RankSky: Spectral Skyline Ranking

• RankSky is a spectral skyline ranking method that orders Pareto-optimal tuples using a PageRank-inspired stationary distribution computed on a transformed skyline matrix.
• It constructs a row-stochastic matrix from the inner-product interactions of skyline attribute vectors and applies the IPL algorithm for convergence.
• RankSky overcomes limitations of methods like dp-idp by ensuring distinctive rankings and improved efficiency, as evidenced by empirical evaluations on varied datasets.

RankSky is a skyline-query ranking method that orders Pareto-optimal tuples when a skyline result set is too large to be directly actionable. It was introduced as part of a comparative study of skyline-ranking methods and is explicitly presented as an adaptation of PageRank to skyline points, using a square stochastic matrix, a teleportation matrix, a damping factor, a row score eigenvector, and the IPL algorithm. Rather than ranking tuples by additional pairwise dominance among skyline points, RankSky derives a stationary importance distribution from a matrix built from the skyline tuples themselves after all preferences have been unified to $MAX$ (Martin-Nevot et al., 29 Jul 2025).

1. Problem setting and motivation

RankSky is defined in the setting of skyline queries over a relation $r$ with attributes $A_1,\dots,A_n$. For two tuples $t$ and $t'$, the dominance relation is written $t \prec_d t'$ when $t$ is at least as good as $t'$ on every criterion and strictly better on at least one, with the comparison direction determined by the preference on each attribute: $$(\preceq_d, \prec_d) = \left\lbrace \begin{array}{l} (\le, <) \equiv Pref(A_j) = MIN \ (\ge, >) \equiv Pref(A_j) = MAX \end{array} \right.$$ A skyline point is then a tuple in the skyline set $S$, meaning that no other tuple dominates it. The motivation for additional ranking is that skyline output can have high cardinality, especially when the dataset is large, the number of dimensions is high, or the dimensions are weakly correlated; the cited study remarks that in such cases “more Skyline points often lead to less differences between them,” making the raw skyline difficult to exploit (Martin-Nevot et al., 29 Jul 2025).

Within that context, RankSky was proposed against the background of earlier skyline-ranking approaches, especially dp-idp. The paper characterizes dp-idp as a skyline ranking method inspired by tf-idf, where a skyline point is scored from the dominated points beneath it by combining a dominance-power term and an inverse-dominance-power term. Two limitations are emphasized. First, dp-idp has efficiency problems because it repeatedly evaluates dominance layers and computes scores for all skyline points even if only top-$r$0 is desired. Second, it does not ensure a distinctive rank: in the paper’s Pokémon example, skyline points with RowId 2 and 4 both obtain score $r$1, so dp-idp cannot distinguish them. RankSky is introduced specifically because dp-idp “does not ensure a distinctive rank” (Martin-Nevot et al., 29 Jul 2025).

2. Mathematical construction

The objects ranked by RankSky are the skyline points themselves. If the skyline contains $r$2 tuples, RankSky constructs an $r$3 matrix. Before that construction, all preferences must be unified to $r$4; for attributes with $r$5 preference, the values are inverted so that the skyline matrix can be interpreted uniformly in a maximization setting. Let $r$6 denote the skyline matrix after this preference unification (Martin-Nevot et al., 29 Jul 2025).

The central matrix is

$r$7

This square matrix does not encode direct dominance among skyline points. The paper is explicit that skyline points are mutually non-dominating by definition, so direct skyline-to-skyline dominance is unavailable as a ranking structure. Instead, the induced relation comes from the matrix product $r$8, whose entries capture inner-product interactions between skyline tuples. The result is an implicit weighted complete graph over skyline points, not a dominance graph (Martin-Nevot et al., 29 Jul 2025).

