temporal-gcn-lstm

Code for Characterizing and Forecasting User Engagement with In-App Action Graphs: A Case Study of Snapchat

Temporal-gcn-lstm model encodes temporal evolving action graphs to predict future user engagement. The end-to-end, multi-channel neural model also encodes acitivity sequences and other macroscopic features to reach best performance.

Requirements

DGL, NetworkX, PyTorch, Pandas, Numpy, SciKit-Learn, tqdm

Deep Graph Library (DGL) https://www.dgl.ai/

Pytorch https://pytorch.org/

Building action graphs

build_graphs.py: build static graphs for time period

build_temporal.py: build temporal graphs per day

python3 build_graphs.py INPUT_PATH OUTPUT_PATH

python3 build_temporal.py INPUT_PATH OUTPUT_PATH

Models

utils.py: supporting functions

activity_seq_model.py: baseline activity sequence model

gcn_model.py: model structure of our graph convolutional network

multi_channel.py: To run our final best performance temporal graph model

To Run

python3 multi_channel.py

Load custom data with df_path graphs_path macro_path flags

Set variants of model with --activity --macro flags to inlcude or leave out these information. ex. --activity False. Default for both are True for best enhanced performance of model.

Hyperparameters were set to optimal for our dataset, they can be modified as input arguments.