Single-Source Personalized PageRank (SSPPR)

• SSPPR is a graph query that computes the full source-conditioned personalized PageRank distribution using an α-discounted random walk model.
• It employs iterative, local-push, and hybrid methods to balance accuracy and efficiency in handling large, dynamic graphs.
• Analytical and dynamic studies of SSPPR reveal tight complexity bounds and effective update techniques for real-time recommendation and graph learning.

Single Source Personalized PageRank (SSPPR) is the query that, for a fixed source node $s$, computes the Personalized PageRank values $\pi(s,t)$ for all targets $t\in V$. In the standard random-walk semantics, $\pi(s,t)$ is the probability that an $\alpha$-discounted random walk starting at $s$ terminates at $t$; equivalently, SSPPR returns the full source-conditioned proximity profile rather than a single pairwise score. Within the broader Personalized PageRank literature, SSPPR is a principal query type because it underlies recommendation, search, local graph analysis, and several matrix-based or learning-oriented constructions (Wei et al., 2024, Yang et al., 2024).

1. Formal definition and source-centered semantics

For a fixed source $s$, the standard fixed-point formulation writes the PPR vector as

$$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$

where $\vec{e_s}$ is the one-hot vector for $\pi(s,t)$0, and $\pi(s,t)$1 is the transition matrix with

$\pi(s,t)$2

Under the equivalent random-walk interpretation, a walk starts at $\pi(s,t)$3, and at each step it either terminates at the current node with probability $\pi(s,t)$4, or moves to a uniformly chosen out-neighbor with probability $\pi(s,t)$5; $\pi(s,t)$6 is then the probability that the walk from $\pi(s,t)$7 terminates at $\pi(s,t)$8 (Wang et al., 2020). The resulting vector is a probability distribution, satisfying

$\pi(s,t)$9

(Wang et al., 2020).

The survey taxonomy distinguishes SSPPR from two adjacent tasks. Single-pair PPR asks for one value $t\in V$0, while single-target PPR asks for $t\in V$1 for all sources $t\in V$2 toward a fixed target $t\in V$3. Non-personalized PageRank differs in that the teleportation distribution is uniform rather than concentrated on one source (Wei et al., 2024, Yang et al., 2024). This distinction is algorithmically consequential rather than merely terminological.

A more general personalization viewpoint replaces the single source by a personalization vector. In directed graphs with dangling nodes, one formulation is

$t\in V$4

with

$t\in V$5

The source-specific case is recovered when the personalization mass is concentrated near one seed (Garcia et al., 2012). A complementary Markov-chain view writes PageRank as the stationary distribution of the perturbed chain

$t\in V$6

with personalized PageRank corresponding to the rank-one perturbation $t\in V$7; in that sense SSPPR is a particular restart-biased perturbation of a possibly reducible base walk (Borkar et al., 7 Mar 2025).

2. Core algorithmic paradigms

Classical SSPPR algorithms fall into three broad families: global iterative methods, local-push methods, and hybrid methods. In the survey account, standard Power Iteration and Cumulative Power Iteration solve the source-specific fixed-point equation by repeated graph-wide matrix-vector updates, yielding $t\in V$8-error $t\in V$9 in

$\pi(s,t)$0

time (Yang et al., 2024). These methods are deterministic and high-precision, but they scan the graph globally.

Forward Push is the canonical local alternative. It maintains a reserve vector and a residue vector, pushes only from active nodes, and exploits the decomposition theorem

$\pi(s,t)$1

in source-centered form (Yang et al., 2024). A long-standing open question concerned whether Forward Push could match Power Iteration’s logarithmic dependence on the absolute error threshold $\pi(s,t)$2. That question was answered positively: a common FIFO implementation of Forward Push has running time

$\pi(s,t)$3

for high-precision SSPPR, matching the asymptotic bound of Power Iteration (Wu et al., 2021). The same work introduced PowerPush, which combines the strengths of Power Iteration and Forward Push, and SpeedPPR, an approximate SSPPR algorithm with overall expected time

$\pi(s,t)$4

on scale-free graphs, improving the $\pi(s,t)$5 bound of FORA (Wu et al., 2021).

