PAT: A Personality-Driven Augmentation and Transfer Learning Framework for Depression Detection on Social Media

Shengwu Zheng; Yuzhe Ding; Xiaohui Liang; Fei Li; Chong Teng; Donghong Ji

cs.CLOct 10, 2026

PAT: A Personality-Driven Augmentation and Transfer Learning Framework for Depression Detection on Social Media

Authors: Shengwu Zheng, Yuzhe Ding, Xiaohui Liang, Fei Li, Chong Teng, Donghong Ji

Organizations: School of Cyber Science and Engineering, Wuhan University · School of Public Health, Wuhan University

Abstract

Depression is a prevalent mental disorder that affects millions worldwide, with early detection crucial for effective intervention. While exist?ing methods have achieved promising results in automated depression detection, they face two key limitations: (1) noise in post screening: symptom-based semantic screening is easily misled by superficial lexical patterns, thereby introducing false-positive posts; (2) weak op?timization of the post encoder: relying solely on coarse user-level supervision signals pre?vents models from capturing depressive cues within individual posts. To address these limita?tions, we propose PAT, a Personality-driven Augmentation and Transfer learning frame?work. In the first phase, PAT optimizes a post?level encoder via a fine-grained contrastive learning objective, utilizing LLM-based data augmentation driven by classic personality the?ories to alleviate data scarcity. In the second phase, the trained post-level scorer is deployed for high-quality post filtering, while its en?coder parameters are seamlessly transferred for user-level tasks. Extensive experiments demon?strate that PAT significantly outperforms exist?ing methods. Furthermore, PAT delivers multi?granular interpretability by tracing post-level mental-state trajectories and highlighting key symptoms, thereby offering reliable diagnostic evidence for clinical practice.

Figures & tables

Figure 1: A typical Screen-and-Classify framework for social media user-level depression detection. The hierarchical model is trained solely on binary supervision (depressed or non-depressed), which prevents the post encoder from capturing fine-grained depressive cues in each post and thus limits its optimization. Moreover, the screening step may introduce false positives. For instance, the word “cry” in the fourth sentence of the figure expresses a positive emotion, misleading the model.

Figure 2: An overview of the PAT framework: it first performs post-level optimization and then transfers the post scorer’s encoder to the user-level task for screening and encoding. For post-level optimization, PAT introduces Personality-Driven Augmentation and novel Fine-grained Contrastive Learning . Unlike standard supervised contrastive learning, it distinguishes three data types derived from an anchor post: (1) Augmented Posts, generated under personality guidance, serving as strong positives; (2) Same-Label Posts, which share the ground-truth label, serving as weak positives; and (3) Different-Label Posts, which have different labels, serving as negatives.

Methods	Dataset	F1	Precision	Recall	AUC
HAN	eRisk2017	0.570 $\pm$ 0.013	0.564 $\pm$ 0.030	0.575 $\pm$ 0.036	0.879 $\pm$ 0.002
DeCapsNet		0.579 $\pm$ 0.013	0.563 $\pm$ 0.038	0.586 $\pm$ 0.024	0.891 $\pm$ 0.010
E2LPS		0.575 $\pm$ 0.009	0.559 $\pm$ 0.018	0.594 $\pm$ 0.040	0.884 $\pm$ 0.010
Mental-Bert		0.629 $\pm$ 0.012	0.595 $\pm$ 0.034	0.667 $\pm$ 0.018	0.891 $\pm$ 0.002
GPT-5.4		0.632 $\pm$ 0.009	0.702 $\pm$ 0.016	0.575 $\pm$ 0.028	-
Qwen3.5-Plus		0.527 $\pm$ 0.014	0.389 $\pm$ 0.024	0.815 $\pm$ 0.036	-

Table 1: Main results of experiments in the user-level depression detection.

Method	eRisk2017	eRisk2018	SWDD
	F1	F1	F1
PAT	0.640	0.665	0.866
w/o T	0.597	0.593	0.834
w/o Scorer	0.625	0.649	0.812
w/o PA	0.611	0.641	0.843

Table 2: The results of the ablation study

Figure 3: Experimental results (F1 score) of PAT as a plug-in module on baseline methods.

