ResiGateNet for Depression Detection

Title

Who

Cynthia Zhang (czhang91), Zhangqi Liu (zliu258), Tongzhao Liu (tliu121), Jinyu Wang (jwang628)

Introduction

In our project, we aim to tackle the pressing issue of detecting depression and suicidal ideation through social media content. We will be performing classification on the “Suicide and Depression Detection” dataset from Kaggle, which comprises Reddit posts labeled for potential suicidal thoughts. We want to build upon the paper “MGL-CNN: A Hierarchical Posts Representations Model for Identifying Depressed Individuals in Online Forums” by integrating depth-enhancing capabilities of Residual Networks (ResNet) with the Multi-Gated LeakyReLU CNN. The gating mechanism allows the model to focus on the most relevant features, and the residual connections capture the long-range dependencies in textual contexts. The goal is to build an innovative model, “ResiGateNet”, that has the capacity to understand the nuanced language indicative of depression and suicidal ideation, providing a critical tool for the early detection of mental health issues and timely intervention.

Related Work

Language Modeling with Gated Convolutional Networks: https://arxiv.org/pdf/1612.08083.pdf

Blog on Residual Gated Graph Convolutional Networks: https://wandb.ai/graph-neural-networks/ResGatedGCN/reports/A-Brief-Introduction-to-Residual-Gated-Graph-Convolutional-Networks--Vmlldzo1MjgyODU4

ResNet: https://arxiv.org/abs/1512.03385

Summary: [TODO]

Implementations of the "Language Modeling with Gated Convolutional Networks" paper:

https://github.com/jojonki/Gated-Convolutional-Networks

https://github.com/anantzoid/Language-Modeling-GatedCNN

data

The "Suicide and Depression Detection" dataset that contains labeled Reddit posts:

https://www.kaggle.com/datasets/nikhileswarkomati/suicide-watch

There are 232074 posts, and the two classes are balanced. The data should not require significant preprocessing except for text preprocessing on the posts, such as tokenization and stop-words removal.

Methodology:

(This might be subject to change) What is the architecture of your model?

Multi-Gated LeakyReLU CNN model with residual connections

How are you training the model? [TODO] If you are implementing an existing paper, detail what you think will be the hardest part about implementing the model here. [TODO] If you are doing something new, justify your design. Also note some backup ideas you may have to experiment with if you run into issues.

The gating mechanism allows the model to focus on the most relevant features, and the residual connections capture the long-range dependencies in textual contexts.

Metrics

(This might be subject to change) What experiments do you plan to run?

Comparative Analysis: Evaluating the performance of the proposed Multi-Gated LeakyReLU CNN with Residual Networks models against baseline models like Bag of Words (BoW) combined with SVM or MNB, and other deep learning models such as LSTM or standard CNNs. Feature Sensitivity: Evaluating how different types of input features (e.g., text embeddings, n-gram features) influence the model's effectiveness.

For most of our assignments, we have looked at the accuracy of the model. Does the notion of “accuracy” apply for your project, or is some other metric more appropriate?

The notion of “accuracy” apply for this project.

If you are implementing an existing project, detail what the authors of that paper were hoping to find and how they quantified the results of their model.

The authors of the paper aimed to develop a more effective method for identifying depressed individuals in online forums, improving upon the limitations of existing depression detection models. Specifically, they sought to capture critical sentiment information from the vast volume of posts published by each user, which is challenging due to the nuanced and diverse nature of language indicative of depression.

Because of imbalance of data, the authors use Precision, Recall, F-1 score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC)

If you are doing something new, explain how you will assess your model’s performance. Since the dataset we are using is balanced, accuracy should be a valid reflection of the model's performance, however, we will also be using Precision, Recall, F-1 score, and Area Under the Receiver Operating Characteristic Curve (AUC-ROC) so that we can compare it to other models.

What are your base, target, and stretch goals? Achieve significantly high recall without substantially sacrificing precision. This would indicate the model's capability to identify most depressed users with minimal false positives, which is critical for practical applications of depression detection in online forums.

Ethics

What broader societal issues are relevant to your chosen problem space?

Our project aims to detect early signs of depression on social media, providing early support avenues for those reluctant to seek help. Yet, it raises privacy concerns and risks related to mislabeling and the potential stigma for users. We're committed to ethical data use, prioritizing user consent, and anonymity.

Why is Deep Learning a good approach to this problem?

Deep learning is well-suited for depression detection in social media due to its ability to learn complex patterns from large volumes of unstructured text data. The nuanced language of mental health, often subtle and embedded in personal narratives, requires the kind of intricate model architectures that deep learning provides. These models can discern latent patterns and sentiments in language that may indicate depression, beyond what simpler statistical or rule-based methods can achieve. Furthermore, deep learning models can improve over time with more data, adapt to the evolving nature of online communication, and personalize detection based on individual language use, making them a robust choice for this application.

Division of labor:

In our team of four, we have decided to adopt a collaborative approach where all members contribute equally to each part of the project. Our collective efforts will span across all key areas: