MD-HITS: Multi-Dimensional Network Ranking

• MD-HITS is a nonlinear eigenvector framework that generalizes classical HITS by ranking entities in temporal, multilayer, weighted networks.
• It computes five interdependent centrality vectors (hubs, authorities, layer broadcast, layer receive, and temporal scores) using a power-iteration algorithm with global convergence guarantees.
• The model overcomes classical HITS limitations by ensuring unique, robust rankings even in disconnected or sparse network structures through a multi-homogeneous, order-preserving approach.

The Multi‐Dimensional HITS (MD‐HITS) model is a nonlinear eigenvector-based framework for ranking entities in temporal, multilayer, weighted, and directed networks. Developed by Arrigo and Tudisco (2018), MD-HITS generalizes Kleinberg’s classical HITS algorithm to networks structured by multiple interaction layers and time stamps. It defines five interdependent centrality vectors—accounting for nodes, layers, and temporal epochs—via the Perron eigenvector of a multi-homogeneous, order-preserving operator, guaranteeing well-posedness under broad conditions and enabling robust, globally convergent computation on arbitrary nonnegative networks (Arrigo et al., 2018).

1. Network Structure and Model Formulation

MD-HITS targets data naturally encoded as sequential interactions across distinct “layers” and discrete time stamps. The foundational objects are:

• Nodes: A finite set $V = \{1, ..., n_V\}$ (e.g., authors, airports).
• Layers: A finite set $L = \{1, ..., n_L\}$ representing modes (e.g., journals, airlines).
• Time stamps: A set $T = \{1, ..., n_T\}$ labeling discrete time intervals.
• Edges: Defined as quintuples $(i,\ell) \rightarrow (j,k)$ at time $t$, representing an interaction from node $i$ in layer $\ell$ to node $j$ in layer $k$ occurring at $t$.
• Adjacency tensor: $L = \{1, ..., n_L\}$0, with $L = \{1, ..., n_L\}$1 if the corresponding interaction exists.

This structure accommodates complex scenarios such as research citation flows between authors across journals over years, or multilayer social/transportation networks with temporal evolution.

2. Multi-Homogeneous Order-Preserving Mapping and Centrality Vectors

MD-HITS introduces five centrality vectors, each normalized to unit $L = \{1, ..., n_L\}$2-norm and composed as $L = \{1, ..., n_L\}$3 with:

• $L = \{1, ..., n_L\}$4: node hub-scores,
• $L = \{1, ..., n_L\}$5: node authority-scores,
• $L = \{1, ..., n_L\}$6: layer broadcast centrality,
• $L = \{1, ..., n_L\}$7: layer receive centrality,
• $L = \{1, ..., n_L\}$8: temporal stamp scores.

Raw (multiway) tensor-vector products define interactions for each centrality type: $L = \{1, ..., n_L\}$9 Nonlinearity is injected by entrywise exponentiation: with $T = \{1, ..., n_T\}$0,

$T = \{1, ..., n_T\}$1

The MD-HITS centralities are given by $T = \{1, ..., n_T\}$2, with $T = \{1, ..., n_T\}$3, satisfying the nonlinear eigen-equation: $T = \{1, ..., n_T\}$4 component-wise, $T = \{1, ..., n_T\}$5 for $T = \{1, ..., n_T\}$6.

3. Interpretation of Centralities

Each of the five centrality vectors has a distinct semantic role:

• Hub-score $T = \{1, ..., n_T\}$7: Quantifies how strongly node $T = \{1, ..., n_T\}$8 points to influential nodes (with high authority $T = \{1, ..., n_T\}$9), on important broadcast layers $(i,\ell) \rightarrow (j,k)$0 (high $(i,\ell) \rightarrow (j,k)$1), at important times $(i,\ell) \rightarrow (j,k)$2.
• Authority-score $(i,\ell) \rightarrow (j,k)$3: Measures how strongly node $(i,\ell) \rightarrow (j,k)$4 receives from prominent hubs $(i,\ell) \rightarrow (j,k)$5, over high-broadcast layers $(i,\ell) \rightarrow (j,k)$6, at critical times $(i,\ell) \rightarrow (j,k)$7.
• Layer broadcast $(i,\ell) \rightarrow (j,k)$8: Captures the extent to which layer $(i,\ell) \rightarrow (j,k)$9 enables powerful hubs to reach top authorities.
• Layer receive $t$0: Assesses how layer $t$1 accumulates influential flows between hubs and authorities.
• Temporal stamp $t$2: Rates the importance of time $t$3 in conveying significant interactions between hubs and authorities across layers.

