maxiaoba/grape

grape's Issues

I want to use this code for regression analysis only Can I?

Hello @maxiaoba I have a dataframe includes features and target variable as given in your case. Now I want to create a regressor using your code. what changes I need to make, as in your code you filling these missing values and then predicting the output right?
Secondly if my data has already some missing values in it then Can I use your code ? where are you saving the missing values or completely filled feature matrix?

How to run GRAPE with other data?

I have some data and want to use them to run GRAPE.
Is it possible to run GRAPE with data other than UCI and MC?
What is the data format that GRAPE requires as input?

Train with custom dataset

Hi!
I am trying to apply your approach on my dataset with a lot of missing values, stored as NaN values. I am trying to customize your code, but I think that some your suggestions could be useful to speed up my work.

I am currently with two dataframes df_X and df_Y as input parameters, df_X being a feature matrix with numerical and NaN values, df_Y being an array of numbers yi = 1, 2 or 3.

I already tried to use the "get_data" function used for the UCI dataset: the code works, but the training fails (loss is always nan) due to nan values contained in data. I think I should not add the "nan" edges during training, but the lack of comments in code makes it difficult to understand what to change to accomplish this purpose.

Thanks in advance for your help! :)

The reproduced result is not the same as the paper

Do I need to modify the initial hyperparameters for the concrete dataset?

How to get a complete data set after filling？

Requeriments (versions)

Hello!

I'm having difficulties running the code, even though I have all the requirements installed. This is likely a version mismatch.
This is what I am seeing:

File "/usr/PycharmProjects/GRAPE/models/egsage.py", line 55, in forward
    return self.propagate(edge_index, x=x, edge_attr=edge_attr, size=(num_nodes, num_nodes))
  File "/usr/anaconda3/envs/grape/lib/python3.7/site-packages/torch_geometric/nn/conv/message_passing.py", line 235, in propagate
    msg_kwargs = self.inspector.distribute('message', coll_dict)
  File "/usr/anaconda3/envs/grape/lib/python3.7/site-packages/torch_geometric/nn/conv/utils/inspector.py", line 58, in distribute
    raise TypeError(f'Required parameter {key} is empty.')

Which version of pytorch-geometric did you use?

Edge embedding enquiry

Hi there,

I would like to clarify the value of the edge embedding when the value in D (the feature matrix) of that specific edge is missing in the bipartite graph. And what is the network utilised for O_edge to obtain a prediction of D_hat.

Thanks

The reproduced result is not the same as the paper

Do I need to modify the initial hyperparameters for the concrete dataset?

Why use x_j? not x_i in EGraphSage class

Hi, I am learning about grape for data imputation
I think that the paper's pseudo code and python implementation is not matching.
Is it right to use x_i instead of x_j? for consistency of codes?
x_j : source
x_i : destination

please let me known if I understand grape wrong way.

About the edge embedding

As the paper writes, the edge values are either continuous or discrete. But how are the edge values transfered into edge embeddings? what does the edge embedding look like? I dont see this clearly in the paper.
Hoping for the authors kindly reply.

Change the edge dropout rate and the missing data ratio

I‘d like to consult some questions about the code. When I want to change the edge dropout rate and the missing data ratio, which parameter should I change?The default of the '--dropout' or the '--valid' or the '--known'?
Hoping for your kindly reply.

How to use result?

How to use result as it produced different shape from the original input shape

batch training

Hi,

There is no mention of a batch size in the code or paper. Is it possible to do batch wise training for GRAPE?

Thanks,

VR

About train set and test set