Figure 4: A case study demonstrating the comprehensive interpretability of the PAT framework.

Method	Data	Acc	F1
BaseModel	ORI	0.725	0.713
BaseModel	AUG	0.733	0.719
BaseModel + SCL	AUG	0.732	0.717
BaseModel + FGCL	AUG	0.751	0.738

Table 3: Results of post-level depression detection. ORI and AUG denote the original dataset and the augmented dataset, respectively. SCL and FGCL represent Supervised Contrastive Learning and Fine-grained Contrastive Learning.

Panel of Figure 5 — Figure 5: Visualization of the embedding space of different contrastive learning methods.

Type	SemCons	StyCons
BIG5-0	4.46 $\pm$ 0.92	4.86 $\pm$ 0.13
BIG5-1	4.32 $\pm$ 0.79	4.74 $\pm$ 0.32
BIG5-2	4.31 $\pm$ 0.91	4.70 $\pm$ 0.38
BIG5-3	4.38 $\pm$ 1.03	4.68 $\pm$ 0.37
BIG5-All	4.36 $\pm$ 0.91	4.74 $\pm$ 0.31
MBTI-0	3.93 $\pm$ 1.72	4.47 $\pm$ 0.63

Table 4: Data quality assessment. BIG5-

n

denotes the augmented sentences derived from original posts labeled with a depression severity of

n

using the Big Five-driven scheme. BIG5-All represents the complete set of all augmented sentences generated via the Big Five framework

Metric	BIG5	MBTI	UNGUIDED
Self-BLEU-4 $\downarrow$	0.018	0.101	0.151
Distinct-1 $\uparrow$	0.388	0.377	0.334
Distinct-2 $\uparrow$	0.857	0.812	0.725
Distinct-3 $\uparrow$	0.965	0.935	0.872

Table 5: Linguistic diversity of different augmentation schemes. Lower Self-BLEU-4 and higher Distinct-

n

scores indicate greater diversity. Bold values denote the best result for each metric.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

Figure A.1: The prompt used by large language model methods in the user-level task

Panel of Figure 12 — Figure A.2: Prompt templates for the two-step personality-driven augmentation procedure. (a) Symptom identification and symptom-level severity estimation based on the 21 BDI-II dimensions. (b) Generation conditioned on the original post and the identified symptom–severity pairs, using BIG5 or MBTI personality-style guidance.

	Train	Test	Val
No. of posts(None)	1922	218	210
No. of posts(Slight)	1193	151	158
No. of posts(Moderate)	1107	135	144
No. of posts(Serious)	670	107	100
Sum	4892	611	612

Appendix

Table A.1: Label Distribution and Data Split of the Post-Level Depression Severity Classification Dataset.

From	Trait	Content
Original		I’m always the one comforting others, so why is no one comforting me?
BIG5	High Extraversion	I’m always cheering my friends up, so why does no one care about how I feel?
	Low Extraversion	Comforting others is so exhausting; I need to be cared for too.
	High Agreeableness	I am willing to listen to everyone’s troubles, but I also hope someone can understand my feelings.
	Low Agreeableness	Why is it always me comforting others? Who cares about my pain?
	High Conscientiousness	I’m used to helping others solve their problems, but I should also learn to seek support for myself.

Appendix

Table A.2: Augmented data examples. Shown here is the English translation of the original Chinese data.

	SWDD			eRisk 2017			eRisk 2018
	Train	Test	Val	Train	Test	Val	Train	Test	Val
No. of Users (Dep)	2,227	742	742	81	27	27	128	43	43
No. of Users (Con)	11,716	3,905	3,905	451	151	150	896	299	298
No. of Total Users	13,943	4,647	4,647	532	178	177	1,024	342	341

Appendix

Table A.3: Label Distribution and Data Split of the User-Level Datasets.

	Symptom	Descriptions
1	Sadness	I feel sad.
2	Pessimism	I am discouraged about my future.
3	Past Failure	I always fail.
4	Loss of Pleasure	I don’t get pleasure from things.
5	Guilty Feelings	I feel quite guilty.
6	Punishment Feelings	I expected to be punished.

