Feedback-PageRank: Mechanisms and Applications

• Feedback-PageRank is a family of methods that incorporate recursive feedback into ranking, dynamically adjusting scores based on prior ranks, rewards, or external signals.
• The approach enables applications like efficient feedback arc set computation, reverse-time reward aggregation, and fixed-point personalization analysis in cyclic graphs.
• Recent systems-and-control formulations extend Feedback-PageRank to distributed computation and dynamic teleportation, providing proven convergence and sensitivity guarantees.

Feedback-PageRank denotes a family of PageRank-derived constructions in which ranking is coupled to a feedback mechanism rather than treated solely as the stationary distribution of a fixed random walk. In recent work, the expression has been used for several related procedures: scoring edges through PageRank on a directed line graph in order to compute a feedback arc set, propagating exogenous rewards backward through a graph by a reverse Bellman equation, iterating personalized PageRank until the personalization vector coincides with the induced ranking, and driving teleportation by time-varying external activity signals (Geladaris et al., 2022, Yao et al., 2013, Aleja et al., 22 Jul 2025, Rossi et al., 2012). A complementary systems-and-control perspective interprets classical PageRank itself as a feedback-stabilized linear system, clarifying convergence, damping, aggregation, and distributed implementation (Ishii et al., 2013).

1. Conceptual range and common structure

Classical PageRank appears in the surveyed literature in two equivalent conventions. In a column-stochastic formulation, the PageRank vector satisfies

$r = \alpha P_{\mathrm{col}}^{T} r + (1-\alpha) v,$

where $P_{\mathrm{col}}$ is a link-following transition matrix and $v$ is a personalization vector. In a row-stochastic formulation, the Google matrix is written

$G = \alpha (P + d u^T) + (1-\alpha) e v^T,$

and the PageRank vector $\pi$ satisfies $\pi^T G = \pi^T$ (Aleja et al., 22 Jul 2025). In the systems-and-control formulation, the same computation is written as the affine LTI iteration

$x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$

with teleportation parameter $m \in (0,1)$ and positive stochastic matrix $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$ (Ishii et al., 2013).

Across the feedback variants, the structural modification is not uniform. In reinforcement ranking, the feedback variable is an exogenous reward $r$ propagated backward along in-links through

$P_{\mathrm{col}}$0

In fixed-point personalized PageRank, feedback is introduced by the operator $P_{\mathrm{col}}$1, which maps a personalization vector to the PageRank vector it induces. In evolving-teleportation models, feedback enters through a time-varying teleportation vector $P_{\mathrm{col}}$2. In the feedback-arc-set heuristic, PageRank is not applied to the original vertex set at all, but to a directed line graph whose nodes correspond to edges of the original digraph (Yao et al., 2013, Aleja et al., 22 Jul 2025, Rossi et al., 2012, Geladaris et al., 2022).

A plausible unifying description is that Feedback-PageRank replaces the one-shot ranking of a fixed graph by a recursive mechanism in which rankings, rewards, or external signals are re-injected into the ranking process. The surveyed papers differ principally in where that reinjection occurs: the teleportation term, a reverse-time reward term, the personalization map, or an auxiliary edge-space graph.

2. Edge-space Feedback-PageRank for feedback arc sets

In "Computing a Feedback Arc Set Using PageRank" (Geladaris et al., 2022), the objective is the minimum Feedback Arc Set problem: given a directed graph $P_{\mathrm{col}}$3, find $P_{\mathrm{col}}$4 such that $P_{\mathrm{col}}$5 is acyclic while minimizing $P_{\mathrm{col}}$6. The problem is NP-hard, and it is directly relevant to hierarchical graph drawing, where edges in the computed FAS are typically reversed rather than deleted.

The key construction is the directed line graph $P_{\mathrm{col}}$7. Its vertex set contains one node for each edge of $P_{\mathrm{col}}$8, so $P_{\mathrm{col}}$9. It has a directed edge from $v$0 to $v$1 if and only if the head of $v$2 equals the tail of $v$3; equivalently, edges of $v$4 encode length-2 walks in $v$5 through a middle vertex. Its size satisfies

$v$6

The paper constructs $v$7 by a DFS-inspired procedure in time

$v$8

where $v$9.

