zyunnn / startup_prediction

Startup Prediction

This repository contains solution for HKUST-UBS 'China Startup Prediction' Project from Team (Andrew and Zhiyun).

1. How to use

1.1 Setup

conda create -n startup_prediction python=3.6
source activate startup_prediction
pip install -r requirements.txt

1.2 Training

To train and evaluate model, run:

cd src/model
python <INSERT FILENAME>

For example, to train and evaluate lightGBM model, run:

cd src/model
python multi_lightgbm.py 

2. Methodology

2.1 Feature Engineering

Some features extracted are:

• Company overview
  • includes sector, management team, location
• Funding event
  • includes normalized funding amount, funding round code, number of investors
• Investor
  • includes top investors
• News
  • includes positive and negative sentiment of news

2.2 Model

The models implemented are as follows:

• LightGBM
  • as baseline model
• Temporal Convolutional Network (TCN)
  • with dilations, causal network and skip connections

3. Implementation

3.1 Modules and Repository Structure

• Codes are separated into three main modules: data_loader,model,data.
• Subections in docs specifically addresses the criterias of the evaluation rubics.
  • For Design - quality of background research, please refer to resources.md and model_description.md
  • For Code - organization of code, please refer to preprocessing_file_organization.jpeg and file_description.md

4. Results

The following results are computed as precision, recall and f1-score for test set:

• LightGBM
  • precision: 0.7652
  • recall: 0.5787
  • f1-score: 0.6076
• Temporal Convolutional Network (TCN)
  • precision: 0.8495
  • recall: 0.8539
  • f1-score: 0.8517

5. Future Work

• Experiment with additional graph propagation layer (similar to Graph Convolutional Network[1]) added between TCN block for feature propagation
• Finetune TCN

References

[1] T. N. Kipf and M. Welling, “Semi-supervised classification with graph convolutional networks.”