Z-Score Normalization

• Z-score normalization is a data standardization technique that centers values around zero with unit variance, supporting robust statistical analysis.
• Sliding-window z-score normalization computes moving mean and standard deviation to adapt to real-time data changes, improving applications like EMG motion prediction with notable accuracy gains.
• The method underpins time series clustering and scientometrics, where log-transformed normalization reduces bias and facilitates fair cross-field evaluations.

Z-score normalization, also known as standard-score normalization or standardization, is a transformation that centers and scales data by subtracting the mean and dividing by the standard deviation. This procedure produces normalized values with zero mean and unit variance. Z-score normalization is fundamental to numerous statistical and machine learning algorithms, enabling robust comparison, clustering, and classification across datasets with differing scales and distributions.

1. Mathematical Definition and Fundamental Properties

Given a sequence of observations $x = (x_1, x_2, \ldots, x_n)$, standard z-score normalization is defined by:

$z_i = \frac{x_i - \mu}{\sigma}$

where $\mu = \frac{1}{n}\sum_{i=1}^n x_i$ is the sample mean and $$\sigma = \sqrt{\frac{1}{n}\sum_{i=1}^n (x_i - \mu)^2}$$ is the sample standard deviation, each computed over the calibration dataset (Tanaka et al., 2022, Berthold et al., 2016). The transformation ensures that the distribution of $z$ is centered at zero and scaled to unit variance, facilitating algorithms that assume comparable ranges or distributions.

2. Sliding-Window Z-Score Normalization for Real-Time Applications

Sliding-window z-score normalization (SWN) extends the classical approach by dynamically computing $\mu$ and $\sigma$ over a rolling buffer of fixed length $L$:

$W_i = \{ x_{i-L+1}, ..., x_i \}$

$\mu_{w,i} = \frac{1}{L} \sum_{j=i-L+1}^i x_j$

$z_i = \frac{x_i - \mu}{\sigma}$0

$z_i = \frac{x_i - \mu}{\sigma}$1

This approach, as demonstrated for EMG-based motion prediction, eliminates the need for an offline calibration phase and adapts to slow shifts in signal baseline and amplitude (Tanaka et al., 2022). Empirical results show substantial improvements in classification accuracy for both within-user (+15.0%) and cross-user (+11.1%) motion prediction when replacing non-normalized inputs with SWN. Window length $z_i = \frac{x_i - \mu}{\sigma}$2 selection is robust within the 100–500 ms range; accuracy was largely insensitive to the precise choice in this interval.

3. Application in Time Series Analysis and Distance-Based Methods

Z-score normalization is pivotal for time series analysis, particularly in clustering and nearest-neighbor algorithms utilizing Euclidean distance. For any two time series $z_i = \frac{x_i - \mu}{\sigma}$3 and $z_i = \frac{x_i - \mu}{\sigma}$4, their z-scored forms $z_i = \frac{x_i - \mu}{\sigma}$5 satisfy:

$z_i = \frac{x_i - \mu}{\sigma}$6

where $z_i = \frac{x_i - \mu}{\sigma}$7 is the Pearson correlation coefficient computed on the normalized series (Berthold et al., 2016). Thus, average squared Euclidean distance between z-scored sequences is mathematically equivalent to a correlation-based dissimilarity. This property underpins the equivalence between Euclidean-based k-means clustering and “Pearson‐k‐means” with appropriate prototype (centroid) renormalization, though in practice omission of this renormalization produces negligible deviation relative to random initialization variability.

4. Z-Score Normalization in Field-Normalized Scientometrics

In scientometric analyses, especially cross-disciplinary citation benchmarking, citation distributions are highly right-skewed. Z-score normalization is effectively applied after a logarithmic transformation to produce field-neutral metrics:

$z_i = \frac{x_i - \mu}{\sigma}$8

where $z_i = \frac{x_i - \mu}{\sigma}$9 is the citation count for paper $\mu = \frac{1}{n}\sum_{i=1}^n x_i$0 in field $\mu = \frac{1}{n}\sum_{i=1}^n x_i$1, and $\mu = \frac{1}{n}\sum_{i=1}^n x_i$2 are the mean and standard deviation of $\mu = \frac{1}{n}\sum_{i=1}^n x_i$3 over all papers $\mu = \frac{1}{n}\sum_{i=1}^n x_i$4 in field $\mu = \frac{1}{n}\sum_{i=1}^n x_i$5 (Lu et al., 20 Apr 2025). Log transformation regularizes the distribution towards normality, enabling meaningful standardization. Empirical comparison using the Mahalanobis distance-based $\mu = \frac{1}{n}\sum_{i=1}^n x_i$6 metric shows that $\mu = \frac{1}{n}\sum_{i=1}^n x_i$7 substantially reduces cross-field ranking bias relative to alternatives such as raw ratios or source-side weighting (SNIP).

5. Implementation and Computational Considerations

For real-time applications (e.g., EMG motion prediction), SWN incurs minimal computational overhead (~18.6%), remaining well within critical time budgets (Tanaka et al., 2022). Initialization can involve zero-padding or accumulating early samples for the normalization buffer; subsequent updates merely shift the window by adding new data and removing the oldest sample. For field-normalized scientometric pipelines, the procedure involves grouping entities by field, computing transformation and normalization per field, and ranking or evaluating metrics globally (Lu et al., 20 Apr 2025).

Example Table: Z-Score Normalization Use Cases

Context	Formula/Approach	Key Outcome
Time series clustering	$\mu = \frac{1}{n}\sum_{i=1}^n x_i$8	Pearson-equivalent
EMG-based motion prediction (SWN)	$\mu = \frac{1}{n}\sum_{i=1}^n x_i$9	Increased accuracy
Field-normalized citation analysis	$$\sigma = \sqrt{\frac{1}{n}\sum_{i=1}^n (x_i - \mu)^2}$$0	Reduced bias

6. Hyperparameter Selection and Methodological Guidelines

For SWN, window lengths of 100–500 ms are recommended; feature-extraction windows in the same range generally produce optimal accuracy when combined with normalization. All common EMG features benefit similarly, and no additional window overlap is necessary. In citation normalization, field assignment and accurate computation of within-field $$\sigma = \sqrt{\frac{1}{n}\sum_{i=1}^n (x_i - \mu)^2}$$1 and $$\sigma = \sqrt{\frac{1}{n}\sum_{i=1}^n (x_i - \mu)^2}$$2 are critical for robust cross-field comparability. When combining source-side and target-side normalization (e.g., SNIP weighting followed by log + z), further reduction in field bias is achieved (Lu et al., 20 Apr 2025).

7. Limitations, Extensions, and Future Directions

Standard z-score normalization presumes stationarity of the sample distribution; shifting baselines can invalidate the calibration reference, motivating methods like SWN for nonstationary data streams. In time series clustering, prototype renormalization maintains strict equivalence with correlation-based metrics; omission introduces minor deviations. Peak classification accuracy in EMG motion prediction using SWN does not reach the optimal within-user performance in cross-user models, suggesting the need for further integration of transfer-learning or domain-adaptation approaches. In scientometrics, combining log-transformed z-score normalization with source-side weighting yields the lowest measured cross-field bias; yet, uneven paper growth rates and citation overlaps remain challenges for absolute neutrality.

In summary, z-score normalization and its extensions, such as SWN and log-transformed field normalization, offer theoretically sound and empirically validated frameworks for standardizing data, enabling fair comparison, improving robustness, and supporting advanced analytics in diverse domains including time series mining, real-time biomedical sensing, and evaluative bibliometrics.