PageRank is then run on $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$0 with no damping: $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$1, uniform initialization $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$2, and update

$G = \alpha (P + d u^T) + (1-\alpha) e v^T,$3

If a node has no outgoing links, it keeps its score to itself. The paper reports that about five iterations gave good convergence on small and medium-sized graphs typical of graph drawing. Each iteration costs $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$4 time and $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$5 memory.

The heuristic, PageRankFAS, is SCC-driven. At each outer iteration, Tarjan’s algorithm is used to decompose $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$6 into strongly connected components. Only edges inside SCCs are relevant because edges between SCCs cannot lie on directed cycles. For each SCC $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$7, the algorithm builds $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$8, runs PageRank for a fixed number of iterations, selects the node of $G = \alpha (P + d u^T) + (1-\alpha) e v^T,$9 with maximum PageRank, maps that node back to the corresponding edge $\pi$0 of $\pi$1, adds $\pi$2 to the FAS, removes $\pi$3 from $\pi$4, and repeats until no SCC contains a cycle. The pseudocode removes one highest-scored edge per SCC per outer iteration; tie-breaking is not specified.

The intuition is explicitly edge-centric. Cycles in $\pi$5 become cycles in $\pi$6, and an edge of $\pi$7 that participates in many cyclic walks generates many adjacency relations in $\pi$8. PageRank mass on $\pi$9 therefore acts as a heuristic indicator of how strongly an original edge is embedded in cyclic structure. The method has no approximation guarantees, but it is designed to remove high-impact edges early.

3. Reverse-time reward aggregation and reinforcement ranking

"Reinforcement Ranking" (Yao et al., 2013) formulates authority as reverse-time feedback aggregation rather than as a stationary distribution of a teleportation-augmented random walk. The web graph has $\pi^T G = \pi^T$0 pages and a policy-induced transition matrix $\pi^T G = \pi^T$1 that is row-substochastic: rows corresponding to non-dangling pages sum to $\pi^T G = \pi^T$2, while dangling pages have zero rows. No patching is introduced for dangling pages.

The central object is a bounded reward function $\pi^T G = \pi^T$3, which may encode relevance, clicks, dwell time, trust, topicality, or other exogenous signals. For a surfing policy $\pi^T G = \pi^T$4, the authority score $\pi^T G = \pi^T$5 is defined by discounted aggregation of rewards arriving from predecessors:

$\pi^T G = \pi^T$6

where $\pi^T G = \pi^T$7 controls the propagation horizon. This leads to the reverse Bellman equation

$\pi^T G = \pi^T$8

with fixed point

$\pi^T G = \pi^T$9

and, when it exists,

$x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$0

The existence and convergence theory differs sharply from classical PageRank. Because $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$1 is row-substochastic, $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$2. If $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$3 is bounded and $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$4, then $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$5 is nonsingular, the solution is unique and finite, and the reverse Bellman operator $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$6 is a contraction in the $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$7 norm with modulus $x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$8. Thus irreducibility, ergodicity, and teleportation are not required for well-posedness.

The conceptual distinction from PageRank is explicit. Standard PageRank computes a stationary distribution

$x(k+1) = (1-m) A x(k) + \frac{m}{n}\mathbf{1},$9

where teleportation is used to enforce irreducibility and aperiodicity. Reinforcement ranking instead computes a reward-driven fixed point on the sparse link graph itself, with rewards flowing backward through incoming links. Dangling pages require no special correction, and the method is reported to be less sensitive to graph topology and more local in its response to graph changes. Two locality properties are highlighted: disjoint independence, under which disconnected subgraphs decouple, and altruistic independence, under which a subgraph with only outgoing links is unaffected by external changes so long as no new incoming links are added.

4. Fixed points of personalized PageRank

"Fixed points of Personalized PageRank centrality: From irreducible to reducible networks" (Aleja et al., 22 Jul 2025) gives a precise mathematical characterization of feedback-PageRank as a fixed-point problem for personalized PageRank. With row-stochastic transition matrix $m \in (0,1)$0 and damping $m \in (0,1)$1, define

$m \in (0,1)$2

If $m \in (0,1)$3 is a personalization vector, then the row-vector PageRank is

$m \in (0,1)$4

The feedback operator is

$m \in (0,1)$5

that is, personalization is mapped to the PageRank vector it generates.