This five-fold structure encodes mutually reinforcing relationships between nodes, layers, and temporal windows in a unified nonlinear eigenvector framework.

4. Existence, Uniqueness, and Theoretical Guarantees

MD-HITS achieves broad well-posedness by leveraging a multi-homogeneous order-preserving map and nonlinear contraction principles. Define the $t$4 “homogeneity” matrix: $t$5 If $t$6 and the spectral radius $t$7, then there is exactly one positive solution $t$8 (with $t$9 for all $i$0) and unique $i$1 to $i$2, with $i$3 maximal in a Perron-type sense. This result utilizes a higher-order Hilbert-metric contraction. A sufficient condition is: $i$4 With all $i$5, $i$6, but uniqueness and existence remain valid.

5. Power Iteration Algorithm and Computational Considerations

MD-HITS employs a globally convergent, power-iteration-like procedure for centrality computation. At iteration $i$7:

1. Compute raw scores:

$i$8

2. Normalize each to unit $i$9-norm:

$\ell$0

3. Check convergence: e.g., $\ell$1.

Each iteration costs $\ell$2, requiring only sparse tensor-vector multiplies per mode. Under $\ell$3, the map contracts in a suitable projective metric, yielding linear convergence at rate $\ell$4: $\ell$5

6. Special Cases, Comparisons, and Empirical Assessment

MD-HITS encompasses classical and multilayer centrality frameworks as special cases:

• Standard HITS reduction: When $\ell$6 and $\ell$7, MD-HITS yields the standard HITS equations. However, classical HITS requires strong connectivity ($\ell$8 irreducible)—MD-HITS with $\ell$9 admits a unique solution for any $j$0, mitigating the “curse of disconnectedness.”
• Nonlinear HITS: Setting $j$1 in the monolayer setting gives a unique nonlinear HITS pair without irreducibility requirements.
• Other extensions (Co-HITS, MultiRank, HAR, tensor-SVD): These approaches need irreducibility assumptions for layers or supra-adjacency structure and are locally (not globally) convergent; moreover, they cannot seamlessly incorporate inter-layer edges or temporal stamps.

Empirical evaluations demonstrate:

• On a small toy graph, standard HITS produces multiple fixed points depending on initialization; MD-HITS (e.g., $j$2) yields a unique, theoretically justified ranking.
• For random sparse tensors, MD-HITS scales linearly in $j$3; convergence observed in dozens of iterations (with $j$4) even for networks with millions of nodes/layers/times.
• Applied to a real citation network ($j$5k authors, $j$6k journals, $j$7 years), MD-HITS generates five robust rankings in $j$8 seconds; centralities vary minimally for perturbations in $j$9 (intersection similarity $k$0), and temporal centrality picks up established citation bursts.
• In the FAO trade dataset ($k$1 countries, $k$2 goods, no temporal dimension), standard vector-based methods fail on the highly disconnected aggregate; MD-HITS assigns positive broadcaster/receiver scores to all but true isolates and plausibly ranks traded commodities, e.g., soybeans as most central in 2010.

7. Significance and Impact in Network Science

MD-HITS offers a principled, universally well-posed, and computationally tractable generalization of eigenvector-based centrality to complex temporal multilayer structures. Its nonlinear, multi-homogeneous architecture establishes unique, positive, and interpretable rankings for nodes, layers, and time stamps, addressing critical limitations of prior linear and tensor-based centrality models—particularly with respect to disconnected or non-irreducible interaction patterns. It provides a foundation for the analysis of compositional, time-evolving, and loosely coupled networks in domains ranging from bibliometrics to infrastructure modeling (Arrigo et al., 2018).