RankSky then row-normalizes $r$9 into a row-stochastic matrix $A_1,\dots,A_n$0, with $A_1,\dots,A_n$1 and $A_1,\dots,A_n$2. A teleportation matrix $A_1,\dots,A_n$3 is defined as the all-ones matrix of size $A_1,\dots,A_n$4, and the damping factor is fixed to the standard PageRank value

$A_1,\dots,A_n$5

The final transition matrix is

$A_1,\dots,A_n$6

This damped construction is used to ensure irreducibility and aperiodicity, yielding existence and uniqueness of the stationary distribution (Martin-Nevot et al., 29 Jul 2025).

The ranking vector is a row score vector $A_1,\dots,A_n$7 satisfying

$A_1,\dots,A_n$8

Its components are interpreted as normalized importance weights over skyline points. The paper stresses that RankSky does not use dominance relations directly, does not use logarithms, and relies instead on matrix construction, stochastic normalization, teleportation, and eigenvector computation (Martin-Nevot et al., 29 Jul 2025).

3. Algorithmic procedure and computational profile

The RankSky pipeline begins by computing the skyline $A_1,\dots,A_n$9 of the input relation $t$0 using BBS (Branch-and-Bound Skyline). The skyline matrix $t$1 is then constructed, with one row per skyline point and one column per dimension, after all preferences have been unified to $t$2. The transpose $t$3 is formed, and the square matrix $t$4 is computed. Row-wise normalization yields the stochastic matrix $t$5, after which the Google-style matrix $t$6 is assembled with $t$7 and $t$8 (Martin-Nevot et al., 29 Jul 2025).

Score computation uses IPL. The initial vector is uniform,

$t$9

and iteration proceeds according to

$t'$0

until convergence under an $t'$1-norm stopping rule. The precision parameter is $t'$2, hence

$t'$3

The paper’s algorithmic presentation uses an auxiliary vector $t'$4, normalizes after each multiplication, and stops when the norm difference is below $t'$5. Final scores are attached to skyline points and sorted in descending order to produce the ordered score table (Martin-Nevot et al., 29 Jul 2025).

The paper gives RankSky’s time complexity as

$t'$6

This combines skyline extraction using BBS, matrix-related preprocessing, and iterative multiplication over the $t'$7 transition matrix. The memory footprint is correspondingly dominated by dense quadratic structures: $t'$8, $t'$9, $t \prec_d t'$0, and $t \prec_d t'$1 are each $t \prec_d t'$2, while $t \prec_d t'$3 and $t \prec_d t'$4 contribute rectangular skyline-by-dimension storage. The authors explicitly state that RankSky is not embeddable in a relational DBMS, in contrast to CoSky (Martin-Nevot et al., 29 Jul 2025).

4. Position within skyline-ranking research

In survey treatments of skyline ranking, the broader problem is framed as a compromise between classic top-$t \prec_d t'$5 and skyline querying. Top-$t \prec_d t'$6 requires a scoring function, often a weighted sum, whereas skyline returns all non-dominated tuples without ranking them. Skyline ranking is presented as a hybrid approach intended to extract interesting skyline points and rank them without demanding a precise user utility function; in that literature, importance is often associated with dominance behavior across subspaces and recursive importance propagation in a skyline graph (Fabris, 2022).

That broader branch of work is distinct from RankSky’s specific mechanism. Survey papers identify skyline ordering/ranking as a family that includes Top-K Dominating, SKYRANK, size-constrained skyline, and related methods that rank skyline tuples through dominance counts, subspace dominance, or skyline layers (Cuzzucoli, 2022). Another survey presents flexible skyline methods such as $t \prec_d t'$7-skyline, F-skyline, UTK, and ORD/ORU as intermediate compromises between skyline and ranking, emphasizing relaxed preference input, generalized dominance, and output-size control (Tonnarelli, 2022).

RankSky belongs to this general landscape of skyline ordering, but its construction differs sharply from subspace-dominance or F-dominance schemes. The 2025 paper states that RankSky ignores dominance in the ranking phase and instead applies a PageRank-style stationary-distribution computation to a matrix induced by skyline attribute vectors. This suggests that RankSky is best understood not as a dominance-count method, nor as a restricted-skyline method, but as a spectral skyline-ranking method whose ranking signal is derived from the geometry of the skyline matrix rather than from dominance edges or explicit user preference regions.

