SubsCoRe: Contrast Subgraph Mining

• SubsCoRe is a method that systematically extracts contrast subgraphs from coherent cores by optimizing coherence and contrast scores via binary search and min-cut algorithms.
• It detects significant temporal and contextual shifts in pairwise weighted graphs by leveraging precise mathematical formulations and scalable optimization procedures.
• SubsCoRe has been empirically validated across diverse scenarios, demonstrating effective performance in areas such as social network analysis and urban mobility event detection.

SubsCoRe is a term denoting distinct advanced methodologies across multiple domains, each addressing a specific challenge through carefully constructed algorithmic and statistical strategies. The acronym commonly refers to either "Contrast Subgraph Mining from Coherent Cores" in graph mining, "Sub-spectrogram Segmentation" in environmental sound classification, "Core-elements Subsampling" for alternating least squares in large-scale recommender systems, or "Subspace-in-Confident-Region" adaptive observation cost control in variational quantum eigensolvers. Each variant employs unique theoretical formulations, optimization routines, and practical strategies. This article systematically presents the methodology, mathematical foundations, optimization procedures, theoretical guarantees, empirical validations, and operational considerations for the most prominent instance: "Contrast Subgraph Mining from Coherent Cores" as introduced in (Shang et al., 2018).

1. Formal Problem Statement and Mathematical Definitions

SubsCoRe, in the context of contrast subgraph mining, addresses the detection of node-subsets whose edge structures differ markedly between two weighted graphs that share a common vertex set. Formally, consider undirected, non-negatively weighted graphs $G_A=(V,E_A)$ and $G_B=(V,E_B)$ defined over identical node sets $V$, with $E_A(u,v), E_B(u,v)\geq 0$ denoting edge weights. Optionally, a seed set of nodes $\mathrm{seeds}\subseteq V$ and a neighborhood radius $r\in\mathbb{N}$ are provided, leading to an $r$-neighborhood:

$$N_r(S) = \{u\in V \mid d_A(u,S) \leq r \text{ or } d_B(u,S)\leq r\},$$

where $d_A(u,S)$ is the shortest path length in $G_A$ from $u$ to $S$.

• A coherent core $c$ satisfies $$\mathrm{seeds}\subseteq c \subseteq N_r(\mathrm{seeds})$$, maximizing a similarity-based coherence score:

$$\mathrm{Coherence}(S) = \frac{\sum_{u < v,\, u,v\in S} \mathrm{coherence}(u,v)}{\sum_{u\in S}\mathrm{penalty}(u)}.$$

• A contrast subgraph $g$ is any superset $c\subseteq g\subseteq N_r(c)$, selected to maximize a difference-based contrast score:

$$\mathrm{Contrast}(S) = \frac{\sum_{u < v,\, u,v\in S} \mathrm{contrast}(u,v)}{\sum_{u\in S}\mathrm{penalty}(u)}.$$

In the canonical instantiation, edgewise scores are set as follows: $$\mathrm{coherence}(u,v) = \min\{E_A(u,v), E_B(u,v)\}$$, $\mathrm{contrast}(u,v) = |E_A(u,v)-E_B(u,v)|$, and $\mathrm{penalty}(u)=1$. The problem reduces to nested maximizations:

• Find the coherent core: $$\hat{c} = \arg\max_{\mathrm{seeds}\subseteq c\subseteq N_r(\mathrm{seeds})} \mathrm{Coherence}(c)$$.
• Find the contrast subgraph: $$\hat{g} = \arg\max_{c\subseteq g\subseteq N_r(c)} \mathrm{Contrast}(g)$$.

2. Algorithmic Framework and Optimization Procedure

The SubsCoRe method operationalizes these definitions through a two-phase, max-flow/min-cut approach:

Phase A: Coherent Core Extraction

• Restrict candidate cores $c$ to $$\mathrm{seeds}\subseteq c \subseteq N_r(\mathrm{seeds})$$ if seeds are specified, otherwise search globally.
• Employ binary search over candidate coherence thresholds; for each test, reduce to a min-cut problem in an auxiliary flow network, ensuring polynomial-time exactness.

