DISRL: Dual-Scale Incremental Semi-Supervised Reinforcement Learning for Dynamic Community Tracking

Tianming Zhang; Jinxu Wang; Yixuan Zhang; Zhengyi Yang; Chenyu Hou; Bin Cao; Jing Fan

cs.LGOct 21, 2026

DISRL: Dual-Scale Incremental Semi-Supervised Reinforcement Learning for Dynamic Community Tracking

Authors: Tianming Zhang, Jinxu Wang, Yixuan Zhang, Zhengyi Yang, Chenyu Hou, Bin Cao, Jing Fan

Organizations: College of Computer Science and Software Engineering Zhejiang University of Technology Hangzhou, China · School of Computer Science The University of Sydney Sydney, Australia

Abstract

Dynamic local community detection tracks the evolving community of designated seed nodes across graph snapshots. However, it faces two critical challenges: temporal supervision uncertainty, where sparse and outdated labels fail to provide seed-specific guidance under topology changes, and unbounded error cascades, where sequential policy transfer accumulates exploration errors over long horizons. To address these challenges, we propose DISRL, a Dual-scale Incremental Semi-supervised Reinforcement Learning framework. DISRL introduces a supervision construction module that dynamically generates seed-specific guidance from sparse known communities, and a dual-scale community exploration module that stabilizes policy evolution by decoupling short-term adaptation within temporal windows from long-term parameter consolidation across windows. Furthermore, we construct a dynamic graph benchmark with diverse community evolution scenarios. Extensive experiments demonstrate that DISRL consistently outperforms state-of-the-art baselines under various dynamic settings.

Figures & tables

Figure 1. Overall framework of DISRL, consisting of supervision construction and dual-scale community exploration. The framework passes seed-specific guidance to a community explorer. Fast policy parameters adapt over snapshots within each temporal window; a master parameter consolidates information between windows.

Real-world dynamic dataset
Dataset	$T$	$\overline{\|\mathcal{V}\|}$	$\overline{\|\mathcal{E}\|}$	$\overline{\|\mathcal{C}\|}$	$\overline{Q}$
Football	24	137	653.9	19.8	0.5867
Expansion & shrinkage setting
Amazon	25	18256.2	56573.5	1291.0	0.9301
DBLP	25	132172.3	540405.8	4520	0.9105
Twitter	13	99566.3	1334284.8	2651	0.6313

Table 1. Dataset statistics.

T

denotes the number of snapshots;

\overline{|\mathcal{V}|}

\overline{|\mathcal{E}|}

\overline{|\mathcal{C}|}

, and

\overline{Q}

denote the average numbers of nodes, edges, and communities, and the average modularity ( Newman, 2006 ) , respectively. All averages are taken over snapshots.

Dataset	Static	Prec.	Rec.	F1	Jac.	Dynamic	Prec.	Rec.	F1	Jac.
Real-world dynamic dataset
Football	DISRL	97.00 $\pm$ 0.21	94.32 $\pm$ 0.63	95.24 $\pm$ 0.36	91.33 $\pm$ 0.62	DISRL	97.00 $\pm$ 0.21	94.32 $\pm$ 0.63	95.24 $\pm$ 0.36	91.33 $\pm$ 0.62
	SLRL	91.99 $\pm$ 0.29	88.55 $\pm$ 0.90	87.17 $\pm$ 0.67	81.03 $\pm$ 0.88	CS-TGN	91.02 $\pm$ 4.27	96.22 $\pm$ 0.85	92.35 $\pm$ 1.63	87.53 $\pm$ 3.40
	CLARE	38.86 $\pm$ 1.10	47.49 $\pm$ 1.39	42.21 $\pm$ 1.19	32.02 $\pm$ 1.14	MFC-TopoReg	65.38 $\pm$ 3.40	62.18 $\pm$ 3.20	63.42 $\pm$ 3.28	52.92 $\pm$ 3.74
	ProCom	29.90 $\pm$ 1.58	54.08 $\pm$ 3.00	34.41 $\pm$ 1.63	23.58 $\pm$ 1.42	DevDynaP	88.11	96.14	91.80	86.17
	SLSS	82.09	90.00	85.51	77.42	QTCS	89.59	84.51	86.72	83.08

Table 2. Main comparison on simulated evolutionary settings and a real-world dynamic dataset. All metrics are reported in percentages (%). Best and second-best results are marked in bold and underlined within each comparison group, respectively.

Variant	Real-world dynamics	Expansion&shrinkage		Merging&splitting
Variant	Football	Amazon	DBLP	Amazon	DBLP
w/o cc	93.42 $\pm$ 0.27 / 87.94 $\pm$ 0.44	88.02 $\pm$ 0.12 / 80.32 $\pm$ 0.16	75.89 $\pm$ 0.13 / 67.39 $\pm$ 0.23	81.48 $\pm$ 0.52 / 71.40 $\pm$ 0.87	52.54 $\pm$ 1.18 / 40.81 $\pm$ 1.40
w/o rcs	94.23 $\pm$ 0.30 / 90.18 $\pm$ 0.49	88.28 $\pm$ 0.15 / 80.63 $\pm$ 0.19	76.53 $\pm$ 0.24 / 67.67 $\pm$ 0.32	82.86 $\pm$ 0.44 / 73.35 $\pm$ 0.31	53.58 $\pm$ 1.08 / 41.65 $\pm$ 1.24
w/o sim	88.29 $\pm$ 1.86 / 81.95 $\pm$ 2.15	87.28 $\pm$ 1.00 / 79.27 $\pm$ 1.42	75.29 $\pm$ 1.71 / 66.80 $\pm$ 1.66	81.76 $\pm$ 1.94 / 71.46 $\pm$ 1.95	53.25 $\pm$ 2.24 / 41.72 $\pm$ 2.49
w/o win	92.97 $\pm$ 0.46 / 88.20 $\pm$ 0.62	87.14 $\pm$ 0.54 / 78.92 $\pm$ 0.64	74.81 $\pm$ 0.41 / 65.07 $\pm$ 0.76	79.96 $\pm$ 1.14 / 69.68 $\pm$ 1.17	49.07 $\pm$ 1.65 / 37.66 $\pm$ 1.53
w/o decay	94.86 $\pm$ 0.30 / 90.67 $\pm$ 0.49	88.53 $\pm$ 0.46 / 81.17 $\pm$ 0.80	77.08 $\pm$ 0.84 / 68.57 $\pm$ 1.12	82.46 $\pm$ 0.36 / 72.84 $\pm$ 0.54	54.98 $\pm$ 1.58 / 43.20 $\pm$ 1.40

