This repository contains our work for the Kaggle challenge the Advanced Learning for Text and Graph data (ALTEGRAD) course : https://www.kaggle.com/c/altegrad-2021/overview

Work by :

• Akesbi Kamil
• Beauté Rémy
• Dutot Grégoire

Link prediction is the problem of predicting the existence of a link between two entities in a network.

The problem of link prediction has recently attracted a lot of attention in many domains. For instance, in social networks, one may be interested in predicting friendship links among users, while in biology, predicting interactions between genes and proteins is of paramount importance.

In this challenge, we had to deal with the problem of predicting whether a research paper cites another research paper. More specifically, we were given a citation network consisting of several thousands of research papers, along with their abstracts and their list of authors. The pipeline that is typically followed to deal with the problem is similar to the one applied in any classification problem, the goal is to use edge information to learn the parameters of a classifier and then to use the classifier to predict whether two nodes are linked by an edge or not.

In this repository, we present the main approaches we tested for the MVA Altegrag 2022 Kaggle challenge. The goal is to predict wheter a research paper cites another research paper. We embedded information about the papers thanks to unsupervised framework. We also implemented a GNN encoder to produce more fine-tuned embeddings. These embeddings are finally fed to a Multilayer Perceptron that predicts the probability of two nodes being linked to one another. We achieve quite good performance on the prediction task and observe the efficiency of fixed-size neighborhood sampling to adapt the GAE framework to large graphs.

More details about our work can be found in the reports.

pip -r install requirements