5. Experimental evaluation and integration in DeepSky

The evaluation reported for RankSky was performed on synthetic datasets with independent dimensions, cardinalities ranging from 10 to 1 billion tuples, and dimensionalities 3, 6, and 9. For runtime comparison, the main baselines were SkyIR-UBS from the dp-idp family, an improved dp-idp with dominance hierarchy, RankSky itself, and CoSky in both algorithmic and SQL forms (Martin-Nevot et al., 29 Jul 2025).

Method	Worst-case time at 50,000 tuples, 3 dimensions	Note
RankSky	1 min 46 s	PageRank-style skyline ranking
Algorithmic CoSky	2 min 16 s	TOPSIS/cosine-to-ideal-point
Improved dp-idp	< 3 hours	Dominance-hierarchy variant
SkyIR-UBS	> 6 hours	dp-idp-family baseline

These results are presented as the strongest empirical argument for RankSky: it is much faster than dp-idp-family methods and slightly faster than algorithmic CoSky, though slower than SQL CoSky. For datasets up to 200,000 tuples, the paper reports worst-case times of about 34 minutes in 3 dimensions, about 44 minutes in 6 dimensions, and about 1 h 20 min in 9 dimensions. Algorithmic CoSky is slightly slower in each of those settings, whereas SQL CoSky remains much faster (Martin-Nevot et al., 29 Jul 2025).

The paper also reports a ranking-behavior example on a Pokémon skyline. RankSky yields distinct scores for the three skyline points shown:

• RowId 1: $t \prec_d t'$8
• RowId 4: $t \prec_d t'$9
• RowId 2: $t$0

This differentiates points 2 and 4, which dp-idp had assigned the same score $t$1. The authors treat that example as evidence that RankSky can provide distinctive rankings where dp-idp cannot, but they do not state or prove a theorem that RankSky always guarantees strict pairwise score distinctness (Martin-Nevot et al., 29 Jul 2025).

RankSky also appears inside the multilevel framework called DeepSky. DeepSky couples multilevel skyline layers to dp-idp, RankSky, or CoSky. It first takes skyline $t$2, then continues with $t$3, $t$4, and further layers if fewer than $t$5 points are available. In that setting, RankSky functions as one ranking back-end within a broader top-$t$6 skyline framework (Martin-Nevot et al., 29 Jul 2025).

6. Interpretation, limitations, and significance

A RankSky score is interpreted in the paper as a stationary importance weight over skyline points. Larger score means that the point is more central or more important in the induced skyline transition structure, and the scores form a normalized probability-like vector. The theoretical guarantees imported from PageRank are standard: a nonnegative stochastic matrix has eigenvalue $t$7; with irreducibility and aperiodicity, the eigenvalue $t$8 is simple; the associated positive eigenvector is unique up to scaling; and normalization to sum 1 yields a unique stationary distribution (Martin-Nevot et al., 29 Jul 2025).

The paper is correspondingly careful about limitations. RankSky requires all preferences to be unified to $t$9, and that transformation can introduce roundoff errors, especially with high cardinality. The method is sensitive to the representation because the transition matrix depends on transformed attribute values. As with PageRank, it may bias toward already structurally central points and may overlook isolated but high-quality skyline points. It is computationally intensive for very large datasets because of the dense $t'$0 matrices, and it is not embeddable in a relational DBMS. The damping factor affects both convergence speed and fairness, but the paper adopts $t'$1 without a sensitivity study. The authors also state that RankSky is insensitive to semantic context or user intent and suggest personalized PageRank variants as a future extension (Martin-Nevot et al., 29 Jul 2025).

These properties make RankSky significant within skyline-query research for a specific reason: it offers a spectral ranking method for skyline points that is neither a dominance-layer heuristic nor a direct ideal-point score. A plausible implication is that RankSky is most appropriate when one wants a global structural ordering over a large skyline, accepts an algorithmic rather than DBMS-embedded implementation, and prefers a PageRank-style stationary interpretation over dominance-local or user-personalized semantics.