Hybrid methods combine local propagation with Monte Carlo refinement. FORA uses Forward Push to accumulate reserve and reduce the residue mass, then estimates the remaining contribution by random walks. Its core identity is

$\pi(s,t)$6

which allows the residual term to be sampled rather than fully propagated (Wang et al., 2019). This construction yields approximate SSPPR algorithms with rigorous guarantees on result quality, an effective indexing scheme, and a top-$\pi(s,t)$7 selection module with high pruning power (Wang et al., 2019).

Weighted graphs expose a weakness of node-level LocalPush: one push from a node distributes probability to all neighbors, even when only a few heavy edges dominate the transition mass. EdgePush addresses this by decomposing node pushes into edge-level pushes with per-edge thresholds $\pi(s,t)$8. It maintains node income $\pi(s,t)$9, edge expense $\alpha$0, and edge residue

$\alpha$1

and proves an improvement over LocalPush by an order of up to $\alpha$2 on unbalanced weighted graphs, both for $\alpha$3-error and normalized additive error (Wang et al., 2022).

3. Approximation models, upper bounds, and lower bounds

SSPPR is studied under several approximation regimes. A common thresholded model requires constant relative error for targets with $\alpha$4, while permitting additive error $\alpha$5 below threshold. In undirected graphs, the formal requirement can be written as

$\alpha$6

for fixed constants $\alpha$7 (Bertram et al., 11 Feb 2026). This model separates statistically meaningful entries from the long tail.

For undirected graphs, the complexity of thresholded SSPPR is now tightly characterized. The main result is

$\alpha$8

time in the worst case, with the same asymptotic bound in the average case, in the adjacency-list model and its variants (Bertram et al., 11 Feb 2026). The analysis exploits reversibility,

$\alpha$9

which does not hold in directed graphs and rules out asymmetry-based lower-bound constructions used in the directed setting (Bertram et al., 11 Feb 2026).

Absolute-error SSPPR admits a different complexity landscape. For the query that requires

$s$0

with high probability, a hybrid Monte Carlo plus Adaptive Backward Push algorithm achieves expected running time

$s$1

for directed graphs,

$s$2

for undirected graphs with maximum degree $s$3, and

$s$4

for power-law graphs with $s$5 (Wei et al., 2024). The same work also studies a degree-normalized absolute-error guarantee,

$s$6

with expected complexity

$s$7

on undirected graphs (Wei et al., 2024).

Recent lower bounds substantially tighten the SSPPR complexity picture. For relative-error SSPPR-R, any algorithm in the arc-centric graph access model requires

$s$8

queries. For additive-error SSPPR-A, the lower bound is

$s$9

under the sparsity assumption $t$0 for fixed $t$1 (Jiang et al., 19 Jul 2025). These results improve earlier bounds by logarithmic factors and bring lower bounds closer to the best known upper bounds.

4. Dynamic and evolving graphs

Dynamic graphs motivate SSPPR schemes that avoid recomputing the full source-conditioned distribution after every edge update. One early line of work stores short random-walk segments starting from every node and reuses them to answer top-$t$2 personalized PageRank queries from a seed node. Under the assumption that personalized PageRanks follow a power law with exponent $t$3, if

$t$4

random-walk segments are stored at each node, then the expected number of database fetches for top-$t$5 personalized PageRanks is

$t$6

(Bahmani et al., 2010). In the same framework, insertions and deletions can be maintained efficiently; for example, the expected total update work over $t$7 edge arrivals is $t$8 with the $t$9-dependence explicit in the derivation, and deletions are handled with comparable efficiency (Bahmani et al., 2010).

A more recent development is FIRM, an indexing scheme for evolving graphs built on complete random-walk paths rather than only walk terminals. FIRM stores full paths

$s$0

and incrementally repairs only the walks affected by an insertion or deletion. Under the random arrival model and the condition

$s$1

both Update-Insert and Update-Delete run in expected $s$2 time per edge update (Hou et al., 2022). The query pipeline remains FORA / SpeedPPR style: a Forward-Push phase computes reserve and residue vectors, and precomputed walks are consumed as needed for random-walk refinement (Hou et al., 2022).

These dynamic results establish a recurring SSPPR pattern: precompute or cache source-independent walk primitives, update them incrementally, and defer source-specific computation to a lightweight query-time composition step. This suggests that the practical bottleneck in evolving-graph SSPPR is often index maintenance rather than the final source-conditioned estimation itself.

