Citation Prediction for NIPS Papers (1987-2017)

This repo includes our codes and report for our project in CSCI-544: Applied Natural Language Processing taught by Prof. May.

Group members:

• Seyed Mohammad Asghari ([email protected])
• Soheil Nazar Shahsavani ([email protected])
• Ahmad Fallahpour ([email protected])
• Mehrdad Kiamari ([email protected])

Project Description

Goal

The goal of this project was to predict the number of citations for the papers published in Neural Information Processing Systems (NIPS) conference, from 1987-2017 based on some information of these papers.

Data collection

We collected information about all the papers published in Neural Information Processing Systems (NIPS) conference, from 1987-2017.

The papers published in each year, are listed on the nips website here. Each of the papers listed for an specific year, contains the following information:

• Title of the paper
• List of authors
• Abstract
• BibTex

One of the challenges of our project was that NIPS website does not have the abstract for some of the papers (all the papers from 1987-2007). Additionally, we need the total number of citations for each paper and the affiliation of the authors. Therefore, we wrote a python script to extract these information from google scholar.

The details of the data collection step can be found in here.

The outputs of this steps are 4 .csv files as follows.

• Author.csv that contains these columns: paper id, #citations, year, Author0, Author1, …
• Abstract.csv that contains these columns: paper id, #citations, year, Abstract
• Affiliation.csv that contains these columns: paper id, #citations, year, Affiliation0, Affiliation1
• Title.csv that contains these columns: paper id, #citations, year, Title

Data processing

We use this IPython Notebook split_train_test.ipynb to divide the data set (the above 4 .csv files) into the following training and test sets:

Training sets:

Author_training.csv, Affiliation_training.csv, Title_training.csv, Abstract_training.csv.

Test sets:

Author_test.csv, Affiliation_test.csv, Title_test.csv, Abstract_test.csv.

Then, we use these two IPython Notebooks author.ipynb and affiliation.ipynb to generate following csv files:

Author_train_predicted.csv, Author_test_predicted.csv, Affiliation_train_predicted.csv, Affiliation_test_predicted.csv.

Next, we use the IPython Notebook title.ipynb to generate following files:

Title_training_predicted.csv, Title_test_predicted.csv.

Next, we use the IPython Notebook abstract.ipynb to generate following files:

Abstract_training_predicted_[method].csv, Abstract_test_predicted_[method].csv,

where mode can be nltk, textblob, or rake.

Classification

We use a variety of models from Scikit-learn package for implementing the classifier. The inputs to this classfication problem are:

Training dataset:

• Author_train_predicted.csv
• Affiliation_train_predicted.csv
• Title_training_predicted.csv
• Abstract_training_predicted_[mode].csv

Test dataset:

• Author_test_predicted.csv
• Affiliation_test_predicted.csv
• Title_test_predicted.csv
• Abstract_test_predicted_[method].csv

where mode can be nltk, textblob, or rake.

The details of the classification step can be found in the IPython Notebook classification.ipynb. In our classification task, since the number of labels is 3, one baseline approach is to choose of the 0, 1, or 2 labels randomly. This approach results in the accuracy of %33.33 for the training and test sets. A better baseline can be obtained by looking at the training set and find the most frequent label and use it for the test. We call this approach, Dumb Classifier.

Todos

• Write the code structure
• ...