FastQC: γ-Quasi-Clique Enumeration

• FastQC is a branch-and-bound algorithm that efficiently enumerates maximal γ-quasi-cliques in undirected graphs through advanced pruning and branching rules.
• It employs size–disconnection pruning and symmetry-exploiting candidate refinement to limit the search space and ensure the QC maximality property.
• The divide-and-conquer variant, DCFastQC, accelerates processing on large or dense graphs, achieving up to 100× speedup over previous approaches.

FastQC is a branch-and-bound (BB) algorithm designed for the efficient enumeration of all large maximal $\gamma$-quasi-cliques (MQCs) in undirected graphs. The method introduces co-designed pruning and branching rules to overcome time complexity barriers of prior approaches, complemented by a divide-and-conquer enhancement. FastQC enables the practical discovery of cohesive subgraph structures at scales and densities previously infeasible due to both theoretical worst-case bounds and empirical performance bottlenecks (Yu et al., 2023).

1. Problem Formulation: Maximal γ-Quasi-Clique Enumeration

Let $G = (V, E)$ be an undirected graph with $n = |V|$ vertices. Fix a density parameter $\gamma \in [0.5, 1]$ and a size threshold $\sigma > 0$.

• γ-Quasi-Clique (γ-QC): An induced subgraph $G[H]$ ($H \subseteq V$) is a $\gamma$-quasi-clique if:
  1. $G[H]$ is connected; and
  2. For every $v \in H$, $\deg_{G[H]}(v) \geq \gamma \cdot (|H| - 1)$.

Equivalently, defining the maximum number of disconnections in $G[H]$ as

$$A(H) \equiv \max_{v \in H} (|H| - 1 - \deg_{G[H]}(v)),$$

we require $A(H) \leq T(|H|)$, where $T(x) = \lfloor (1 - \gamma)x + \gamma \rfloor$.

• Maximal γ-Quasi-Clique (MQC): A $\gamma$-QC $G[H]$ is maximal if no $H' \supset H$ exists such that $G[H']$ is also a $\gamma$-QC.
• Enumeration Problem (MQCE): Output all MQCs in $G$ of size at least $\sigma$.

Since QC maximality checking is NP-hard and the QC family is non-hereditary, enumeration is typically staged:

1. Enumerate a superset $S$ containing all MQCs.
2. Filter $S$ for maximality.

FastQC addresses the superset enumeration stage via an advanced BB framework.

2. Branch-and-Bound Framework in FastQC

FastQC explores the space of vertex subsets using a recursive BB search, where each node (branch) is represented by a triple $(S, C, D)$:

• $S$: current partial set (must-include vertices)
• $C$: candidate set (potential additions to $S$)
• $D$: excluded set (must-not-include vertices)

Subsets $H$ satisfying $H \supseteq S$, $H \subseteq S \cup C$, $H \cap D = \emptyset$ are explored under each node.

The high-level recursive process FastQC-Rec(S, C, D) operates as follows:

• Apply pruning (described in Section 3).
• If candidate set $C$ is empty, test $S$ for being a $\gamma$-QC (output if $|S| \geq \sigma$).
• When an early-stop maximality criterion is satisfied, output $S \cup C$ as a found MQC.
• Otherwise, select a pivot $u \in S \cup C$ and branch via the selected branching strategy (see Section 4).

3. Pruning Mechanisms: Size–Disconnection Space

FastQC introduces pruning rules in the "size–disconnection (SD)" space, efficiently restricting search to promising branches. For any $H \subseteq V$:

• $A(H) = \max_{v \in H} (|H| - 1 - \deg_{G[H]}(v))$
• $T(x) = \lfloor (1 - \gamma)x + \gamma \rfloor$

For each node $B = (S, C, D)$:

• Lower bound on QC size: $|S|$
• Upper bound: $$o(B) = \min\big\{|S \cup C|,~ \lfloor d_{\min}(B)/\gamma \rfloor + 1\big\}$$, where $d_{\min}(B) = \min_{v \in S} \deg_{G[S \cup C]}(v)$

Branch $B$ can only contain a $\gamma$-QC if:

$A(S) \leq T(o(B))$

Vertex refinement in $C$ is performed by progressive pruning:

• (R1) Remove $v$ if $A(S \cup \{v\}) > T(o(B))$ — inclusion of $v$ violates QC bounds.
• (R2) Remove $v$ if $\deg_{G[S \cup C]}(v) < \sigma - T(o(B))$ — no large QC could include $v$.