A fixed point $m \in (0,1)$6 satisfies $m \in (0,1)$7, hence

$m \in (0,1)$8

Rearranging the PageRank equation yields the key condition

$m \in (0,1)$9

Therefore, fixed points of personalized PageRank are exactly stationary distributions of the underlying Markov chain $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$0, and they are independent of $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$1.

This identification makes the SCC structure decisive. If the graph is strongly connected, then for every $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$2 there is a unique fixed point, equal to the unique left Perron vector $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$3 of $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$4. More generally, after permuting nodes into block upper-triangular SCC form, the final or sink SCCs determine the full fixed-point set. If $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$5 are the sink SCC blocks with left Perron vectors $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$6, then the feedback iteration

$M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$7

converges to

$M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$8

where $M = (1-m)A + \frac{m}{n}\mathbf{1}\mathbf{1}^\top$9 and the coefficients depend on the initial vector $r$0. The set of fixed points is therefore the convex hull of the embedded Perron vectors of the sink SCCs. Uniqueness holds if and only if there is a single final SCC.

The paper also clarifies the role of dangling nodes. In the Google matrix they are handled by the usual rank-one correction

$r$1

which guarantees a positive row-stochastic matrix and hence a unique PageRank vector for any positive $r$2. For the feedback analysis, however, the fixed-point characterization remains the stationary-distribution equation on the link-following chain. Computationally, the paper notes that SCC decomposition followed by local Perron-vector computation on sink SCCs is more direct and exact than repeatedly solving PageRank subproblems.

5. Dynamic, distributed, and control-theoretic feedback formulations

The systems-and-control account in "The PageRank Problem, Multi-Agent Consensus and Web Aggregation -- A Systems and Control Viewpoint" (Ishii et al., 2013) treats PageRank as a closed-loop affine system. With column-stochastic hyperlink matrix $r$3 and damping parameter $r$4,

$r$5

and the Google matrix

$r$6

is positive and column-stochastic. Perron–Frobenius theory then yields a unique positive eigenvector $r$7 with $r$8. The convergence rate of the centralized power method is governed by the second eigenvalue, with

$r$9

The paper also reports the sensitivity bound

$P_{\mathrm{col}}$00

That same paper develops two operational feedback mechanisms. The first is a distributed randomized algorithm based on local link matrices $P_{\mathrm{col}}$01 and i.i.d. page activation $P_{\mathrm{col}}$02. With design parameter

$P_{\mathrm{col}}$03

the update

$P_{\mathrm{col}}$04

is paired with the time average

$P_{\mathrm{col}}$05

Mean-square ergodic convergence holds:

$P_{\mathrm{col}}$06

The second is aggregation. After a coordinate transform $P_{\mathrm{col}}$07, one obtains a reduced-order recursion on group totals and intra-group deviations. If the grouping sparsity parameter $P_{\mathrm{col}}$08 satisfies

$P_{\mathrm{col}}$09

then the reduced solution $P_{\mathrm{col}}$10 obeys

$P_{\mathrm{col}}$11

"Dynamic PageRank using Evolving Teleportation" (Rossi et al., 2012) inserts feedback directly through a time-varying teleportation vector. With column-stochastic $P_{\mathrm{col}}$12 and normalized external signal $P_{\mathrm{col}}$13, dynamic PageRank is governed by the ODE

$P_{\mathrm{col}}$14

Its exact solution is

$P_{\mathrm{col}}$15

When $P_{\mathrm{col}}$16 is constant, the solution reduces to

$P_{\mathrm{col}}$17

where $P_{\mathrm{col}}$18 is the static PageRank solution of $P_{\mathrm{col}}$19, and therefore $P_{\mathrm{col}}$20. Forward Euler with step size $P_{\mathrm{col}}$21 gives

$P_{\mathrm{col}}$22

which becomes the classical Richardson iteration when $P_{\mathrm{col}}$23 and $P_{\mathrm{col}}$24 is held fixed.