Table 3. Ablation results. Each cell reports F1/Jaccard.

Figure 2. Long-horizon tracking performance on Football.

Figure 3. Effect of the retrieved community number

K

. {F1 and Jaccard scores for different retrieval sizes on Football, Amazon, and DBLP.}

Figure 4. Sensitivity to iGPN propagation. {F1 and Jaccard scores on Football, Amazon, DBLP, and Twitter for different propagation depths.}

Figure 5. Sensitivity to temporal window size. {F1 and Jaccard scores on Football, Amazon, and DBLP for window sizes from two to six.}

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Component	Configuration	Value
Supervision Retrieval	Known communities	100,all except target
	Top- $K$	50,5
Community Explorer	Batch size	16
	Epochs per snapshot	50
	Hidden dimension	64
	iGPN layers	3,2

Appendix

Table 4. Hyper-parameter configurations of DISRL.

Figure 6. DBLP case study over four snapshots. Rows show snapshots and columns show methods; each panel reports F1. The target community grows from 7 to 16 nodes. DISRL, ProCom, and CS-TGN predictions over four DBLP snapshots, with true positives, false positives, false negatives, and the target seed identified.

Figure 7. Efficiency comparison under the expansion&shrinkage setting. The reported runtime is normalized per target seed per snapshot over 50 independent queries. Running times for DISRL, CS-TGN, QTCS, SLRL, and SLSS on Amazon, DBLP, and Twitter, plotted on a logarithmic scale.

Dynamic Graph Learning

Dynamic graph learning models how networks and their node features change over time to predict links, events, labels, and system behavior.

See all 124 papers

Diffusion-Induced Spatial Attention Overlapping Community Detection
Sep 22, 2026Kosti Koistinen, Vesa Kuikka, Joni Herttuainen +3Graph Attention Networks Graph Clustering
Aalto University School of Science Computer Science Department P.O.Box 11000, 00076 AALTO, Finland · Lockheed Martin USA
Forget Less, Generalize More: Unifying Temporal and Structural Adaptation for Dynamic Graphs
May 28, 2026Qian Chang, Ciprian Doru Giurcaneanu, Runsong Jia +6Temporal GNNs Graph Representation Learning
University of Auckland, Auckland, New Zealand · University of Technology Sydney, Sydney, Australia · Central China Normal University, Wuhan, China
Accelerating Dynamic Graph Clustering on GPU Architectures with cuGraph
Aug 4, 2026Nelson Aloysio Reis de Almeida Passos, Emanuele Carlini, Salvatore TraniGPU Acceleration Clustering
University of Pisa, Dept. of Computer Science, 56127 Pisa PI, Italy · National Research Council, 56124 Pisa PI, Italy

Sep 22, 2026J/K move · Enter open · S save

Diffusion-Induced Spatial Attention Overlapping Community Detection

Kosti Koistinen, Vesa Kuikka, Joni Herttuainen, Matthew Hendren+2

Aalto University School of Science Computer Science Department P.O.Box 11000, 00076 AALTO, Finland · Lockheed Martin USA

Graph Attention Networks Graph Clustering

Open paper AskPDF arXiv

Detection of overlapping communities is essential for modelling networks in which nodes participate simultaneously in multiple structural or functional groups. Existing graph neural network approaches commonly rely on local message passing, which can obscure community boundaries through smoothing and limit the representation of structurally relevant long-range dependencies. We introduce Diffusion-Induced Spatial Attention Community Detection (DISCO), a deep-learning framework that combines a structural prior derived from influence spreading dynamics, sparse multi-head attention, and non-negative community-affiliation learning. The prior identifies candidate interactions beyond immediate graph neighbours and biases attention according to their structural proximity, while a Bernoulli-Poisson edge-reconstruction objective enables overlapping community inference from node attributes and structural profiles, or both. Benchmark experiments show that DISCO performs competitively against established graph convolutional and graph attention approaches across different input configurations. To demonstrate its practical applicability, we present a proof-of-concept cybersecurity use case in which changes between community assignments inferred from consecutive communication-network snapshots provide an interpretable anomaly signal. Temporal community similarity identifies structural deviations, while node-level contributions help locate the devices associated with them. DISCO therefore provides both a flexible method for overlapping community detection and a foundation for analysing structural change in dynamic networks.

Abstract

Dynamic Graph Learning

Explore similar work

Diffusion-Induced Spatial Attention Overlapping Community Detection

Forget Less, Generalize More: Unifying Temporal and Structural Adaptation for Dynamic Graphs

Accelerating Dynamic Graph Clustering on GPU Architectures with cuGraph