ignore rendered HTML

[social-media-chapter.git] / README.md

---
title: Software and data for "A Computational Analysis of Social Media Scholarship"
output: html_document
--- 

> **Authors:** [Jeremy Foote](http://jeremydfoote.com/), [Aaron Shaw](http://aaronshaw.org/), [Benjamin Mako Hill](https://mako.cc/academic/)<br />
> **Archival copies of code and data:** <https://dx.doi.org/10.7910/DVN/W31PH5><br />
> **License:** see [COPYING file](COPYING): code is released under [GNU GPLv3](https://www.gnu.org/licenses/gpl-3.0.en.html) or any later version for code; paper is released as [CC BY-NC-SA](https://creativecommons.org/licenses/by-nc-sa/4.0/legalcode).

<figure style="float:right">
<img src="./hairball.png"/>
<figcaption>Network "hairball" of social media papers published from 2004-2016</figcaption>
</figure>

Dramatic increases in large-scale data generated through social media, combined with increased computational power, have enabled the growth of computational approaches to social media research, and social science in general. While many of these approaches require statistical or computational training, they have the great benefit of being inherently transparent—allowing for research that others can reproduce and learn from.

To that end, we wrote a book chapter in the [Sage Handbook of Social Media](https://us.sagepub.com/en-us/nam/the-sage-handbook-of-social-media/book245739) in which we obtain a large-scale dataset of metadata about social media research papers which we analyze using a few commonly-used computational methods. This document is designed to tell you exactly how we did that and to walk you through how to reproduce our results and our paper by running the code we wrote.

This document is meant to be read alongside our chapter. You can find the chapter here:

> Foote, Jeremy D., Aaron Shaw,  and Benjamin Mako Hill. 2017. “A Computational Analysis of Social Media Scholarship.”  In The SAGE Handbook of Social Media, edited by Jean Burgess, Alice Marwick, and Thomas Poell, 111–34. London, UK: SAGE. [[Official Link](https://uk.sagepub.com/en-gb/eur/the-sage-handbook-of-social-media/book245739)] [[Preprint PDF](http://mako.cc/academic/foote_shaw_hill-computational_analysis_of_social_media.pdf)]

The rest of this document describes how to download our code and data and how to reproduce the analyses in the paper on your own computer.

### Requirements and Expectations

We will be as explicit as possible in this document, and try to make it accessible to less-technical readers. However, we do make a few assumptions:

* You have access to and basic familiarity with [a POSIX command line interface](https://en.wikipedia.org/wiki/POSIX). The instructions here are written for and tested using [Debian](https://www.debian.org/) and [Ubuntu](https://www.ubuntu.com/) GNU/Linux. That said, these instructions should work without modification on most Linux systems.  Although MacOS users may need to tweak a few things, they should work there, too. Microsoft Windows users will likely need to tweak more things. This is particularly true for the last step—building the paper itself. If you can get a simple example like [this one](https://github.com/yihui/knitr-examples/blob/master/005-latex.Rtex) working, then there's a decent chance you can get the paper to build.
* You have [Python 3.x](https://www.python.org/downloads/) installed. For many users, you will already have it installed. Debian and Ubuntu users can install it with `apt install python3`. Others can download it from [the Python download page](https://www.python.org/downloads/)
* You have [GNU R 3.x](https://www.r-project.org/) installed. Debian and Ubuntu users can install it with `apt install r-base`. Others can install it from [the R homepage](https://www.r-project.org/). In our testing we used versions GNU R versions 3.3.2 and 3.4.1.
* To conduct the bibliometric network analysis, you'll need the [igraph library](http://igraph.org/). To install it on Debian or Ubuntu you can run `apt install libigraph0v5`.
* You will also need the following Python libraries:

> * [Requests](http://docs.python-requests.org/en/master/)
> * [numpy](http://www.numpy.org/)
> * [pandas](https://pandas.pydata.org/)
> * [igraph](http://igraph.org/)
> * [sklearn](http://scikit-learn.org/stable/)
> * [scipy](https://www.scipy.org/)
> * [rpy2](https://rpy2.bitbucket.io/)
> * [matplotlib](https://matplotlib.org/)