These pruning operations are iterated, updating $o(B)$ and re-checking the pruning condition until no further removal is possible. If $A(S) > T(o(B))$ at any step, the branch is safely pruned. This ensures the correctness and safety of the enumeration process, as only subsets $H \supseteq S$ with $A(H) \leq T(|H|) \leq T(o(B))$ can form $\gamma$-QCs.

4. Branching Strategies: Sym-SE and Hybrid-SE

Once pruning and candidate refinement stall, FastQC branches the current search node using symmetry-exploiting rules.

• Symmetric SE branching (Sym-SE): Order candidates $C = (v_1, \ldots, v_m)$. Branch into $m+1$ child nodes, each enforcing inclusion/exclusion decisions on $v_i$. For increasing $i$, $A(S_i)$ increases; if $A(S_i) > T(o(B))$, all further siblings are pruned.
• Hybrid-SE branching: Triggered when a pivot $u$ has $\deg_{dis}(u) > T(o(B))$ and $\deg_S(u) = 0$. The "exclude $u$" branch uses SE branching and a separate necessary condition for maximality; the "include $u$" uses Sym-SE to prune non-QC branches—boosting pruning efficacy.

Pivot selection prefers $u \in S \cup C$ with maximal disconnections to maximize the utility of subsequent pruning.

5. Worst-Case Time Complexity

Let $k = \max_{B} T(o(B))$, with the bound $k \leq T(n)$ and typically $k \ll n$. Define $\alpha_k$ as the maximal real root of $x^{k+2} - x^{k+1} - 2x^{k} + 2 = 0$. FastQC explores at most $O(n d \alpha_k^n)$ recursion nodes, with $\alpha_k < 2$ for all $k \geq 1$ (e.g., $\alpha_2 \approx 1.769$, $\alpha_3 \approx 1.899$, $\alpha_4 \approx 1.953$).

All pruning and branching rules are proven, via induction and recurrence analysis, to satisfy this exponential node growth, improving on the $O^*(2^n)$ worst-case bounds of prior BB approaches (Yu et al., 2023).

6. Divide-and-Conquer Enhancement: DCFastQC

To scale FastQC to massive or dense graphs, a divide-and-conquer variant ("DCFastQC", Editor's term) is employed:

• Compute a degeneracy ordering $v_1, \ldots, v_n$.
• For each $v_i$, define $$V_i = \{v_j : \mathrm{dist}_G(v_i, v_j) \leq 2,~ j \geq i\}$$ and form subgraph $G_i = G[V_i]$.
• Each MQC appears in exactly one $G_i$ where $v_i$ is minimal by degeneracy rank.
• Algorithm outline:
  • Form $G_i = G[V_i]$;
  • Iteratively prune via one-hop/two-hop pruning;
  • Call `FastQC-Rec(S = {v_i}, C = V_i \setminus {v_i}, D = {v_1, \ldots, v_{i-1}})$.</li> </ul></li> </ul> <p>For degeneracy$\delta$and maximum degree$d$, each$G_i$has$O(\delta d)$vertices; aggregate runtime is$O(n \delta d \alpha_k^{\delta d})$$, yielding substantial acceleration on sparse or locally bounded graphs.</p> <h2 class='paper-heading' id='empirical-performance-and-observations'>7. Empirical Performance and Observations</h2> <p>Empirical evaluation on both real (12 network datasets, up to$$23$M vertices and$261$M edges) and synthetic (Erdős–Rényi, up to$10^6$$vertices, varied densities) graphs demonstrates:</p> <ul> <li>Against Quick+ (state-of-the-art BB), DCFastQC achieves up to$$100\times$$speedup on real-world networks (e.g., Enron, WordNet, Pokec), solving instances in seconds where Quick+ requires minutes or fails due to memory exhaustion.</li> <li>Speedup increases with$$\gamma$and$\sigma$$as the search space becomes more amenable to aggressive pruning.</li> <li>On dense Erdős–Rényi graphs, DCFastQC successfully handles$$10^5$–$10^6$vertices and edge-densities up to$70$, where Quick+ fails beyond$10^5$$vertices.</li> <li>DCFastQC and FastQC return identical sets of MQCs, but enumerate far fewer non-maximal QCs, significantly reducing the burden of the subsequent maximality filtering step and further improving overall enumeration efficiency.</li> </ul> <p>These results establish FastQC (and DCFastQC) as among the fastest and most theoretically robust algorithms for maximal$$\gamma$-quasi-clique enumeration in large-scale graphs (Yu et al., 2023).