Personalized Recommendations for Authors in Research Communities of the RDGraph

This project, developed by the DARElab team at the ATHENA Research Center, is part of the FAIRCORE4EOSC European project. Our goal is to enhance the discovery of research within the RDGraph by providing personalized recommendations to authors across different research communities.

Methodology

The RDGraph organizes research products into distinct communities (subgraphs), each representing a unique research topic. Our approach centers on analyzing citation edges within the RDGraph to discover relationships between authors. We extract these edges to construct an author-paper interaction matrix for each research community. This matrix is then processed using a collaborative filtering recommender system, which generates personalized recommendations for authors within each community.

Database

[TODO]

Components Overview

This section details the essential elements of this repository, each integral to the functionality and operation of our recommendation system:

• database/rec_data.sql: This SQL script configures the PostgreSQL schema for the recommender system. It establishes the database foundation necessary for processing and generating personalized recommendations by performing key operations:

  • Creates a new schema named recsys_schema and defines an ENUM type for interaction types (authorship and cited) that categorizes the nature of interactions within the database.
  • Creates and populates the interactions table for tracking author-product interactions by community, using ORCID identifiers, product identifiers, and linked foreign keys to the main tables.
  • Optimizes query performance through materialized views and indices. This includes materialized views to summarize interaction frequencies and a table to store and rank recommendations for authors.
  • Creates triggers and functions to automate the filling of the recommendations table with bibliographic details.

• database/rec_data.py: This Python script provides the functionality needed to interact with the the database schema defined in rec_data.sql. It supports data retrieval, processing, and insertion for generating personalized recommendations. It contains several key functions:

  • get_citations_by_community and get_authorships_by_community: Retrieve citation and authorship interactions from the database for a specific community, returning a DataFrame with relevant data.
  • prepare_recommendation_data: Prepares data tuples for database insertion, organizing author IDs and recommended result IDs with corresponding ranks and community acronyms.
  • write_recommendations: Inserts recommendation records into the database based on prepared data tuples.
  • get_recommendations_by_author: Fetches and returns detailed recommendations for a specific author, including titles, types, publication dates, and publishers, organized by community and rank.

• recommenders: This directory contains the available recommendation algorithms.

• train.ipynb: A Jupyter notebook for training the recommender system across different research communities, leveraging the data extracted from the RDGraph. Note that we currently use only interactions of the type cited to train a recommender.

• communities: This directory contains subfolders for each research community. Each subfolder contains the results of optimizing the recommender and the final results with the optimal parameters in json files.

• main.py: Serves as the entry point for the application, managing API interactions with the database that stores the recommendations.

Execution Steps

Follow these steps to set up and run the recommendation system:

1. Initialize the Database Schema

   • Run the SQL file database/rec_data.sql to setup the recsys_schema. This step creates the necessary database structures like tables, views, and functions for the recommender system. It also populates the database with data on author-product interactions across various research communities, which will be used to train the recommendation algorithms.

2. Train the Recommender System

   • Open and run the Jupyter notebook train.ipynb to train the recommender system of your choice. Each community has its own model that is trained with its respective data.