> If you use Debian or Ubuntu, you install them with the following command run as root in your shell: 

            apt install python3-requests  python3-numpy python3-pandas python3-igraph python3-sklearn python3-rpy2 python3-matplotlib
                        
> Another way to install them is using `pip3` (or just `pip`) with the following command in your shell: 

            pip3 install requests numpy pandas python-igraph sklearn scipy rpy2 matplotlib

* You will also need to install the following *R* libraries:

> * [ggplot2](http://ggplot2.org/)
> * [data.table](https://r-datatable.com/)
> * [reshape2](https://github.com/hadley/reshape)
> * [glmnet](https://cran.r-project.org/web/packages/glmnet/)
> * [txtplot](https://cran.r-project.org/web/packages/txtplot/)
> * [Matrix](https://cran.r-project.org/web/packages/Matrix/)
> * [xtable](https://cran.r-project.org/web/packages/xtable/)
> * [dplyr](http://dplyr.tidyverse.org/)
> * [knitr](https://yihui.name/knitr/)

> One way to install them is by running the command *from within a running copy of R*: 

            install.packages(c("ggplot2", "data.table", "reshape2", "glmnet", "txtplot", "Matrix", "xtable", "dplyr", "knitr"))

We have ensured that every piece of software used in this analysis is [free/open source software](https://www.gnu.org/philosophy/free-sw.en.html) which means it is both available at no cost and, like our analysis code, is transparent and inspectable.

### 1. Setting up Your Environment

You'll want to start by creating a new directory. Download and extract following file from the [the Harvard Dataverse repository for the paper](https://dx.doi.org/10.7910/DVN/W31PH5):

* [`code_and_paper.tar.gz`](https://dataverse.harvard.edu/file.xhtml?fileId=3107305) (this file is publicly accessible)

If this is successful, you should have two subdirectories: `code` and `paper`.

### 2. Getting the Scopus Data

The data for this project came from the [Scopus API](https://dev.elsevier.com/). An API ([Application Programming Interface](https://en.wikipedia.org/wiki/Application_programming_interface)) is basically a way for computers to talk to each other directly. In our case, we asked the Scopus API to give us metadata about a set of research papers. In order to replicate the paper, you have two options for getting this metadata.

#### 2.1. Option 1: Downloading Collected Data from Harvard Dataverse

As part of writing this paper, we did the work of downloading the metadata from Scopus and it is available in the [`raw_data.tar.gz`](https://dataverse.harvard.edu/file.xhtml?fileId=3106986) file on the Dataverse page. This dataset includes metadata about all of the papers that include the term "*social media*" in their title, abstract, or keywords, as well as all of the papers which cite these papers, as of February 2017.

Scopus won't let us make that dataset publicly accessible on the web so you'll need to request access to it through [the Harvard Dataverse](https://dataverse.harvard.edu/) and we'll ask that you only use the data for the purpose of reproducing this paper.

Once you have downloaded the data, unpack it in the same directory that you unpacked the `code_and_data.tar.gz` file. Now you should have a third subdirectory: `raw_data`.

#### 2.2. Option 2: Getting the Data Yourself

If you want to download the data yourself instead of using the `raw_data.tar.gz` file we have prepared, follow the instructions in this section.  There are several reasons you might want to do this. For example, you might want to retrieve a new version of the dataset with details of more recent papers. Or you might want to do a similar analysis with a different search term.

In order to do this:

1. You must belong to an institution that has access to [Scopus](https://www.scopus.com/). 
2. You will need to be patient. Scopus has a weekly limit on requests to their API, and it may take multiple weeks to download all of the results.

You will also need to get an API key from <https://dev.elsevier.com/index.html>. In the `code/data_collection` directory, edit the `scopus_api.py` file so that it has your key. The file should look something like `key = 'XXXXXXXXXXXXXXXXX'` where your key replaces the Xs.

