Weighted Kendall-τ Coefficient

• Weighted Kendall-τ is a generalization of the classic Kendall’s τ by incorporating nonnegative rank-based weights to emphasize top-ranked items.
• It offers both position-based and pairwise weighting schemes, enabling metric and pseudometric formulations with efficient O(n log n) computation.
• Applications span information retrieval, machine learning, and network analysis, with kernelized extensions enhancing ranking discrimination.

The weighted Kendall-$\tau$ coefficient is a family of ranking correlation measures that generalize the classical Kendall’s $\tau$ by introducing nonnegative weights to emphasize or de-emphasize specific rank positions or item pairs. These generalizations address the need to capture the differential importance of top-ranked items, a phenomenon prevalent in applications such as information retrieval, machine learning, and large-scale network analysis.

1. Definition and Foundational Formalisms

Let $a = (a_1, \dots, a_n)$ and $b = (b_1, \dots, b_n)$ be two rank-lists representing permutations of $\{1, \ldots, n\}$. The weighted Kendall-$\tau$ coefficient is defined as

$\tau_w(a, b) = \frac{\sum_{i \neq j} w_i w_j\, \sgn(a_j - a_i) \sgn(b_j - b_i)}{\sum_{i \neq j} w_i w_j}.$

Let $\mathcal{C}$ and $\mathcal{D}$ be the sets of concordant and discordant pairs, respectively; then equivalently,

$$\tau_w(a, b) = \frac{\sum_{(i,j)\in \mathcal{C}} w_i w_j - \sum_{(i,j)\in \mathcal{D}} w_i w_j}{\sum_{i \neq j} w_i w_j}.$$

Weighted Kendall-$\tau$ distances can also be defined using a strictly upper-triangular matrix of pairwise weights $w_{i,j}$: $$d_W(a, b) = \frac{\sum_{i<j} w_{i,j} \, \mathbf{1}_{(\pi_j - \pi_i)\,(\sigma_j - \sigma_i) < 0}}{\sum_{i<j} w_{i,j}},$$ where discordance is determined for each pair $(i, j)$. This construction admits both metric ($w_{i,j} > 0$ for all $i < j$) and pseudometric forms, with metric betweenness and pseudolinear quadruple properties established combinatorially (Piek et al., 2024).

2. Position-Based and Pairwise Weighting Schemes

Position-based weighting assigns to each position $i$ in a ranking a monotone decreasing “base” weight $f(i)$ and then combines $f(a_i)$, $f(b_i)$ for each item $i$ into a single $w_i$:

• Additive: $w_i = \frac{f(a_i) + f(b_i)}{2\sum_{j=1}^n f(j)}$
• Multiplicative: $$w_i = \frac{f(a_i)f(b_i)}{\sum_{j=1}^n f(a_j)f(b_j)}$$

Frequently used $f$ include harmonic decay $f(i) = 1/i$ and inverse-quadratic $f(i) = 1/(i+n_0)^2$, which yield top-heavy regimes where the upper ranks dominate the coefficient (Lombardo, 11 Apr 2025).

A more general scheme employs pairwise weighting $w_{i,j}$, such as positional $w_{i,j} = u_i u_j$ or distance-based weights $w_{i,j} = |i-j|^p$. Pairwise kernels $W((a, b), (c, d))$ further allow non-factorized, application-specific control (Jiao et al., 2018).

For datasets admitting ties, weights may be defined symmetrically $w(i,j) = w(j,i)$ and extended to the full calculation of joint-ties, left-ties, right-ties, concordant and discordant pairs, along with their respective total weights, as detailed in (Vigna, 2014).

3. Statistical Properties and Standardization

Unweighted Kendall-$\tau$ is symmetric: the distribution of the coefficient across all pairs of independent random rankings is centered at zero, establishing $E[\tau] = 0$. Introducing nonuniform or rank-dependent weights breaks this symmetry; the expected value for $\tau_w$ on random rankings becomes nonzero (typically positive unless special weighting symmetry is enforced), leading to spurious apparent concordance between random rankings (Lombardo, 11 Apr 2025). The mathematical basis is the absence of an involution $\pi \leftrightarrow \tilde{\pi}$ mapping $\tau_w(\pi)$ to $-\tau_w(\pi)$ when $w_i$ depend on $a_i, b_i$.

To remedy this, a standardization procedure is introduced: define a strictly monotonic, continuous, piecewise-quadratic shift-rescale function $g(x)$ so that $\Gamma_{\text{std}} = g(\Gamma)$ for a raw statistic $\Gamma$ (e.g., $\tau_w$) yields

• $E[g(\Gamma)] = 0$ on random lists,
• $g$ maps $[-1, 1] \rightarrow [-1, 1]$ with $g(\pm 1) = \pm 1$,
• strict monotonicity and preservation of the order of concordance values,
• unbiasedness under the null.

