Introduction

The source code of AAAI2023 paper "Orders Are Unwanted: Dynamic Deep Graph Convolutional Network for Personality Detection". The document describes how to reproduce the experiment result. We also provide a Chinese version document.

Requirements

• PyTorch version >= 1.7
• transformers version >= 2.9.0

Datasets and post-trained BERT are available in Google Drive and Baidu Netdisk (verification code: 96ap). Preprocessing files are provided in folders pandora and kaggle, respectively.

• Pandora MBTI datasets are not public datasets, please contact the author on the offical website for permission.

Main Results

There are 2 datasets, Kaggle and Pandora MBTI datasets for our evaluations.

Example cmd for running experiments on Kaggle

python main.py \
    --all_gpu_train_batch_size 8 \
    --all_gpu_eval_batch_size 32 \
    --num_train_epochs 25.0 \
    --task kaggle \
    --l0

apply --l0 to enable L0 constraint on L2C, corresponding to Eq(15) in the paper, otherwise the training objective function is cross-entroty loss.

all_gpu_train_batch_size is set to 8 for Kaggle, and 4 for Pandora.

Best checkpoint is available in Google Drive.

Ablation Experiments

Reproduing some ablation experiments need to modify source code, we will improve it in the future work.

• single-hop: apply --single_hop in cmd
• remove special nodes(-u): --no_special_node in cmd
• undirected graph: change False to True for line 35 in graphmask.py
• switch to GCN/GAT: change self.dgcnx to Dynamic_GCN/Dynamic_GAT for Multi_DGCN in model_utils.py
• no DART: apply --no_dart in cmd