This is the source code for paper xxx.
Given start-ups' high-risk and high-reward nature, identifying the ones that will eventually succeed is literally a million-dollar question for practitioners in the $63-billion venture capital industry and for policy makers worldwide, especially at an early stage such that investment returns can be exponential, and policies can better guide and promote the innovation ecosystem for long-term economic growth.
Although various empirical studies and data-driven modeling work have been done, the predictive power of complex networks of stakeholders including venture capital investors, start-ups, and start-ups' managing members has not been thoroughly explored. We design an effective graph representation learning model where node embeddings are incrementally updated by unsupervised graph self-attention and optimized with fine-tuning by supervised link prediction and node classification. Our model uses network structures, temporal dependencies among time periods, and rich node-level attributes for success prediction. Overall, our method achieves superior performance on a real dataset of global venture capital investments, almost twice as human investors. In addition, our model excels at prediction for start-ups in industries such as healthcare and IT. Meanwhile, we shed light on the impacts on start-up success from observable factors including gender, education, and networking, which can be of value for practitioners as well as policy makers when they screen ventures of high growth potential.
The code has been successfully tested in the following environment. (For older PyG versions, you may need to modify the code)
- Python 3.8.12
- PyTorch 1.11.0
- Pytorch Geometric 2.0.4
- Sklearn 1.0.2
- Pandas 1.3.5
We provide samples of our data in the ./Data folder. The input of our model is as follows:
graph_edgesincludes the edges of each time step. The shape is [Time_num x 2 x Edge_num]. Time_num is the number of time steps. Edge_num is the number of the edge in this time step.edge_dateis the time step corresponding to each edge and the length is equal to the number of all edges.edge_typeis the edge type corresponding to each edge and the length is equal to the number of all edges.all_nodesis the number of nodes.new_companiesis the index of the newly added node at each time. The shape is [(Time_num - 1) x new_add_node_length].labelsis the label of the newly added node at each time. The shape is [(Time_num - 1 ) x new_add_node_length].nodetypesis the set of node types corresponding to all nodes.
Node Representation Learning
node_representation_learning.py : File for generating node representations in VC networks by node classification and link prediction tasks
python node_representation_learning.py --embedding_dim 64 --n_layers_clf 3 --train_embed --loss_type 'LPNC'Start-up Success Prediction
startup_success_prediction.py : Code that dynamically updates newly added nodes and predicts the success of startups
python startup_success_prediction.py --dynamic_clf --gpus 'cuda:0'File Statement
Run the node_representation_learning.py file to generate the representation of the nodes and save the embedding in the file Save_model. Then run the startup_success_prediction.py file to make predictions about the success of the startups.
Model/Convs.py contains MGTConvs, which is the layer to update the nodes dynamically. Predict_model in Model/Model.py is the model for startup success prediction.
Please cite our paper if you find this code useful for your research:
citation