Deep Incomplete Multi-View Learning Network with Insufficient Label Information

This repo contains the code of our AAAI24 paper Deep Incomplete Multi-View Learning Network with Insufficient Label Information. To tackle the double missing of features and labels problem, we propose a novel Deep Incomplete Multi-view Learning Network (DIMvLN) by incorporating graph networks and semi-supervised learning in this paper.

Requirements

• scikit-learn==1.1.1
• scipy==1.8.1
• torch==1.12.0+cu116
• torch-cluster==1.6.0+pt112cu116

To install the package in an existing environment, run

pip install -r requirements.txt

or to create a conda environment for running this repo, run

conda env create -f environment.yaml

Content

• classification.py: Main function to run incomplete multi-view semi-supervised classification.
• config/config_with_gcn_completion.py: contains the hyper-parameters (e.g., $\alpha$ and $\lambda_1$) and the training options (e.g., structure of network and learning rate).
• models/model.py: Class of the DIMvLN can tackle the dataset with 2 to 6 views.
• models/model_coach.py: Class of the pipeline to run the training process.

Usage

Before you start the experiment, please make sure that the constant MAIN_DIR in the models/utils/data_loader.py script points to the path of the parent folder of the data folder.

To run the incomplete multi-view semi-supervised classification with missing rate=50% and unlabeled rate={95%, 90%, 85%}, simply type:

>> cd <Path_to_DIMvLN_code_folder>
>> python classification.py --dataset 0 --device 0 --info_freq 20 --test_time 10 --missing_rate 0.5 --unlabeled_rate 0.95,0.9,0.85 --save_result y

where --dataset is the selected dataset index; --info_freq means the information update frequency; --test_time means the number of independent runs.

Start pre-train GCN...
pre_train_loss in 0: 0.4602
pre_train_loss in 1: 0.3740
pre_train_loss in 2: 0.3520
pre_train_loss in 3: 0.3194
...
Epoch
**********************
 1/200    Train: L_rc==>1.0239 L_icl==>-0.0053 L_ccl==>0.3493 L_cel==>2.2867 ACC==>0.473 || Validation: Loss==>0.0000 ACC==>0.505
 20/200   Train: L_rc==>0.2678 L_icl==>-0.0053 L_ccl==>0.8352 L_cel==>0.0105 ACC==>0.835 || Validation: Loss==>0.0000 ACC==>0.809
 40/200   Train: L_rc==>0.2703 L_icl==>-0.0053 L_ccl==>0.8783 L_cel==>0.0066 ACC==>0.948 || Validation: Loss==>0.0000 ACC==>0.903
 ...
 180/200  Train: L_rc==>0.2610 L_icl==>-0.0053 L_ccl==>0.8835 L_cel==>0.0001 ACC==>0.999 || Validation: Loss==>0.0000 ACC==>0.948
 200/200  Train: L_rc==>0.2612 L_icl==>-0.0053 L_ccl==>0.8940 L_cel==>0.0001 ACC==>0.999 || Validation: Loss==>0.0000 ACC==>0.942
**********************
>>>>> Best validation Accuracy score:
     0.9502617801047121 best_acc
>>>>> Test(1/10) Report with best validation Accuracy model parameters:
Accuracy score:  0.9205020920502092
Macro-Precision score:  0.8545993164857755
Macro-F1 score:  0.8227701485690322

Reference

If you find our work useful in your research, please consider citing:

@inproceedings{jiang2024deep,
  title={Deep Incomplete Multi-View Learning Network with Insufficient Label Information},
  author={Jiang, Zhangqi and Luo, Tingjin and Liang, Xinyan},
  booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
  volume={38},
  number={11},
  pages={12919--12927},
  year={2024}
}