6. Empirical behavior, applications, and limitations

The edge-space feedback-arc-set heuristic of (Geladaris et al., 2022) is reported to produce substantially smaller FAS than GreedyFAS and SortFAS on graph-drawing instances. On random graphs with 100 to 4000 nodes and average out-degree $P_{\mathrm{col}}$25, $P_{\mathrm{col}}$26, and $P_{\mathrm{col}}$27, PageRankFAS typically yields FAS that are less than $P_{\mathrm{col}}$28 of GreedyFAS and SortFAS, and for 4000-node graphs with about 12,000 edges the reduction is almost $P_{\mathrm{col}}$29 versus GreedyFAS. Runtime is reported as less than $P_{\mathrm{col}}$30 second for graphs with at most 1000 nodes and at most $P_{\mathrm{col}}$31 seconds for 4000-node instances, while the comparison heuristics run in roughly $P_{\mathrm{col}}$32 to $P_{\mathrm{col}}$33 seconds. On the webgraphs wordassociation-2011 and enron, the reported FAS percentages are $P_{\mathrm{col}}$34, $P_{\mathrm{col}}$35, and $P_{\mathrm{col}}$36 for GreedyFAS, SortFAS, and PageRankFAS on the former, and $P_{\mathrm{col}}$37, $P_{\mathrm{col}}$38, and $P_{\mathrm{col}}$39 on the latter. The same study emphasizes the principal limitation: runtime and memory are dominated by the size of the line graph, with $P_{\mathrm{col}}$40 and worst-case $P_{\mathrm{col}}$41 behavior on dense digraphs. Weighted graphs are not considered.

The reverse-time reinforcement model of (Yao et al., 2013) is evaluated on four Wikipedia snapshots from Oct-2008 to Jan-2011, the largest with about $P_{\mathrm{col}}$42M pages and about $P_{\mathrm{col}}$43M links, and on the DBLP citation graph with about $P_{\mathrm{col}}$44M nodes and about $P_{\mathrm{col}}$45M links. When updating rankings after graph changes, initialization from the previous snapshot gives Reinforcement Ranking about $P_{\mathrm{col}}$46 better accuracy than uniform initialization at the same iteration budget, whereas PageRank’s convergence from a previous solution is reported not to be significantly better than from uniform initialization. On DBLP, the top-ranked results are said to align better with later-observed citation counts and to avoid low-citation outliers that appear among top PageRank results. The limitations emphasized by the authors are reward design, policy specification, and the fact that the authority vector is not a probability distribution.

The dynamic-teleportation model of (Rossi et al., 2012) is tested on a Wikipedia graph with about $P_{\mathrm{col}}$47M nodes and about $P_{\mathrm{col}}$48M edges using hourly pageviews, and on a Twitter follower graph with about $P_{\mathrm{col}}$49k nodes and about $P_{\mathrm{col}}$50k edges using monthly tweet counts. The paper reports that cumulative pageviews and in-degree on Wikipedia are essentially uncorrelated, with correlation about $P_{\mathrm{col}}$51, which motivates explicit external-input modeling. For one-step-ahead forecasting, the dynamic PageRank model outperforms a baseline using only the external signal. Reported sMAPE values are $P_{\mathrm{col}}$52 versus $P_{\mathrm{col}}$53 on non-stationary Wikipedia nodes and $P_{\mathrm{col}}$54 versus $P_{\mathrm{col}}$55 on stationary ones; on Twitter, the corresponding values are $P_{\mathrm{col}}$56 versus $P_{\mathrm{col}}$57 and $P_{\mathrm{col}}$58 versus $P_{\mathrm{col}}$59. The paper also notes practical limitations: the graph is static, the number of Euler steps per update period is application-dependent, and smoothing of noisy $P_{\mathrm{col}}$60 is identified as a promising extension.

The systems-and-control formulation of (Ishii et al., 2013) extends the feedback theme to distributed computation and bibliometrics. It gives a distributed randomized PageRank estimator with mean-square ergodic convergence and an aggregation-based approximation with a provable $P_{\mathrm{col}}$61 error bound. It also presents a citation-ranking application for control journals through Eigenfactor-style scores based on a cross-citation matrix, with damping typically set to $P_{\mathrm{col}}$62 and personalization vector given by article-volume fractions.

A recurring misconception is that Feedback-PageRank names a single standard algorithm. The literature surveyed here instead exhibits several distinct mechanisms that share a feedback interpretation but solve different problems: cycle breaking in directed graphs, authority propagation without teleportation, fixed-point analysis of personalized PageRank, and temporal adaptation to exogenous attention signals. What they have in common is not a single update rule, but the replacement of one-way ranking by recursive dependence on prior ranks, rewards, or external inputs.