Team: Hear Me Roar
- Ashraf GHIYE
- Hussein JAWAD
The goal of this project is to study and apply machine learning techniques to a real-world regression problem. In this regression task, each sample corresponds to a researcher, i.e. an author of research papers, and the goal is to build a model that can predict accurately the h-index of each author. To build the model, we have access to two types of data:
- Co-author Graph that models the collaboration intensity of two researchers: i.e. whether they have co-authored any papers.
- Abstracts of the top-cited papers for each author.
Quantifying success in science plays a key role in guiding funding allocations, recruitment decisions, and rewards. Recently, a significant amount of progresses have been made towards quantifying success in science.
We focus here on using artificielle intelligence to learn a predictive function based on two types of data that can assess the quality of a researcher's work (his abstracts) and the quantity of his work (collaboration intensity).
To get the data of this project you will have to run get_data.sh.
$ chmod +x get_data.sh
$ ./get_data.sh
All the necessary files will be downloaded in data folder, these files include:
abstracts.txt: for each paper, this file contains the ID of the paper along with the inverted index of its abstract. Roughly speaking, an inverted index is a dictionary where the keys correspond to words and the values are lists of the positions of the corresponding words in the abstract. For example, the abstract “that is all that I know” would be represented as follows:
paper_ID----{"IndexLength":5,"InvertedIndex": {"that":[0,3],"is":[1],"all":[2],"I":[4],"know":[5]}}
author_papers.txt: list of authors and the IDs of their top-cited papers.collaboration_network.edgelist: a co-authorship network where nodes correspond to authors that have published papers in computer science venuestrain.csv: 23,124 labeled authors. Each row of the file contains the ID of an author and his/her h-index.test.csv: this file contains the IDs of 208,115 authors unlabelled.paper_embeddings_64.csv: Doc2Vec embeddings of each paper.Text.json: plain texts for each abstract. The format is paperID: abstract in plain text, it is useful to performe NLP downstream tasks on this file like text embeddings.
$ pip install -U -r requirements.txt
We have extracted some new features and created a new weighted graph of collaboration. To get all of these features into features folder you can run the script feature_generators.sh which extract the hand-crafted features explained in the report.
$ chmod +x feature_generators.sh
$ ./feature_generators.sh
After you have run the previous two scripts, you can run model.py to train the model and see some benchmarks scores. For more details about the model and other benchmarks refer to final_report.pdf.
After training the model we can run an inference step is made by running inference.py which generates a .csv prediction for all the authors in the testing set.