srpablino/gcnsm

Graph Convolutional Networks - Schema Matching (GCNSM)

For the dataset similarity experiments (based on meta-features extracted from openML datasets, see this work), and the attribute matching based on the monitor dataset from the DI2KG challenge, you can skip step1 and step2 by downloading the files indicated in the folder step2/output/

If you want to run all steps, you must provide the input-files for the involved datasets in your experiments. For the example experiments (openML and monitor), you can download the input files as indicated in the readme file inside step1/input

Step 1:

Extract metafeatures from datasets with pandas profiling tool You need to create a folder (/step1/input/folder_experiment_name/) with the datasets of your experiment inside

Step 2:

Build graph from the datasets and metafeatures + encoded features of the datasets using the fastText word embeddings model. It uses step1/output/folder_experiment_name/ files.

Step 3:

Use the encoded graph to train a NN which will learn how the input datasets are related, and be able to relate new unseen datasets to the ones you already have. This step requieres that you have the graph with the fasttext-encoded features of your input datasets (see step2/output/readme.txt to download the encoded graph for the example experiments). If you started from step1, this step uses the files generated at: step2/output/folder_experiment_name/

This step also requeres that you have a csv ground_truth (with a header) following the format:
id1,id2,match
xx,yy,1
xx,zz,-1

In this simple example, nodes xx and yy are similar (1), while nodes xx and zz are not similar (-1)

There 3 different cases according to the experiment performed:

hold_out experiment:

You will need to create 2 separated files (train.csv, and test.csv) and put them inside step3/ground_truth/folder_experiment_name/hold_out

cv_10 and random_subsampling

For the 10-fold cross validation and random subsampling experiments, you will need just one file inside step3/ground_truth/folder_experiment_name/experiment_name.csv (or .json etc), which later will be processed to get the train/test splits, according the specific experiment (10-fold cv or random_sub_sampling). Here you need a 4th column, with the topic that would group the nodes that are related:
id1,id2,match,topic
xx,yy,1,topic1
xx,zz,-1,Null

Running the experiment

You can run the experiment step by step, using the corresponding jupyter notebooks inside each step folder