Cooperative Multimodal Approach to Depression Detection in Twitter

Authors: Tao Gui, Liang Zhu, Qi Zhang, Minlong Peng, Xu Zhou, Keyu Ding, Zhigang Chen110-117

AAAI 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results demonstrate that the proposed method outperforms state-of-the-art methods by a large margin (over 30% error reduction). In several experiments and examples, we also verify that the selected posts can successfully indicate user depression, and our model can obtained a robust performance in realistic scenarios.
Researcher Affiliation Collaboration 1Shanghai Key Laboratory of Intelligent Information Processing, Fudan University 2i Flytek Co., Ltd.
Pseudocode Yes Algorithm 1 COMMA for Depression Detection
Open Source Code Yes We release all the codes and results for further research.
Open Datasets Yes Textual Depression Dataset D1. Shen et al. (2017) constructed a textual depression dataset on Twitter. Inspired by the work reported in (Coppersmith, Dredze, and Harman 2014), the authors labeled users as depressed if their anchor tweets satisfied a strict pattern, i.e., (I m/ I was/ I am/ I ve been) diagnosed depression.
Dataset Splits Yes With 1,402 depressed users in total, we randomly selected 1,402 non-depressed users in D1 to make the percentage of depressed users equal to 50%, but we did so in a more difficult manner by removing all the anchor tweets. After obtaining the dataset, we trained and tested these methods using five-fold cross validation.
Hardware Specification No The paper does not provide specific details about the hardware used for running experiments (e.g., CPU/GPU models, memory).
Software Dependencies No The paper mentions software like GRU, VGG-Net, Adam, and scikit-learn, but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes We initialized the word embeddings and other parameters related to the deep learning models by randomly sampling from a standard normal distribution and a uniform distribution in [-0.05, 0.05], respectively. We set the dimensionality of the word embedding to 128. In addition, we used one layer of bidirectional GRU to model the post text, and set the hidden neurons of each GRU to 64. To extract image features, we applied VGG-Net, and mapped the output of the first fully connected layer to a vector with a dimensionality of 128. Each of text and image selectors use a GRU with 50 hidden neurons. Our model can be trained end-to-end with backpropagation, and we performed gradient-based optimization using the Adam update rule (Kingma and Ba 2014), with a learning rate of 0.001.