5. Structural and analytical viewpoints

Beyond algorithm design, SSPPR has several analytical descriptions that clarify how source personalization shapes the output distribution. In directed graphs with dangling nodes, the dependence on the personalization vector is linear: $s$3 For a node $s$4, the set of attainable personalized PageRank values is exactly

$s$5

an open interval determined by the $s$6-th column of $s$7 (Garcia et al., 2012). The same matrix yields criteria for effective competitors and leaders: sign changes between two columns characterize competitive pairs, while row-wise strict maxima identify leadership nodes (Garcia et al., 2012). In SSPPR terms, source bias changes rankings within explicit structural bounds rather than arbitrarily.

A second viewpoint concerns compressibility across sources. For a certain class of random graphs, the effective dimensionality of the family of PPR vectors scales sublinearly in $s$8 with high probability, even though the full PPR matrix has rank $s$9. The analysis uses a Jeh–Widom-style hub decomposition in which a source-specific vector $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$0 is approximated from a small hub set $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$1 and a modified-chain vector $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$2, and bounds the total number of PPR values computed by

$$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$3

(Vial et al., 2018). This suggests that, on suitable graph families, many source-specific PPR vectors can be reconstructed from a sublinear set of hub vectors rather than computed independently.

The singular-perturbation view adds a coarse-grained asymptotic interpretation. For

$$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$4

with closed communicating classes $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$5, the limiting PageRank mass factorizes into within-class stationary distributions and a class-level stationary distribution induced by $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$6. In the personalized case $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$7, the class-level chain satisfies

$$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$8

so the zero-noise limit biases mass toward the communicating classes favored by the personalization vector $$\vec{PPR_s}=(1-\alpha)\vec{PPR_s}\cdot \mathbf{P}+\alpha \cdot \vec{e_s},$$9 (Borkar et al., 7 Mar 2025). For SSPPR, this formalizes the idea that restart mass first selects relevant classes and only then distributes mass internally.

6. Related formulations, distinctions, and applications

A persistent distinction in the literature is between source-centric and target-centric computation. Single-source PPR asks, in effect, “which nodes are important from $\vec{e_s}$0?”, whereas single-target PPR asks “which nodes contribute to the importance of $\vec{e_s}$1?” The two are not computationally symmetric, even though they are built from the same quantity $\vec{e_s}$2 (Wang et al., 2020). In particular, Backward Search yields a clean additive approximation guarantee for single-target queries, whereas Forward Search gives only a weak additive guarantee on undirected graphs in that context (Wang et al., 2020).

A related misconception is that target-centric PPR can be obtained simply by reversing edges and solving the usual source-centric problem. The target-node literature explicitly rejects that identification: target-centric PPR is not simply “reverse the graph and compute source-based PPR from $\vec{e_s}$3” because the walk semantics and recurrence differ (Lofgren et al., 2013). This distinction matters when SSPPR is used as a building block inside more elaborate systems.

Applications of SSPPR are broad but technically specific. In social networks, personalized recommendation and search are standard use cases; approximate top-$\vec{e_s}$4 SSPPR has been evaluated on Twitter-scale graphs, where FORA answers a top-500 approximate SSPPR query within 1 second on a billion-edge Twitter dataset using a single commodity server (Wang et al., 2019). In dynamic social networks, Monte Carlo–based maintenance was reported to be fast enough for real-time queries over a dynamic social network (Bahmani et al., 2010).

SSPPR also appears as an internal primitive in other models. ProPPR performs first-order probabilistic inference by a personalized PageRank variant over a query-specific proof graph, augmented with restart edges back to the query root; grounding time is independent of DB size, and order-of-magnitude speedups are possible by parallelizing learning (Wang et al., 2013). In graph mining and learning, target-oriented PPR algorithms have been used to improve heavy hitters PPR query, single-source SimRank computation, and scalable graph neural networks via approximate PPR matrices (Wang et al., 2020).

Taken together, these developments position SSPPR as both a standalone graph query and a reusable computational primitive. Its modern theory combines random-walk semantics, local linear-algebraic invariants, tight approximation tradeoffs, dynamic maintenance schemes, and structural results about personalization, reversibility, and low effective dimension.