Phase B: Contrast Subgraph Identification

• For fixed $c$, restrict candidate subgraphs $g$ to $c\subseteq g\subseteq N_r(c)$.
• Analogously, perform binary search on the contrast score. Each feasibility test is converted to a single min-cut in a directed flow network, constructed as follows:
  • Nodes: source $S$, sink $T$, $N_r(c)$.
  • Edges: (S→$u$) for $u\in c$ with infinite capacity; (S→$u$) for $u\notin c$ with large capacity $U$; ($u$→$T$) with capacity $U+2\cdot\mathrm{mid}\cdot\mathrm{penalty}(u)-d(u)$, where $d(u) = \sum_{v} \mathrm{contrast}(u, v)$.
  • For each $(u,v)$ with positive contrast, bidirectional arcs with capacity $\mathrm{contrast}(u, v)$.
• For threshold $\delta$, define:

$$h_\delta(g) = \sum_{u\in g} [2\delta\,\mathrm{penalty}(u) - d(u)] + \sum_{u\in g, v\notin g} \mathrm{contrast}(u, v).$$

The min-cut on this network yields a set $g$ that minimizes $h_\delta(g)$, and feasibility amounts to $h_\delta(g)\leq 0$.

• Iterate until the interval for mid converges to desired tolerance $\delta_\mathrm{tol}$.

This binary search plus min-cut reduction achieves a fully polynomial-time exact solution.

3. Theoretical Properties and Computational Complexity

• The number of nodes in the flow network is $|N_r(c)|+2$, with edge count $O(|N_r(c)|^2 + |E_A| + |E_B|)$.
• Each min-cut computation requires $O(nm)$ time (e.g., Orlin’s algorithm), with $n=|N_r(c)|$ and $m$ edges.
• Binary search over the feasible interval requires $O(\log(\text{input-size}))$ iterations, as the step-size is halved until reaching $\delta_\mathrm{tol}$.
• In the worst case ($N_r(c)=V$):

$$O\bigl((|V|+|E_A|+|E_B|) + (|V|\cdot (|V|+|E_A|+|E_B|))\cdot \log(\text{input-precision})\bigr)$$

Scalability is thus polynomial in the total graph size and suffices for large graphs (tens to hundreds of thousands of nodes/edges).

4. Empirical Validation and Application Scenarios

SubsCoRe was systematically validated across diverse large-scale, real-world scenarios:

Application Area	Data	Key Features of SubsCoRe output
Collaboration Change Detection	DBLP (coauthor graphs, ≈7000 nodes)	Seeds select e.g. “Jiawei Han”; core captures long-term collaborators, contrast subgraph distinguishes epoch-specific collaborators (e.g., Jing Gao pre/post 2009), optimal contrast(g)=8.99
Spatio-Temporal Event Detection	Beijing taxi network (148k nodes)	Seeds on urban arteries; coherent core identifies persistently busy roads, contrast subgraph pinpoints event-specific regions (e.g., concert traffic), contrast ≈ 23.9
E-commerce Trend Detection	Amazon product hierarchy (14k nodes)	Seeds in specific categories select enduringly popular nodes, contrast subgraph highlights transient spikes (e.g., new game releases), contrast(g)=1.44

Across all settings, the method highlights meaningful, temporally local, or contextually relevant structural contrasts, corroborated by external event knowledge.

5. Parameters, Tuning Strategies, and Practical Considerations

• Neighborhood Radius ($r$): Enforces locality; typical $r=1$–$2$ on social graphs, $r=10$–$20$ for city-scale mobility graphs.
• Penalty function ($\mathrm{penalty}(u)$): Modulates subgraph size selection; set to $1$ for uniform weighting in experiments.
• Seeds: Constrain search to a region of interest for interpretable, targeted contrast queries; fully unsupervised runs (seeds $=\varnothing$) recover globally maximal coherent/contrast subgraphs.
• Binary search tolerance ($\delta_\mathrm{tol}$): Set to granularity of edge-weights; smaller values yield more precise solutions at additional computational cost.
• Scalability: Empirically tested on graphs of order $10^5$ nodes/edges; larger problems require only polynomial additional time.

6. Broader Impact and Domain-Specific Use Cases

SubsCoRe's contrast-mining framework unifies local structural similarity and dissimilarity in a maximization schema that is adaptable to multiple domains:

• Social Networks: Detects evolving communities, new or dissolving collaborations, and abrupt regime changes.
• Urban Mobility/Event Detection: Isolates localized surges in movement or network flow, actionable in traffic management and anomaly detection.
• E-Commerce/Taxonomy Trends: Pinpoints time-sensitive spikes or declines in user interest, enabling rapid trend tracking.
• Abnormal Substructure Discovery: Identifies anomalous, event-triggered subgraphs, vital for forensic analysis or fraud detection.

The generality and interpretability of the coherence/contrast separation, coupled with exact and scalable optimization, make SubsCoRe foundational for temporal network analysis, comparative structure mining, and high-resolution change-point detection across scientific and business analytics.

---

For extended details and implementation reference, see "Contrast Subgraph Mining from Coherent Cores" (Shang et al., 2018).