Research map / Graph learning / Knowledge and temporal graphs
Knowledge-Enhanced Hierarchical Heterogeneous Graph for Personality Identification with Limited Training Data
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a knowledge-enhanced hierarchical heterogeneous graph with POS, entity, and LIWC knowledge plus user-level attention and multi-label contrastive loss for Big Five personality identification with only 100 training posts.
Problem
Personality identification from social media text when labeled personality data is scarce (only 100 training samples, about 1% of data).
Approach
Combines global multi-view graphs over words, LIWC categories, POS tags, entities, and texts with a per-post post–word–knowledge heterogeneous graph processed by GCN, then optimizes representations with user-level self-attention and a label-level multi-label contrastive loss.
Key results
- Average Macro-F1 of 60.89% on PANDORA vs 57.19% for best baseline Semi-PerGCN, with 100 training samples.
- Average Macro-F1 of 55.26% on PAN2015 and 50.01% on MyPersonality, outperforming baselines by reported margins.
- Removing external knowledge reduces PANDORA average F1 by 9.40%; removing LIWC nodes causes the largest drop among knowledge types (5.19%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use the knowledge-enhanced hierarchical heterogeneous graph method instead of Semi-PerGCN | Personality identification · PANDORA | Macro-F1 (%): 57.19 -> 60.89 | Helped |
| Use the knowledge-enhanced hierarchical heterogeneous graph method instead of IML-CL | Personality identification · PAN2015 | Macro-F1 (%): 52.42 -> 55.26 | Helped |
| Use the knowledge-enhanced hierarchical heterogeneous graph method instead of CNN | Personality identification · MyPersonality | Macro-F1 (%): 48.87 -> 50.01 | Helped |
| Add the knowledge subgraph (ablation) | Ours · PANDORA | Macro-F1 (%): 55.17 -> 60.89 | Helped |
| Add user-level attention (ablation) | Ours · PANDORA | Macro-F1 (%): 58.77 -> 60.89 | Helped |
| Add contrastive learning (ablation) | Ours · PANDORA | Macro-F1 (%): 59.29 -> 60.89 | Helped |
| Add POS tag nodes (ablation) | Ours · PANDORA | Macro-F1 (%): 59.49 -> 60.89 | Helped |
| Add entity nodes (ablation) | Ours · PANDORA | Macro-F1 (%): 58.46 -> 60.89 | Helped |
| Add LIWC nodes (ablation) | Ours · PANDORA | Macro-F1 (%): 57.74 -> 60.89 | Helped |
| Use BERT rather than GloVe or one-hot vectors for initial word embeddings (ablation) | Ours · PANDORA | Macro-F1 (%) | Helped |
| Use GloVe rather than one-hot vectors for initial word embeddings (ablation) | Ours · PANDORA | Macro-F1 (%) | Helped |
| Increase contrastive learning weight λ beyond 0.01 (ablation) | Ours · PANDORA | Macro-F1 (%) | Hurt |
Surprising findings
- POS tag knowledge had minimal effect on performance, attributed to BERT word embeddings already encoding grammatical structure.
- Plain BERT and textCNN are competitive with graph-based baselines such as BertGCN and SHINE, which perform poorly in this low-data personality setting.
Limitations
- Computational cost may be challenging for larger datasets or real-time applications.
- Performance on PAN2015 and MyPersonality is lower, attributed to short, casual, lower-quality posts.
- Relies on external resources (NELL entities, LIWC, POS tagger) and multiple graph constructions, adding complexity.
Experimental setup
Datasets: PANDORA PAN2015 MyPersonality
Benchmarks: PANDORA PAN2015 MyPersonality
Metrics: Macro-F1
Baselines: textCNN BERT BertGCN SHINE Semi-PerGCN Prototypical Network DCKPN IML-CL
Builds on: SHINE BertGCN Semi-PerGCN IML-CL TransE LIWC
Closest papers by meaning
- Data Augmented Graph Neural Networks for Personality Detection (AAAI 2024 · similarity 0.90)
- Leveraging the Dual Capabilities of LLM: LLM-Enhanced Text Mapping Model for Personality Detection (AAAI 2025 · similarity 0.85)
- LLM vs Small Model? Large Language Model Based Text Augmentation Enhanced Personality Detection Model (AAAI 2024 · similarity 0.85)
- Ask, Answer, and Detect: Role-Playing LLMs for Personality Detection with Question-Conditioned Mixture-of-Experts (NeurIPS 2026 · similarity 0.84)
- PersonaX: Multimodal Datasets with LLM-Inferred Behavior Traits (ICLR 2026 · similarity 0.83)
- Hypergraph Representation Learning with Hyperlink Random Effects (NeurIPS 2026 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.