(z,\theta):=\nabla_\theta L(z,\theta)$1, the score decomposes as

$g(z,\theta):=\nabla_\theta L(z,\theta)$2

where $g(z,\theta):=\nabla_\theta L(z,\theta)$3 is the parent set of node $g(z,\theta):=\nabla_\theta L(z,\theta)$4 and $g(z,\theta):=\nabla_\theta L(z,\theta)$5 denotes the data restricted to variables $g(z,\theta):=\nabla_\theta L(z,\theta)$6 (Kuipers et al., 2014).

The prior assumptions are a conditional Gaussian prior on the mean parameter and a Wishart