Fractional Scoring Methods

• Fractional Scoring is a method that assigns score fractions based on precise quantile intervals, eliminating ambiguity at boundaries and ties.
• It employs interval arithmetic to distribute scores proportionally, ensuring the aggregate indicator exactly matches theoretical expectations.
• The approach extends to voting theory, offering continuous, clone-consistent outcomes through algorithmic tie handling and fractional allocation.

Fractional scoring is a set of formally exact, tie-invariant methods for attributing credit or score fractions to individual items—most prominently, scientific publications—in partitioned evaluation schemes such as percentile rank (PR) classes in bibliometrics or multi-candidate preference elections. Fractional scoring eliminates all ambiguity associated with discrete or integer-based class assignments, especially where ties or boundaries are present, and ensures that aggregate indicators exactly match theoretical expectations. The concept has been canonically developed for bibliometric percentile-rank evaluation by Schreiber and extended to related scoring problems in social choice theory and voting.

1. Mathematical Formulation and Definition

Let $N$ items (e.g., publications) be ranked by a measurable attribute (e.g., citation count) and partitioned into $K$ disjoint consecutive PR classes defined by boundaries $0 = p_0 < p_1 < \ldots < p_K = 1$. Each PR class $k$ covers the percentile interval $(p_{k-1}, p_k]$ and is assigned a weight $w_k$.

For each item $i$ (indexed $i=1,\ldots,N$ after sorting), assign the interval

$I_i = [ (i-1)/N, i/N )$

on the unit interval $[0,1]$. The fractional score $w(i)$ attributed to item $i$ is

$$w(i) = \sum_{k=1}^K w_k \cdot | I_i \cap (p_{k-1}, p_k] | \cdot N$$

where $| \cdot |$ denotes the Lebesgue measure (interval length). For intervals $I_i$ that straddle a boundary $p_j$, credit is split according to the fractional overlap. The aggregate indicator is $I_K = \sum_{i=1}^N w(i)$, and the relative indicator is $R = I_K / N$. This construction ensures that the total empirical score equals the theoretical expectation in the continuous limit:

$$I_\text{theory} = N \sum_{k=1}^K w_k (p_k - p_{k-1})$$

This formulation, by incorporating the full quantile interval per item and fractionalizing its class membership according to overlap, eliminates boundary and tie ambiguity (Schreiber, 2013, Schreiber, 2012).

2. Algorithmic Procedure and Tie Handling

The standard algorithm consists of the following steps (Schreiber, 2013, Schreiber, 2012):

1. Sort and Index: Sort all $N$ items by attribute (e.g., citation count), breaking ties with an arbitrary but fixed ordering.
2. Assign Percentile Intervals: Each item receives interval $I_i = [ (i-1)/N, i/N )$.
3. Define PR Class Boundaries: Select PR boundaries $p_k$ for $k=1,\ldots,K$.
4. Compute Overlaps: For each item, determine which PR classes its interval overlaps; for each class, compute the overlap $\ell_{i,k} = | I_i \cap (p_{k-1}, p_k] |$.
5. Fractional Attribution: For each overlap, assign the proportional score $w_k \cdot (\ell_{i,k} \cdot N)$ to item $i$.
6. Sum Indicators: Compute the total and relative indicators.

Handling ties: If several items share the same attribute value, they occupy contiguous indices, and their collective interval is the union $[ (a-1)/N, b/N )$ for indices $i=a,\ldots,b$. Thresholds that fall within this interval split the group's aggregate score fractionally; no item is assigned exclusively to either side of a boundary, and no tie-breaking rule is required. This property ensures tie-invariance and prevents arbitrary assignment at class borders (Schreiber, 2013, Schreiber, 2012, Leydesdorff, 2012).

3. Comparison with Integer and Alternative Percentile Schemes

Traditional integer-based assignment methods rank items and classify each wholly into a single PR class, computing a representative quantile (e.g., $j/N$ or $(j-0.5)/N$). This practice leads to issues:

• Ambiguity at boundaries (e.g., is an item at exactly 50% in the lower or upper half?).
• Instability with ties, as entire tied groups must be placed "en bloc" on one side of boundaries.
• Indicator jumps when small data changes shift items across boundaries.

