Fusion Sampling Validation (FSV)

• Fusion Sampling Validation (FSV) is a hybrid estimator that combines Simple Random Sampling and K-Fold Cross-Validation to achieve low bias and variance in performance estimates.
• The method fuses SRS and KFCV estimates using a weighted fusion factor and optional scaling, balancing representativeness and robustness while reducing computational load.
• Empirical evaluations show that FSV achieves approximately 5% lower variance and mean squared error compared to traditional SRS and KFCV, making it ideal for large datasets and resource-constrained environments.

Searching arXiv for the specified FSV paper and adjacent uses of the term. Fusion Sampling Validation (FSV) is a hybrid data-partitioning technique that blends the strengths of Simple Random Sampling (SRS) and K-Fold Cross-Validation (KFCV) to reduce estimation bias and variance while keeping computational costs manageable. In the formulation introduced in “Fusion Sampling Validation in Data Partitioning for Machine Learning” (Udomboso et al., 2 Aug 2025), FSV combines an SRS-based estimate with a KFCV estimate through a weighted fusion factor and, optionally, a scaling factor on the KFCV component. The resulting framework is presented as a low-bias, low-variance validation procedure intended for effective machine learning implementations, particularly on extensive datasets and in resource-constrained environments (Udomboso et al., 2 Aug 2025).

1. Definition and conceptual basis

FSV is defined as a hybrid model combining SRS and KFCV to optimise data partitioning (Udomboso et al., 2 Aug 2025). Its stated objective is to minimise biases and merge the simplicity of SRS with the accuracy of KFCV. The technique is motivated by two familiar limitations: SRS provides representative samples but can result in non-representative sets with imbalanced data, whereas KFCV enhances model robustness but often compromises generalisation assessment due to high computational demands and extensive data shuffling (Udomboso et al., 2 Aug 2025).

Within the underlying exposition, SRS is defined on a dataset $D$ of size $N$ by drawing a subset $S$ of size $n$ so that each point $x_k \in D$ has equal inclusion probability $p_k = 1/N$ (Udomboso et al., 2 Aug 2025). KFCV is defined by partitioning a sample $S$ into $K$ roughly equal folds $\{S_1,\dots,S_K\}$, training on $S \setminus S_f$ and validating on $N$0 for each fold, then averaging performance (Udomboso et al., 2 Aug 2025). FSV is positioned between these two procedures: it inherits the simplicity and sampling character of SRS while using cross-validation structure to stabilise estimation.

The paper’s discussion attributes the reduction in bias and variance to complementary roles for the two components: the SRS component ensures representativeness, and the KFCV component stabilises outliers and folds’ variability (Udomboso et al., 2 Aug 2025). This suggests that FSV is best understood not as a replacement for either classical method in isolation, but as a weighted estimator intended to trade off statistical stability against computational overhead.

2. Mathematical formulation

The formalism of FSV introduces the notation

• $N$1 for the size of the full data $N$2,
• $N$3 for the sample size per iteration,
• $N$4 for the number of random-sampling iterations,
• $N$5 for the number of folds in KFCV,
• $N$6 for the fusion weight on the SRS estimate,
• $N$7 for the scaling factor on the KFCV estimate (Udomboso et al., 2 Aug 2025).

For each data index $N$8, the SRS selection probability is given by

$N$9

For the fold assignment, the indicator $S$0 is defined so that $S$1 if data point $S$2 is assigned to fold $S$3, and $S$4 otherwise, with

$S$5

The fusion rule itself combines an SRS-based estimate $S$6 and a KFCV estimate $S$7 according to

$S$8

In the paper, the fusion weight is also denoted $S$9 (Udomboso et al., 2 Aug 2025).

The formulation is explicitly weighted rather than hierarchical. FSV does not merely perform SRS followed by KFCV; it produces a fused estimate in which both components contribute quantitatively. The optional scaling factor $n$0 modulates the KFCV term, and the exposition notes that $n$1 is often set to $n$2 (Udomboso et al., 2 Aug 2025). In interpretive terms, this makes FSV a parametric interpolation between pure SRS and a KFCV-driven estimate, though the paper’s emphasis remains on empirical balance rather than on asymptotic optimality proofs.