The Scopus API is a little odd, in that you are authenticated based on your key, but your permissions change based on your IP ([authentication documentation](https://dev.elsevier.com/tecdoc_api_authentication.html)). I noticed that I had to be logged into our university VPN (even if I was on campus) in order to have all of the permissions needed to carry this out.

Getting the data is a three-step process:

1. First, we retrieve data about the papers which include the term "*social media*" in their title, abstract, or keywords. To do this, run the following command from the directory where you downloaded and unpacked our code archives:

        mkdir raw_data
        python3 code/data_collection/00_get_search_results.py -q 'title-abs-key("social media")' -o raw_data/search_results.json

> If you want to change the date range that is considered, then edit the `years` variable in the `code/data_collection/00_get_search_results.py` file.

2. Next, we take each of those results and query the API to get more detailed information about them (e.g., abstracts). To do that, run:

        python3 code/data_collection/01_get_abstracts.py -i raw_data/search_results.json -o raw_data/abstracts_and_citations.json

3. Finally, the search results are then used as input to get a some basic metadata about all of the papers that cite them (we use these data for the network analysis):

        python3 code/data_collection/02_get_cited_by.py -i raw_data/search_results.json -o raw_data/cited_by.json

### 3. Cleaning the Data

Whichever option you used above, you should now have a `raw_data` subdirectory which contains the files `search_results.json`, `abstracts_and_citations.json`, and `cited_by.json`.

These raw data files are "raw" in the sense that they contain lots of data we are looking for as well as lots of things we won't use in this analysis. They are also in JSON format which is not the best format for bringing data into most of the tools we'll be using for our analysis. As a result, we will clean them up to make them easier to work with in the tools that we'll be using.

All of the following commands should be run from the directory where you downloaded the material.

1. First, create a `processed_data` directory in the root directory with `raw_data`, `code`, etc., to store all of these processed data files:

        mkdir processed_data

2. We'll start by changing the abstracts file into a tab-separated value (TSV) table:

        python3 code/data_processing/00_abstracts_to_tsv.py -i raw_data/abstracts_and_citations.json -o processed_data/abstracts.tsv

3. We'll also make a few additional tables to make it easier to work with the data. A social network "edgelist" file which lists which papers cite each other:

        python3 code/data_processing/01_cited_by_to_edgelist.py -i raw_data/cited_by.json -o processed_data/citation_edgelist.txt

4. And a filtered version of this file which only lists the papers which include "social media" in their abstract, title, or keywords:

        python3 code/data_processing/02_filter_edgelist.py -i processed_data/citation_edgelist.txt -o processed_data/social_media_edgelist.txt

5. Finally, we make a few tables that are easy to import into R:

        python3 code/data_processing/03_make_paper_aff_table.py -i raw_data/abstracts_and_citations.json -o processed_data/paper_aff_table.tsv
        python3 code/data_processing/04_make_paper_subject_table.py -i raw_data/abstracts_and_citations.json -o processed_data/paper_subject_table.tsv 

Once we have all the data cleaned and prepared, we are ready to proceed to our analysis. Our general workflow for analysis will be to run our analysis code and then save the output to an RData file in the `paper/data/` subdirectory. We will then import these RData files into the paper to create figures and tables.

Doing this will require two final steps:

6. First, you'll need to create the directory:

        mkdir paper/data

7. Before we get to the analysis, though, we'll save some portions of the processed datasets to the `paper/data` subdirectory by running the following command, which will help us access the processed data directly from our paper in order to report descriptive statistics:

        Rscript code/data_processing/05_save_descriptives.R

### 4. Bibliometric Analysis

#### 4.1 Cluster Analysis and Network Statistics

The code used for our bibliometric analysis is contained within the `code/bibliometrics/` subdirectory.

We've included two copies of our Python code for our bibliometric analysis in the files `00_citation_network_analysis.py` and `00_citation_network_analysis.ipynb`. We will describe using the former in this section. If you have [Juypter](https://jupyter.org/) installed you can open the file in a a notebook format used by many scientists by running `jupyter-notebook citation_network_analysis.ipynb`. If you want to try Jupyter, Debian and Ubuntu users can install it with `apt install jupyter-notebook` and other users can download it [here](https://jupyter.org/install.html).

