PaperlayerView exact revision on arXiv

Source-linked AI summary

Knowledge-Preserving Incremental Social Event Detection via Heterogeneous GNNs

Yuwei Cao, Hao Peng, Jia Wu, Yingtong Dou, Jianxin Li, Philip S. Yu

arXiv:2101.08747v2cs.LG

TL;DR

Social event detection must handle complex, changing streams, while existing incremental methods capture limited knowledge and may forget what they learned. KPGNN models messages with heterogeneous graphs and GNNs, then uses contrastive learning and incremental maintenance; experiments report better performance than various baselines, including gains over EventX in incremental evaluation.

  • Problem

    Existing incremental clustering and community-detection methods capture limited social-data knowledge and cannot memorize previously learned information, despite the complexity and streaming nature of messages.

    Source: §1 · p. 1

  • Method

    KPGNN organizes social-message elements in heterogeneous graphs, uses GNNs to learn representations, and applies contrastive learning to support incremental detection and maintenance.

    Source: §1 · p. 2, §3.2 · p. 4

  • Results

    KPGNN outperforms various baselines on the evaluated corpora; in incremental evaluation, its average relative gain over EventX is 16% in NMI.

    Source: §1 · p. 2, §4.3 · p. 9

  • Takeaways & Limitations

    KPGNN continuously detects events and extends its knowledge using dynamic social streams, and the authors report superiority over baselines in their experiments.

    Source: §6 · p. 12

    The task assumes each social message belongs to at most one event.

    Source: §2 · p. 3

Abstract

from arXiv · show

Social events provide valuable insights into group social behaviors and public concerns and therefore have many applications in fields such as product recommendation and crisis management. The complexity and streaming nature of social messages make it appealing to address social event detection in an incremental learning setting, where acquiring, preserving, and extending knowledge are major concerns. Most existing methods, including those based on incremental clustering and community detection, learn limited amounts of knowledge as they ignore the rich semantics and structural information contained in social data. Moreover, they cannot memorize previously acquired knowledge. In this paper, we propose a novel Knowledge-Preserving Incremental Heterogeneous Graph Neural Network (KPGNN) for incremental social event detection. To acquire more knowledge, KPGNN models complex social messages into unified social graphs to facilitate data utilization and explores the expressive power of GNNs for knowledge extraction. To continuously adapt to the incoming data, KPGNN adopts contrastive loss terms that cope with a changing number of event classes. It also leverages the inductive learning ability of GNNs to efficiently detect events and extends its knowledge from previously unseen data. To deal with large social streams, KPGNN adopts a mini-batch subgraph sampling strategy for scalable training, and periodically removes obsolete data to maintain a dynamic embedding space. KPGNN requires no feature engineering and has few hyperparameters to tune. Extensive experiment results demonstrate the superiority of KPGNN over various baselines.

1 INTRODUCTION

The paper frames social event detection as an incremental-learning problem because social streams are complex and changing, while existing approaches do not adequately acquire, preserve, and extend knowledge. KPGNN combines heterogeneous graph modeling, GNNs, and contrastive learning to detect events and maintain knowledge as messages arrive.

1 INTRODUCTION

  • Social event detection clusters co-related messages from streams, whose volume, varied content, short text, and changing semantics make the task challenging.Social events can provide insights for timely responses, including in crisis management and product recommendation.Source: §1 · p. 1
  • Existing incremental clustering and community-detection methods use limited statistical features, while a GNN-based alternative assumes a fixed output space and requires retraining for new data.The paper identifies these limitations as leaving incremental social event detection unresolved.Source: §1 · p. 2
  • KPGNN organizes varied social-stream elements into heterogeneous graphs and uses GNNs to extract semantic and structural knowledge that its parameters preserve.It resumes training on new messages to extend that knowledge, while contrastive losses address changing event classes and mini-batch subgraph sampling supports scale.Source: §1 · p. 2
  • KPGNN achieves better performance than various baselines in experiments on a large-scale Twitter corpus and an event detection corpus.The corpora are publicly available.Source: §1 · p. 2
  • The paper formalizes social event detection in an incremental-learning setting and presents KPGNN as a heterogeneous GNN model for that task.The authors state that, to their knowledge, this is the first use of GNNs for incremental social event detection.Source: §1 · p. 2

2 NOTATIONS AND PROBLEM FORMULATION

This section defines the social-stream and event-detection task in terms of message blocks and models that extend prior knowledge. It also introduces notation used for the graph, embeddings, and training procedure.

2 NOTATIONS AND PROBLEM FORMULATION

  • A social stream is a continuous temporal sequence of message blocks, each containing messages arriving within a time interval.A message comprises its document, associated users, and timestamp.Source: §2 · p. 2
  • A social event is a set of correlated messages discussing the same real-world happening, with each message assigned to at most one event..Source: §2 · p. 3
  • Event detection maps a message block to its set of events, while incremental detection uses a sequence of models to process blocks over a maintenance window.Each later model extends its predecessor by depending on the predecessor’s parameters.Source: §2 · p. 3
  • The notation includes the message graph, its adjacency matrix and initial features, GNN embeddings, mini-batches, sampled neighbors, and contrastive-loss terms.The algorithm takes a social stream, available labels, window size, GNN layers, and mini-batch count as inputs.Source: §2 · pp. 3–4

3 METHODOLOGY

KPGNN incrementally detects events by organizing social messages into graphs, learning message representations, and periodically updating its model. Its design combines graph-based knowledge retention with contrastive training, scalable sampling, and clustering.

