Context-Compatible Information Fusion for Knowledge Graphs

This repository belongs to our TPDL2020 publication Context-Compatible Information Fusion for Knowledge Graphs. The repository provides our scripts and links to our precomputed data. The project builds upon SemMedDB as a medical knowledge graph. Therefore, you need to setup SemMedDB first. Besides, all results are reported in an Excel sheet (Experiments on SemMed2019.xlsx).

Setting up SemMedDB

We use SemMedDB as a PostgreSQL database. The following steps will help you to setup SemMedDB.

Downloading SemMedDB

Download the predication table from SemMed NIH. We use SemMedDB in version semmedVER40_R. You can directly download the predication table at Predication. Rename the file to predication.sql

Setup Postgres

We use SemMedDB as a PostgreSQL V10 database in our experimental setup and scripts. Details for PostgreSQL can be found at PostgreSQL We recommend to setup a PostgresSQL database. We name the database "semmeddb". We have written a script to convert the MySQL SQL-statements to PostgreSQL (located in sql directory).

python3 sql/semmed_postgres_converter.py predication.sql predication.psql

Subsequently, we use "psql" to load our sql scripts. Create the predication table and load the data into it.

psql -d semmeddb < sql/semmed_create_table.sql 
psql -d semmeddb < sql/predication.psql

Finally, create indexes and materialized views for our experiments (this may take a while).

psql -d semmeddb < sql/semmed_create_index.sql
psql -d semmeddb < sql/materialized_views.sql

Experiments

The experiments belonging to our first experiment (knowledge graph without contexts and with strict contexts) can be found in 01-KG_Context.ipynb. Experiments belomnging to our second experiment (context compatibility) can be found in 02-Context_Compatibility.ipynb. To rerun our experiments, you need to setup a python environment.

Setting up Python

We provide Jupyter notebooks for our experiments. You can run them in an Jupyter environment with the following packages installed:

pip install psycopg2-binary nltk scikit-learn matplotlib

Library Graph (Strict Implicit Context) vs Knowledge Graph

For the first experiment, you need to enter your database credentials for semmeddb in the beginning of the Jupyter notebook.

Context Compatibility

For the second experiment, you need access to the metadata of PubMed. To bypass setting up a metadata database, you can find our precomputed metadata simiarity values and clusterings at OneDrive. Download the zip archiv and unzip it. The files should be stored as an additional directory (inside this repository) called:

preprocessed_context_compatibility

Next, you can rerun the second notebook without having metadata database credentials entered. This line ensures that the metadata processing will be skipped.

do_preprocessing = False

Thanks for reading!