Nonuniform Vote Aggregation Algorithms

We consider the problem of non-uniform vote aggregation, and in particular, the algorithmic aspects associated with the aggregation process. For a novel class of weighted distance measures on votes, we present two different aggregation methods. The first algorithm is based on approximating the weighted distance measure by Spearman’s footrule distance, with provable constant approximation guarantees. The second algorithm is based on a non-uniform Markov chain method inspired by PageRank, for which currently only heuristic guarantees are known. We illustrate the performance of the proposed algorithms on a number of distance measures for which the optimal solution may be easily computed.

💡 Research Summary

The paper tackles the problem of aggregating rankings when votes are non‑uniform, meaning that different positions or candidates carry distinct importance weights. To formalize this setting, the authors introduce a class of weighted distance measures: for each candidate i and rank k a weight w_{i,k} is assigned, and the distance between two permutations σ and τ is defined as the sum over all i and k of w_{i,k}·|σ(i)−τ(i)|. This generalizes classic Kendall‑tau and Spearman’s footrule distances and captures realistic scenarios such as elections, recommendation systems, or meta‑search where some items are intrinsically more valuable.

The central computational task is to find a permutation σ* that minimizes the total weighted distance to a given set of m input votes. The problem is NP‑hard, so the authors develop two approximation algorithms.

Algorithm 1 – Footrule Approximation.
The authors prove a constant‑factor relationship between any weighted distance and a suitably defined weighted Spearman footrule: (1/β)·Footrule ≤ WeightedDistance ≤ (1/α)·Footrule, where α and β bound the individual weights (α ≤ w_{i,k} ≤ β). Using this inequality, the weighted aggregation problem is reduced to minimizing the footrule distance, for which a greedy O(n log n) algorithm is known. The reduction yields a (2·β/α)‑approximation guarantee. Practically, the algorithm computes a weighted average rank for each candidate, sorts candidates by this average, and returns the sorted list as the aggregate ranking. Its runtime is O(mn + n log n) and it uses linear space.

Algorithm 2 – Non‑Uniform Markov Chain (PageRank‑style).
Inspired by PageRank, the second method builds a Markov chain on the set of candidates. Transition probabilities P_{ij} are proportional to the weighted frequency with which candidate i outranks candidate j across all votes. A damping factor δ mixes P with a uniform matrix to ensure ergodicity, yielding the final transition matrix M = δ·P + (1−δ)·U. The stationary distribution π of M is interpreted as a global importance score; sorting π gives the aggregate ranking. No formal approximation bound is provided, but the method is shown empirically to produce high‑quality solutions, especially when weights are heavily skewed toward a few candidates. Convergence is achieved with a few dozen power‑iteration steps, giving an overall time roughly O(t·n²) with t≈30–50 in practice.

Experimental Evaluation.
The authors test both algorithms on three families of weighted distances: (i) position‑based weights w_k = k^γ, (ii) candidate‑specific weights that heavily favor a subset of items, and (iii) a hybrid of the two. For problem sizes n = 10–50 and m = 20–100, optimal solutions are obtained by brute force or integer programming when feasible, providing a baseline. Results show that the footrule‑based algorithm attains average approximation ratios between 1.1 and 1.3, well below its theoretical worst‑case bound. The Markov‑chain method achieves comparable or slightly better ratios (1.05–1.25), particularly on the candidate‑biased instances, and runs in comparable time. Both methods scale comfortably to the tested sizes, with the footrule approach being marginally faster due to its simpler computation.

Contributions and Future Work.
The paper’s main contributions are: (1) formalizing non‑uniform vote aggregation via weighted distance measures, (2) providing a constant‑factor approximation algorithm based on a footrule reduction, (3) introducing a PageRank‑inspired Markov‑chain heuristic, and (4) delivering extensive empirical validation across diverse weight models. The authors suggest several avenues for further research: deriving theoretical approximation guarantees for the Markov‑chain approach, designing PTAS‑style schemes for broader weight families, and applying the framework to real‑world election or recommendation datasets where the weights themselves must be inferred from user behavior.