3.1 Continuous Detection Framework

  • KPGNN’s lifecycle pretrains an initial model, detects events from incoming message blocks, then removes obsolete messages and resumes training on recent data.The maintained model is used for detection in the next window.Source: §3.1 · p. 4

3.2 Heterogeneous Social Message Modeling

  • The model builds a heterogeneous graph from messages, words, named entities, and users, then maps shared-element links into a message-only graph.Message features concatenate average pretrained word embeddings with two-dimensional timestamp encodings.Source: §3.2 · pp. 4–5

3.3 Knowledge-Preserving Incremental Message Embedding

  • Graph-attention message representations fuse language semantics, temporal information, and graph structure, while learned parameters retain event-detection-oriented knowledge.The attention design supports evolving graphs and previously unseen messages.Source: §3.3 · p. 5

3.4 Scalable Training via Contrastive Learning

  • Contrastive training combines triplet and global-local pair losses, uses mini-batch subgraph sampling, and resumes from existing parameters using recent data.Triplet sampling can use labeled messages without full labeling; unlabeled messages still contribute through representation propagation, while the pair loss requires no labels.Source: §3.4 · pp. 5–6

3.5 Message Clustering

  • At detection, KPGNN clusters learned message representations and outputs the clusters as social events; DBSCAN is suitable when the number of classes is unspecified.The text also identifies K-Means as a usable distance-based alternative.Source: §3.5 · p. 6

3.6 Time Complexity of KPGNN

  • KPGNN’s overall running time is O(N_e), where N_e is the number of edges in the message graph.The analysis also gives sampling cost O(∏_{l=1}^L c_l), typically much smaller than N_e.Source: §3.6 · p. 6

4 EXPERIMENTS

The experiments compare KPGNN with offline and incremental event-detection baselines, then examine maintenance strategies and hyperparameter sensitivity.

4 EXPERIMENTS

  • The experiments compare KPGNN with offline and incremental social event-detection models.The section also studies forgetting strategies during maintenance and sensitivity to KPGNN hyperparameters.Source: §4 · p. 6
  • The study investigates how different forgetting strategies affect the maintenance stage of KPGNN’s lifecycle.Source: §4 · p. 6
  • The experiments include a sensitivity analysis of KPGNN’s hyperparameters.Source: §4 · p. 6

4.1 Experiment Setup

The experiments use Twitter and MAVEN datasets, compare KPGNN with representation, similarity, offline-detection, and incremental-detection baselines, and evaluate cluster agreement with NMI, AMI, and ARI.

4.1.1 Datasets.

  • The Twitter dataset contains 68,841 manually labeled tweets across 503 event classes over four weeks; MAVEN is a general-domain event-detection dataset.After filtering, MAVEN contains 10,242 messages across 154 event classes.Source: §4.1.1 · p. 6

4.1.2 Baselines.

  • Baselines include message-representation and similarity methods, offline PP-GCN, incremental EventX, and KPGNN without its global-local pair loss.The ablation KPGNNₜ uses only the triplet loss.Source: §4.1.2 · p. 6

4.1.3 Experiment Setting.

  • Experiments repeat each setting five times and report mean and standard variance; incremental evaluation uses the latest-message strategy.KPGNN uses a three-block maintenance window and 2,000-message mini-batches; several baselines require the ground-truth number of classes.Source: §4.1.3 · p. 6

4.1.4 Evaluation Metrics.

  • Performance is evaluated by NMI, AMI, and ARI, which measure agreement between detected and ground-truth clusters.AMI and ARI adjust for chance.Source: §4.1.4 · p. 7

4.2 Offline Evaluation

In offline evaluation, KPGNN outperforms message-embedding and similarity methods across all metrics on both datasets, and ranks at or near the top against event-detection baselines.

4.2 Offline Evaluation

  • On Twitter and MAVEN, KPGNN exceeds message-embedding and similarity methods across all metrics, with EventX higher only in Twitter NMI.EventX is much lower in AMI and ARI; KPGNN scores highest or second-highest across all metrics.Source: §4.2 · p. 8
  • On Twitter, KPGNN’s margins over message-embedding and similarity methods span 8−141% in NMI, 4−1,200% in AMI, and 29−2,100% in ARI.The comparison covers Word2vec, LDA, BERT, BiLSTM, and WMD.Source: §4.2 · p. 8
  • On MAVEN, KPGNN outperforms general message-embedding methods (Word2vec, LDA, BERT, BiLSTM) and WMD by 13−49% in NMI, 73−375% in AMI, and 150−900% in ARI.The passage attributes this to KPGNN leveraging message semantics and structure, while comparison methods rely on message-element distributions or embeddings and ignore underlying social graph structure.Source: §4.2 · p. 8

4.3 Incremental Evaluation

In incremental evaluation on a Twitter stream, KPGNN consistently outperforms baselines across message blocks and metrics, with gains reported against EventX, WMD, BERT, and PP-GCN.