3. Algorithmic workflow

The algorithmic description given for Fusion Sampling Validation takes as input the full dataset $n$3, sample size $n$4, number of folds $n$5, iterations $n$6, fusion weight $n$7, and scale $n$8, and returns a hybrid performance estimate $n$9 (Udomboso et al., 2 Aug 2025).

The procedure is specified as follows. First, a cumulative estimate $x_k \in D$0 is initialised to $x_k \in D$1. For each iteration $x_k \in D$2, an SRS sample $x_k \in D$3 of size $x_k \in D$4 is drawn from $x_k \in D$5. An SRS estimate $x_k \in D$6 is then computed on $x_k \in D$7. The sample $x_k \in D$8 is partitioned into $x_k \in D$9 folds $p_k = 1/N$0, after which KFCV is performed to obtain

$p_k = 1/N$1

The fused estimate for that iteration is

$p_k = 1/N$2

which is accumulated into $p_k = 1/N$3. After $p_k = 1/N$4 iterations, the final estimate is returned as

$p_k = 1/N$5

This workflow is notable for nesting KFCV inside repeated random sampling rather than applying KFCV directly to the entire dataset (Udomboso et al., 2 Aug 2025). A plausible implication is that the method seeks to moderate the data-shuffling and training burden associated with repeated full-dataset cross-validation while preserving some of its error-estimation benefits.

The computational discussion in the same source gives the total cost approximately as

$p_k = 1/N$6

and states that with moderate $p_k = 1/N$7 and $p_k = 1/N$8, FSV demands approximately $p_k = 1/N$9 model trainings (Udomboso et al., 2 Aug 2025). The same discussion characterises this as dramatically lower than pure KFCV with high $S$0 or nested CV on large $S$1.

4. Evaluation framework and reported metrics

The empirical study evaluates FSV using a fixed set of performance criteria. Let $S$2 denote the performance measure, such as loss, at iteration $S$3. The reported metrics are the mean estimate (ME), variance estimate (VE), mean squared error (MSE), bias, the rate of convergence for mean estimates (ROC_ME), and the rate of convergence for variance estimates (ROC_VE) (Udomboso et al., 2 Aug 2025).

The metric definitions are given as

$S$4

$S$5

$S$6

where $S$7 is the true population loss, and

$S$8

The convergence quantities are described as

$S$9

measured as change in ME per additional iteration, and

$K$0

measured as change in VE per iteration (Udomboso et al., 2 Aug 2025).

The paper’s evaluation protocol uses three datasets of size $K$1 drawn i.i.d. from $K$2, with random seed $K$3, updated across trials for reproducibility (Udomboso et al., 2 Aug 2025). KFCV is performed with $K$4 folds and $K$5 repetitions. For FSV, the reported settings are $K$6, $K$7, and $K$8, with the weight described as empirically found to balance bias-variance (Udomboso et al., 2 Aug 2025).

5. Empirical validation

The study reports aggregated performance across all dataset sizes and iteration settings for SRS, KFCV, and FSV (Udomboso et al., 2 Aug 2025).

Method	ME	VE	MSE
SRS	0.00217	0.99912	1.00134
KFCV	0.00183	0.99927	1.00095
FSV	0.000863	0.949644	0.952127

Method	bias	ROC_ME	ROC_VE
SRS	0.01538	0.0078	0.0112
KFCV	0.01492	0.0062	0.0087
FSV	0.016288	0.005199	0.007137

The results section states that across all sizes and repetitions, FSV achieved the lowest variance and MSE—roughly $K$9 lower than SRS and KFCV (Udomboso et al., 2 Aug 2025). It further states that the bias of FSV remained minimal and comparable to KFCV, and that convergence rates for ROC_ME and ROC_VE outperformed both baselines, indicating faster stability with fewer iterations (Udomboso et al., 2 Aug 2025).

The abstract reports the same FSV summary values: ME of $\{S_1,\dots,S_K\}$0, VE of $\{S_1,\dots,S_K\}$1, MSE of $\{S_1,\dots,S_K\}$2, bias of $\{S_1,\dots,S_K\}$3, ROC_ME of $\{S_1,\dots,S_K\}$4, and ROC_VE of $\{S_1,\dots,S_K\}$5 (Udomboso et al., 2 Aug 2025). On the basis of these reported aggregates, the paper concludes that FSV consistently outperformed SRS and KFCV and demonstrated superior accuracy and reliability in data partitioning, particularly in resource-constrained environments and extensive datasets (Udomboso et al., 2 Aug 2025).