The construction involves moments of $\tau_w$’s distribution: $\bar{\Gamma} = E[\Gamma]$, variance $V$, left-segment variance $V^\ell$, and a uniquely determined $(g_0, g_1)$ parameter pair for $g(x)$, with explicit forms given for the “flat-variance-ratio” and general cases. The standardization exactly recovers the unweighted Kendall-$\tau$ ($g(x) = x$) when $w_i \equiv 1$. Numerical experiments verify centering and preservation of interpretive scale (Lombardo, 11 Apr 2025).

4. Computational Aspects

Naively, weighted Kendall-$\tau$ and its kernelized forms require $O(n^2)$ operations; however, for additive or multiplicative weight schemes, efficient $O(n \log n)$ algorithms generalize classical inversion counting:

• Merge-sort/divide-and-conquer techniques accumulate weighted discordances, maintaining residual sums for each recursion level (Vigna, 2014).
• For kernel versions, quicksort-style pivoting supports the efficient calculation of the weighted sum over concordant pairs, with recursive accumulator updates for low and high partitions (Jiao et al., 2018).

Pairwise and position-based weights with regular structure are especially amenable to these approaches, but arbitrary weight matrices may require full quadratic effort.

5. Kernelization and Machine Learning Applications

Weighted Kendall-$\tau$ coefficients admit positive-definite kernel generalizations. Given a positive-definite weight function $W$ on position pairs, the kernel is

$$K_W(\sigma, \sigma') = \sum_{1 \leq i \neq j \leq n} W((\sigma(i), \sigma(j)), (\sigma'(i), \sigma'(j))) \mathbf{1}_{\sigma(i) < \sigma(j)} \mathbf{1}_{\sigma'(i) < \sigma'(j)},$$

or, when $W$ factorizes via a rank-based $U_{a b}$,

$$K_U(\sigma, \sigma') = \sum_{1 \leq i \neq j \leq n} U_{\sigma(i), \sigma(j)} U_{\sigma'(i), \sigma'(j)} \mathbf{1}_{\sigma(i)<\sigma(j)} \mathbf{1}_{\sigma'(i)<\sigma(j)}.$$

These kernels are right-invariant under item relabeling and support feature maps into matrix- or tensor-valued spaces (Jiao et al., 2018).

Supervised learning of the weight matrix $U$ is supported, either by alternating optimization with SVM (or ridge SVM) objectives or by low-rank tensor factorization strategies (SUQUAN-style). This joint optimization focuses the coefficient on those pairwise or higher-order item sets most discriminative for the application in question.

Extensions to $d$-tuple (higher-order) kernels are theoretically well-posed but computationally feasible only for small $d$, due to $O(n^d)$ scaling.

6. Practical Guidelines for Use and Interpretation

To apply and interpret the standardized weighted Kendall-$\tau$:

1. Choose a position-based base weight $f(i)$ (e.g., $1/i$, $1/(i + n_0)^2$) and a combination rule (additive, multiplicative).
2. Calculate or reference precomputed values $(\bar{\Gamma}, V, V^\ell)$ for the chosen $n$ and weighting.
3. Compute the raw weighted Kendall-$\tau$ on $(a, b)$.
4. Apply the standardization function $g(x)$.
5. Report $g(\tau_w)\in [-1, 1]$.

Empirical studies using hyperbolic or quadratic decay confirm that top-heavy weighting reflects intuitive preferences for top-rank fidelity, correcting artifacts visible in the classical coefficient when applied, for instance, to large-scale network centrality comparisons (Vigna, 2014, Lombardo, 11 Apr 2025).

7. Geometric and Combinatorial Insights

Weighted Kendall-$\tau$ defines a pseudometric (or metric under strictly positive weights) on permutation space and inherits the salient structural properties of the unweighted case:

• Metric betweenness on the permutohedron holds under the weighted metric, mirroring the combinatorics of adjacent transpositions.
• Special quadruples (“pseudolinear quadruples”) realize characteristic distance patterns, extending the geometric analysis of classical permutation spaces (Piek et al., 2024).

This embedded geometric structure underpins both the interpretability of the weighted distance and its extension to higher-order permutation spaces.

---

References:

• "Standardization of Weighted Ranking Correlation Coefficients" (Lombardo, 11 Apr 2025)
• "On a weighted generalization of Kendall's tau distance" (Piek et al., 2024)
• "A Weighted Correlation Index for Rankings with Ties" (Vigna, 2014)
• "The Weighted Kendall and High-order Kernels for Permutations" (Jiao et al., 2018)