ReDepress Dataset

Note: To access this dataset please fill the ReDepress_agreement.docx document and send an email to [email protected]

The ReDepress Dataset originates from the paper
"ReDepress: A Cognitive Framework for Detecting Depression Relapse from Social Media".

This dataset is designed to support research on detecting depression relapse using social media text.
It provides rich temporal user data and cognitive bias annotations, allowing systems to monitor conversational dynamics and make timely inferences.

---

Overview

• Total Users: 204
• File Format: .parquet files (one per user)
• Ground Truth File: ground_truth_labels.csv
• agreement to access dataset: ReDepress_agreement.docx

Each user is assigned a binary label in the ground_truth_labels.csv file:

• 0 → No Relapse
• 1 → Relapse

The dataset contains two categories of users:

1. Relapsed Users
2. Non-relapsed Users

Each user file includes a chronological sequence of posts, containing the user’s own submissions.

---

Data Structure

1. User Files (.parquet)

Each user file contains the following fields:

Column	Description
post_id	Unique string identifier for the post
author	Anonymous identifier of the user who created the post
date	ISO 8601 timestamp of post creation
selftext	Main content (body text) of the submission
title	Title summarizing the submission
avg_attention_bias	Mean of all three annotators’ attention bias scores
avg_interpretation_bias	Mean of all three annotators’ interpretation bias scores
avg_memory_bias	Mean of all three annotators’ memory bias scores
avg_rumination	Mean of all three annotators’ rumination bias scores
majority_attention_bias	Majority label for attention bias
majority_interpretation_bias	Majority label for interpretation bias
majority_memory_bias	Majority label for memory bias
majority_rumination	Majority label for rumination bias

---

2. Label Mapping

The majority label for each cognitive bias was mapped to numerical values as follows:

Memory Bias

Label	Mapped Value
Positive	1
Negative	-1
No Bias	0

Attention Bias

Label	Mapped Value
Positive	1
Negative	-1
No Bias	0

Interpretation Bias

Label	Mapped Value
Positive	1
Negative	-1
No Bias	0

Rumination

Label	Mapped Value
Reflection	1
Brooding	-1
No Rumination	0

---

Ground Truth Labels

The ground_truth_labels.csv file contains the final gold standard relapse annotations:

Column	Description
author	Anonymous user identifier
relapse_label	0 for no relapse, 1 for relapse

---

Use Cases

Researchers can use this dataset to:

• Train and evaluate relapse detection models based on text data.
• Study cognitive bias patterns preceding depression relapse.
• Analyze temporal evolution of users’ social media language and behavior.

---

Citation

If you use this dataset in your research, please cite:

ReDepress: A Cognitive Framework for Detecting Depression Relapse from Social Media
https://arxiv.org/pdf/2509.17991

---

Ethical Considerations

• All user data is anonymized to protect privacy.
• The dataset must be used solely for research purposes related to mental health, language, and computational social science.
• Researchers are encouraged to handle the data with sensitivity and adhere to ethical AI and mental health research guidelines.

---

License

To access this dataset please fill the ReDepress_agreement.docx document and send an email to [email protected]

---

Contact

For questions or collaboration inquiries, please refer to the contact information provided in the ReDepress paper.