Analytic code for the paper: Predicting the longevity of resources shared in scientific publications

Experiment Workflow

1. Preliminary exploration
   Begin with a baseline model using Logistic Regression to understand the initial performance on the dataset.
2. Feature selection
   Apply Lasso regression for feature selection, helping to identify the most impactful features by enforcing sparsity
3. Classic ML Model Exploration
   Test Elastic Net regression, which combines both L1 (Lasso) and L2 (Ridge) regularization, to assess its performance in comparison to the baseline.
4. Experiment with Tobit Model
   Utilize the built-in Tobit model in R to handle censored data and explore how it fits the dataset.
5. Data Preprocessing
   Rebuild the dataset using the scripts provided in the data folder, ensuring that it is formatted correctly for subsequent experiments.
6. Feature Importance Analysis - Random Forest
   Run a Random Forest classifier to evaluate feature importance and compare the results with those obtained from regression models.
7. Feature Importance Analysis - Lasso Regression
   Re-run Lasso regression to calibrate feature importance based on the penalized regression method.
8. Performance Evaluation - Elastic Net
   Evaluate the performance of the Elastic Net regression by calculating its R-squared value to measure how well the model explains the variance in the dataset.
9. Performance Evaluation - Tobit Model with Elastic Net Enhancement
   Introduce the Elastic Net regularization into the Tobit model, combining censoring handling with regularization.
10. Hyperparameter Tuning of the Tobit Model
    Perform hyperparameter tuning on the Tobit model to optimize its performance.
11. In-depth Exploration of Tobit Model
    Conduct a detailed exploration of the Tobit model to understand its behavior and limitations on the dataset.

Code Listing

This section serves as a guide to help you to navigate and understand the project's components. It provides an overview of the project's code organization. In this section, we include detailed explanations of the scripts' purposes in project sub-folders, covering data migration, modeling techniques (such as Lasso, Logistic, and Tobit regression), and analyses for this study.

Folder archived

Scripts utilized during the preliminary exploration phase.

archived/data_migration

• Scripts for migrating data from MongoDB to Spark.

archived/glm

• Modeling scripts that include Lasso, Logistic, and Elastic Net regression.

archived/tobit

• Modeling scripts that utilize Tobit censored regression. The script located in the archived/tobit/spark folder is the Spark implementation of Tobit regression, but this version is currently non-functional. Please do not spend time on it.

archived/R

• R implementation of Tobit censored regression.

Folder iconference_followup_study

Scripts utilized in the paper.

Folder iconference_followup_study/data

• iconference_followup_study/data/datacheck_data.ipynb
  To assess the ratio of alive to dead URLs.
• iconference_followup_study/data/data_proportion_inspection.ipynb
  To analyze the proportion of alive to dead URLs in a spreadsheet format.
• iconference_followup_study/data/data_truncated.ipynb
  To generate a dataset containing only truncated records (selecting for dead URLs).
• iconference_followup_study/data/data_untruncated.ipynb
  To generate a dataset that includes both truncated and untruncated records (selecting for alive and dead URLs). The categorical ordinal feature charset is encoded using a frequency indexer, and the script will standardize the entire dataset.
• iconference_followup_study/data/data_untruncated_charset_viz.ipynb
  In addition to the previous dataset, this script selects the raw URLs along with other variables to analyze the polarity of URL longevity.
  

Folder iconference_followup_study/lasso

• iconference_followup_study/lasso/lasso_truncated.ipynb
  To build a Lasso model using only truncated data (dead URLs).
• iconference_followup_study/lasso/lasso_untruncated.ipynb
  To build a Lasso model using both truncated and untruncated data (all URLs). The categorical ordinal feature charset is encoded with a frequency indexer, and the script will standardize the entire dataset.
• iconference_followup_study/lasso/lasso_untruncated_cleaned.ipynb
  To build a Lasso model using both truncated and untruncated data (all URLs). The categorical ordinal feature charset is encoded using a frequency indexer, and the script will standardize the entire dataset.
• iconference_followup_study/elastic_net/elastic_net_untruncated_cleaned.ipynb
  To build an Elastic Net model using both truncated and untruncated data (all URLs). The categorical ordinal feature charset is encoded with a frequency indexer, and the script will standardize the entire dataset.

Folder iconference_followup_study/elastic_net

• iconference_followup_study/elastic_net/lib
  The dependent library for Elastic Net analysis.
• iconference_followup_study/elastic_net/elastic_net_truncated.ipynb
  To build an Elastic Net model using only truncated data (dead URLs).
• iconference_followup_study/elastic_net/elastic_net_untruncated.ipynb
  To build an Elastic Net model using both truncated and untruncated data (all URLs). The categorical ordinal feature charset is encoded using a frequency indexer. Since the script utilizes the Plotly package for data visualization, the output can be quite large.

Folder iconference_followup_study/random_forest

• iconference_followup_study/random_forest/best_lambda_scrutinize.ipynb
  To fine-tune the hyperparameters of the Random Forest model.
• iconference_followup_study/random_forest/best_lambda_summary.ipynb
  To report the performance of the Random Forest model with its best hyperparameters.

Folder iconference_followup_study/tobit

• iconference_followup_study/tobit/eda
  Scripts for exploratory data analysis (EDA).
• iconference_followup_study/tobit/grid_search
  Scripts for searching optimal hyperparameters.