ink-usc / temporal-gcn-lstm Goto Github PK

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

Python 100.00%

temporal-gcn-lstm's Introduction

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.

Cite

@inproceedings{liu2019characterizing,
  title={Characterizing and forecasting user engagement with in-app action graph: A case study of snapchat},
  author={Liu, Yozen and Shi, Xiaolin and Pierce, Lucas and Ren, Xiang},
  booktitle={Proceedings of the 25th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  pages={2023--2031},
  year={2019}
}

temporal-gcn-lstm's People

Contributors

Stargazers

Watchers

temporal-gcn-lstm's Issues

df_path,macro_path

I have some confuse about these two files.
Are they .pkl file?
What are the contents of this file
I will appreciate for your reply.Thanks

example data

example data,thanks !

my silly question

I'm curious why networkx is already used and dgl is added

Action Graph

Do we have a code in this repo for making an action graph with its probabilities calculation on its edges? (Like Fig1 and Fig3 in your paper)

mock_data seem to be empty

Thank you so much for sharing these codes. However, I was trying to run the multi_channel.py script. All seem to be okay except that the graph_data in the path "mock_data/example_temporal_user_action_graphs.pkl", when loaded, the graphs are empty (zero nodes).

It could be the way I loaded the nx DiGraphs is not correct,

I have used the following,

parser.add_argument('--df_path', type=str,default='mock_data/example_df.pkl')
parser.add_argument('--macro_path', type=str,default='mock_data/example_macro_df.pkl')
parser.add_argument('--graphs_path', type=str,default='mock_data/example_temporal_user_action_graphs.pkl')

Then I tried three types of loading pickled graphs:

1 - graphs_sep = pickle.load(args.graphs_path)
This gives an error "*** TypeError: file must have 'read' and 'readline' attributes"

2 - from networkx.readwrite.gpickle import read_gpickle
graphs_sep = read_gpickle(args.graphs_path)
This seems to read the graph dict of [2 x7] correctly. But when converting the nx_graph to dgl, the graphs are empty. *** networkx.exception.NetworkXError: Graph has no nodes or edges

with open(args.graphs_path, 'rb') as f:
        graphs_sep = pickle.load(f)

This also correctly reads the graph dict. However, same error is produced. *** networkx.exception.NetworkXError: Graph has no nodes or edges

activity sequence

Hi,
In the multi_channel.py module one of the inputs is:

Input datasets
    df: dataframe with userid as index, contains labels, and activity sequence if needed

I am having trouble to understand how exactly would the activity sequence would look like in a dataframe.
can you maybe provide an example?

thank you!