Appendix

Table A.4: Symptom descriptions adopted for screening and the symptom-attention module.

Figure A.3: Sensitivity Analysis of Parameter $K$

Figure A.4: A case of a user labeled as non-depressed.

Figure A.5: A case of a depressed user that was missed in detection.

Figure A.6: A case of a user labeled as non-depressed misclassified as depressed.

Evaluation metric	SWDD	eRisk2018
Top-1 symptom precision	90.0%	61.1%
Top-2 symptom precision	89.0%	57.2%
Top-3 symptom precision	87.3%	52.6%

Appendix

Table A.5: Human evaluation of textual support for the top-ranked symptom explanations produced by PAT.

Explore similar work

Psychologically-Grounded Graph Modeling for Interpretable Depression Detection
Apr 27, 2026Rishitej Reddy Vyalla, Kritarth Prasad, Avinash Anand +4Depression Detection Dynamic Graph Learning
IIIT Delhi, New Delhi, India · Singapore Institute of Technology, Singapore · Nanyang Technological University, Singapore +3
Dep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor with Reliable LLM Reasoning
Jun 9, 2026Yiqing Lyu, Xianbing Zhao, Buzhou Tang +1LLM Inference Depression Detection
School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, Guangdong, China · School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China, Harbin Institute of Technology, Shenzhen, Guangdong, China, and Guangdong Provincial Key Laboratory of Intelligent Information Processing · School of Computer Science and Technology, Harbin Institute of Technology, Shenzhen, Guangdong, China, Pengcheng Laboratory, Guangdong Provincial Key Laboratory of Intelligent Information Processing +1
WPG-MoE: Weak-Prior-Guided Dense Mixture-of-Experts for User-Level Social Media Depression Detection
Jul 5, 2026Xian Li, Yuanhe Tian, Yang Yang +2Social Media Analysis Depression Detection
University of Electronic Science and Technology of China · Zhongguancun Academy · University of Science and Technology of China

Apr 27, 2026J/K move · Enter open · S save

Psychologically-Grounded Graph Modeling for Interpretable Depression Detection

Rishitej Reddy Vyalla, Kritarth Prasad, Avinash Anand, Erik Cambria+3

IIIT Delhi, New Delhi, India · Singapore Institute of Technology, Singapore · Nanyang Technological University, Singapore+3

Depression Detection Dynamic Graph Learning Causal Interpretability

Open paper AskPDF arXiv

Automatic depression detection from conversational interactions holds significant promise for scalable screening but remains hindered by severe data scarcity and a lack of clinical interpretability. Existing approaches typically rely on black-box deep learning architectures that struggle to model the subtle, temporal evolution of depressive symptoms or account for participant-specific heterogeneity. In this work, we propose PsyGAT (Psychological Graph Attention Network), a psychologically grounded framework that models conversational sessions as dynamic temporal graphs. We introduce Psychological Expression Units (PEUs) to explicitly encode utterance-level clinical evidence, structuring the session graph to capture transitions in psychological states rather than mere semantic dependencies. To address the critical class imbalance in depression datasets, we employ clinically approved persona-based data augmentation, enable robust model learning. Additionally, we integrate session-level personality context directly into the graph structure to disentangle trait-based behavior from acute depressive symptoms. PsyGAT achieves state-of-the-art performance, surpassing both strong graph-based baselines and closed-source LLMs like GPT-5, achieving 89.99 and 71.37 Macro F1 scores in DAIC-WoZ and E-DAIC, respectively. We further introduce Causal-PsyGAT, an interpretability module that identifies symptom triggers. Experiments show a 20% improvement in MRR for identifying causal indicators, effectively bridging the gap between depression monitoring and clinical explainability. The full augmented dataset is publicly available at https://doi.org/10.6084/m9.figshare.31801921.

Abstract

Explore similar work

Psychologically-Grounded Graph Modeling for Interpretable Depression Detection

Dep-LLM: Training-Free Depression Diagnosis via Evidence-Guided Structured Multi-factor with Reliable LLM Reasoning

WPG-MoE: Weak-Prior-Guided Dense Mixture-of-Experts for User-Level Social Media Depression Detection