4.3 Incremental Evaluation

  • KPGNN consistently outperforms baselines across all message blocks; against EventX, average gains are 16% in NMI, 319% in AMI, and 589% in ARI.The reported ranges are 6−27%, 164−676%, and 68−1782%, respectively.Source: §4.3 · p. 9, Table 6 · p. 9
  • Against PP-GCN, KPGNN’s average gains are 47% in NMI and 53% in AMI, with gains up to 58% in ARI.The authors say KPGNN preserves up-to-date knowledge while PP-GCN can be distracted by obsolete information as messages accumulate.Source: §4.3 · p. 9
  • Against WMD, KPGNN’s average gains are 22% in NMI, 23% in AMI, and 26% in ARI.The authors attribute the advantage to KPGNN using social-stream structure ignored by WMD.Source: §4.3 · p. 9
  • KPGNN generally outperforms KPGNNₜ, supporting the positive effect of adding the global-local pair loss term.Source: §4.3 · p. 9

4.4 Study on update-maintenance strategies

The latest-message strategy achieves the strongest performance among the tested update-maintenance strategies while using less training time and GPU memory.

4.4 Study on update-maintenance strategies

  • The all-message strategy retains every received message, but is described as impractical because the graph can slow the model and exceed encoder capacity.Source: §4.4 · p. 9
  • The relevant-message strategy removes old messages unrelated to recent arrivals, while preserving their learned knowledge in model parameters.Source: §4.4 · p. 10
  • The latest-message strategy has the strongest performance, despite discarding past messages and retaining only knowledge learned from them.The latest and relevant-message strategies consistently, and the all-message strategy in most blocks, outperform PP-GCN.Source: §4.4 · p. 10
  • The latest-message strategy requires significantly less training time and GPU memory than the other strategies.Its message graph remains lightweight by keeping only the latest message block.Source: §4.4 · p. 10

4.5 Hyperparameter Sensitivity

KPGNN’s NMI and AMI vary only slightly with window and mini-batch sizes, while ARI fluctuations do not clearly track either setting. Overall, the model is insensitive to these hyperparameter changes.

4.5 Hyperparameter Sensitivity

  • Figure 4 compares NMI, AMI, and ARI across message blocks M2, M8, M14, and M20 under different window and mini-batch sizes.Its colors indicate value fluctuations, with sunken areas in blue and convex areas in green.Source: Figure 4 · p. 11
  • KPGNN is insensitive to window and mini-batch size changes: NMI and AMI vary slightly, while ARI fluctuations show no clear relationship to either setting.NMI and AMI standard deviations are 0.01–0.02; block-wise average ARIs range from 0.58–0.59 across window sizes and 0.57–0.59 across mini-batch sizes.Source: §4.5 · pp. 10–11
  • Window sizes 1 and 3 yield block-wise average NMIs of 0.75 each, versus 0.74 each for window sizes 9 and 11.Source: §4.5 · p. 11
  • Mini-batch sizes 1000, 2000, 3000, and 4000 each yield a block-wise average NMI of 0.75.Source: §4.5 · p. 11

5 RELATED WORK

Prior social event detection work includes document-pivot and feature-pivot approaches, as well as offline and online methods. KPGNN is a document-pivot method that continuously adapts to social messages and is described as the first to use GNN inductive learning for incremental social event detection.

Social Event Detection

  • Document-pivot methods cluster social messages by correlation, whereas feature-pivot methods cluster message elements such as words and named entities by distribution.Source: §5 · p. 11
  • Offline methods support retrospective, domain-specific event analysis, while online methods continuously operate on dynamic social streams.Source: §5 · p. 11
  • Incremental clustering and community detection methods learn limited knowledge from social streams and have too few parameters to preserve learned knowledge.A GCN-based method works only offline.Source: §5 · p. 11
  • KPGNN continuously adapts to incoming social messages to acquire, extend, and preserve knowledge, and is described as the first to use GNN inductive learning for incremental social event detection.Inductive learning means generalizing to unseen nodes.Source: §5 · pp. 11–12

6 CONCLUSION

The conclusion presents KPGNN as a knowledge-preserving approach to incremental social event detection that uses social-message semantics and structure. The authors report superiority over baselines and identify event evolution and causal discovery as future directions.

6 CONCLUSION

  • KPGNN incorporates social-message semantics and structure for knowledge acquisition, continuously detects events, and extends knowledge from dynamic streams.Source: §6 · p. 12
  • Experiments demonstrate KPGNN’s superiority over baselines.Source: §6 · p. 12
  • Future work could extend KPGNN to social event analysis, including event evolution, and causal discovery in social data.Source: §6 · p. 12

Related papers

Loading 2101.08747v2…