Fractional scoring negates these issues by ensuring each item's quantile interval is honored and scores are smoothly distributed across class boundaries. As a result, the total indicator precisely matches the theoretical value, unlike conventional counting rules (Leydesdorff–Bornmann, Rousseau, Pudovkin–Garfield), which can deviate by up to 0.06 in relative indicators (Schreiber, 2013, Schreiber, 2012).

The PR₁₀₀ scheme proposed by Leydesdorff considers each item’s quantile interval as $[100 \cdot (R_i-1)/N,\, 100 \cdot R_i/N]$, with the midpoint as a representative score but allows for full fractional aggregation into arbitrary nonlinear bins (e.g., top 1%, top 5%, etc.) via interval arithmetic (Leydesdorff, 2012).

4. Empirical Validation and Worked Examples

Schreiber demonstrates these methods on four datasets comprising thousands of publications, using a $K=6$-class partition: bottom 50%, 50–75%, 75–90%, 90–95%, 95–99%, top 1%, with weights $w=\{1,2,3,4,5,6\}$. For each dataset, the fractional scoring sum $I_3(6)$ matches exactly the theoretical prediction $N \cdot 1.91$ (where $1.91$ derives from the chosen weights and boundaries), while all integer-based methods show deviations (Schreiber, 2013).

A detailed illustrative calculation: in a dataset with $N=2373$, at the 50% boundary ($p_1 = 0.50$), a typical interval $I_{1187}$ straddles the threshold, with the overlap distributed as $\ell_{1187,1}=0.0004$ and $\ell_{1187,2}=0.0008$. The resulting fractional score is $w(1187) = 0.95 + 3.80 = 4.75$, in contrast to an integer assignment of either 1 or 2 (Schreiber, 2013, Schreiber, 2012).

A summary table (K=6, PR-class scheme):

Class Interval	PR Class	Weight ($w_k$)
(0, 0.50]	Bottom 50%	1
(0.50, 0.75]	50–75%	2
(0.75, 0.90]	75–90%	3
(0.90, 0.95]	90–95%	4
(0.95, 0.99]	95–99%	5
(0.99, 1.00]	Top 1%	6

Each item receives fractional credit across intervals in which its quantile range overlaps.

5. Extensions to Voting Theory and Other Domains

Fractional scoring is not limited to bibliometrics. In preferential voting, the CLC + Zermelo scheme establishes a rigorous fractionalization of votes. Here, the CLC (Condorcet–Smith–Lackner Completion) projection replaces a paired-comparison matrix with a "nearby" matrix satisfying majoritarian and consistency properties, and the Zermelo strengths model recovers mixing fractions that sum to one, yielding continuous, clone-consistent, and locally stable social choice outcomes (Camps et al., 2010).

This extension formalizes the “fractional” nature of collective preference decisions, and ensures that plurality, majority, and tie situations are treated using continuous fractions rather than discrete blocks, eliminating common pathologies found in traditional methods.

6. Computational Complexity and Practical Considerations

The primary computational cost in fractional scoring arises from sorting the $N$ items, an $O(N \log N)$ operation. The computation of interval overlaps is $O(K \cdot$ number of tied blocks$)$, which is negligible even for large datasets (Schreiber, 2013). For the PR₁₀₀ scheme, only a single pass is required beyond sorting, making it feasible for datasets with tens of millions of items (Leydesdorff, 2012). In the voting context, the convex quadratic programming required for CLC projection and the fixed-point or Newton solver for Zermelo’s equations are both practical for moderate $N$ (Camps et al., 2010).

7. Limitations, Variants, and Directions for Further Research

Fractional scoring presumes a total order can be established, and computational workload increases with the product of the number of evaluation classes and the number of tied groups. The method is robust for linear indicators but would require modification for nonlinear scoring functions. Potential research directions include:

• Extensions to continuous or sliding-window percentile schemes
• Evaluation of ranking robustness for individuals or institutions under fractional vs. integer-based methods
• Adaptation for interdisciplinary or temporally partitioned reference sets
• Generalization to arbitrary, possibly non-disjoint, class definitions
• Software implementations for automated, robust fractional scoring in large-scale bibliometric or voting datasets

A plausible implication is that as evaluation systems increasingly demand granularity and fairness, fractional scoring offers a foundational mechanism for systematically eliminating ambiguity and enforcing theoretical consistency across domains (Schreiber, 2013, Schreiber, 2012, Leydesdorff, 2012, Camps et al., 2010).