Because the experiments were conducted on synthetic datasets drawn from a normal distribution with fixed sizes and specified seeds, the reported advantage is directly tied to that experimental design. This suggests that the paper’s empirical claim is strongest as a validation of the proposed estimator under controlled sampling conditions rather than as a universal dominance result across arbitrary learning tasks.

6. Hyperparameters, operating regimes, and practical guidance

The paper provides explicit guidance for choosing FSV hyperparameters (Udomboso et al., 2 Aug 2025). For the fusion weight $\{S_1,\dots,S_K\}$6, it recommends values in the range $\{S_1,\dots,S_K\}$7–$\{S_1,\dots,S_K\}$8 based on the desired bias/variance trade-off; higher $\{S_1,\dots,S_K\}$9 favours SRS speed, and lower $S \setminus S_f$0 favours variance reduction. For the scaling factor $S \setminus S_f$1, the recommendation is typically $S \setminus S_f$2, with tuning reserved for cases in which over- or under-weighting of KFCV is needed. For $S \setminus S_f$3, the discussion states that $S \setminus S_f$4 or $S \setminus S_f$5 folds are standard; more folds reduce bias but increase cost. For $S \setminus S_f$6, the paper states that $S \setminus S_f$7–$S \setminus S_f$8 iterations yield diminishing returns beyond approximately $S \setminus S_f$9 (Udomboso et al., 2 Aug 2025).

Practical recommendations are differentiated by resource regime. For resource-constrained settings, the paper recommends using $N$00–$N$01, $N$02, and $N$03 to halve variance at approximately $N$04 training cost (Udomboso et al., 2 Aug 2025). For very large $N$05 or imbalanced data, it suggests considering stratified SRS within FSV to preserve class proportions. It also lists possible extensions: incorporating Monte Carlo CV in place of KFCV, adaptive weighting $N$06 based on iteration variance, and plugging in different sample sizes $N$07 to further reduce cost (Udomboso et al., 2 Aug 2025).

These recommendations indicate that FSV is presented not only as a fixed estimator but as a tunable validation framework. A plausible implication is that its practical identity depends substantially on how aggressively one weights the SRS and KFCV components, so that different parameterisations may occupy different points on an efficiency–stability frontier.

7. Related uses of “FSV” and broader significance

The acronym “FSV” also appears in a distinct sense in the view-synthesis literature summarised in “Sampling for View Synthesis: From Local Light Field Fusion to Neural Radiance Fields and Beyond” (Ramamoorthi, 2024). There, “Fusion Sampling Validation (FSV) guidelines” refer to a validation protocol tied to Local Light Field Fusion (LLFF), including varying adjacent-view disparity, matching MPI depth $N$08 to $N$09, and measuring perceptual error against a Nyquist reference (Ramamoorthi, 2024). The same source states that “the FSV protocol established by LLFF … remains a blueprint for assessing the sampling-error trade-off of any IBR or neural-rendering algorithm” (Ramamoorthi, 2024).

This usage is conceptually different from the data-partitioning method introduced in (Udomboso et al., 2 Aug 2025). In (Ramamoorthi, 2024), FSV refers to a sampling-validation protocol for image-based rendering and neural scene representations; in (Udomboso et al., 2 Aug 2025), it denotes a hybrid estimator for machine-learning data partitioning. The shared acronym therefore does not imply a shared algorithmic lineage.

A broader pattern, however, is visible across recent work that employs “fusion” in sampling design. For example, “Efficient Global-Local Fusion Sampling for Physics-Informed Neural Networks” (Luo et al., 28 Oct 2025) presents a Global–Local Fusion (GLF) sampling strategy that combines a global residual-based distribution with local residual-driven perturbations in order to preserve stability while retaining efficiency. That work concerns collocation-point placement in PINNs rather than validation or data partitioning, but it likewise frames fusion as a way to combine complementary sampling regimes (Luo et al., 28 Oct 2025). This suggests a wider methodological tendency in which hybrid sampling schemes are used to balance robustness, computational cost, and task-specific error control.

Within that broader context, Fusion Sampling Validation in the sense of (Udomboso et al., 2 Aug 2025) occupies a specific niche: it is a validation-oriented fusion method for partitioning and performance estimation, grounded in repeated SRS, embedded KFCV, and explicit bias–variance metrics.