Our bibliometric analysis code does require one additional piece of software called [Infomap](http://www.mapequation.org/) which we use to identify clusters in our citation network. There are some [instructions online](https://github.com/mapequation/infomap) but you can download and install it with the following commands run from the `code/bibliometrics` subdirectory:

            mkdir output_dir
        git clone git@github.com:mapequation/infomap.git
        cd infomap
        make

Once you have Infomap installed, running our bibliometric analysis all done with a single Python command run from the root directory:

        python3 code/bibliometrics/00_citation_network_analysis.py

This will save the output to `paper/data/network_data.RData`. This data is used in the bibliometric section, for example, to create Table 4 and Figure 3.

#### 4.2 Network Visualizations

If you want to create our two network diagrams, you'll need one additional piece of software: [Gephi, the Open Graph Viz Platform](https://gephi.org/). Figure 2 in the paper is a "hairball" network graph of the citation network in our dataset.  Like all the software used here, it is free/open source software and available for [download](https://gephi.org/users/download/).

Gephi is a graphical and interactive tool so, unlike the rest of our analysis, it will take some clicking around to reproduce our graph exactly. 

We created the the big network "hairball" in Figure 2 by starting in the following way:

1. Opening the file `code/bibliometrics/g_sm.graphml`
2. On the top part of the left sidebar, clicking on: *Appearance*, *Node*, the color palette icon, *Partition*, selecting "cluster" from the drop-down, and then *▶Apply*.
3. In the *Layout* tab, Selecting "Fruchterman Reingold" and then clicking *Run*. (This will take a really long time!)

To create the small network cluster in Figure 3 we began in the following way:

1. Opening the file `code/bibliometrics/cluster.graphml`
2. In the *Layout* tab, selecting "Fruchterman Reingold" and then clicking *Run*. 
3. Clicking on: *Appearance*, *Node*, the color palette icon, *Partition*, selecting "name" from the drop-down, and then clicking *▶Apply*.
4. Clicking on: *Appearance*, *Node*, the size icon concentric circles, *Ranking*, selecting "papers" from the drop-down, fiddling with the min and max sizes, and then clicking *▶Apply*.

Getting things just right will take some fiddling! We've included our Gephi files (`code/bibliometrics/clusters.gephi` and `code/bibliometrics/g_sm.gephi`) which are finished or mostly finished versions that you can open up and use if you'd like.


### 5. Topic Modeling Analysis

The topic in the `code/topic_modeling` directory applies [latent Dirichlet allocation](https://en.wikipedia.org/wiki/Latent_Dirichlet_allocation) topic modeling to the social media abstracts.

LDA takes in a set of documents and produces a set of topics and a distribution of topics for each document. The first file takes in the abstract file, and creates two outputs: The abstracts together with their topic distribution and a set of topics and the top words associated with each.

Our topic modeling analysis includes the following steps:

1. Running the following Python script that extracts our topics:

        python3 code/topic_modeling/00_topics_extraction.py

> Note that this will take some time —5-10 minutes on a decent laptop.

2. Running a second file which (a) makes a couple of tables of the top words for each topic and (b) generates some summary statistics for how the topics change over time. These statistics are used, e.g., to create Figure 4. You can run this with:

        python3 code/topic_modeling/01_make_paper_files.py

Topic modeling is a stochastic process, and you may notice differences—potentially large differences—between the results in our paper and the results when you run it. If the order of topics has changed (or if you think other labels would be appropriate for some topics), then you can adjust the topic names by editing the `topic_names` list in the `01_make_papers.py` file.

### 6. Prediction Analysis


For the prediction analysis, we use features of the papers to predict whether or not a paper gets cited. 

These commands require a computer with a large amount of memory (i.e., RAM). We had trouble running this step on our laptops which did not have enough. If you don't have access to such a computer, then you can change the `n_features` variable in the `00_ngram_extraction.py` file from `100000` to something like `3000`. This will change how many terms are included in the prediction analysis, but shouldn't make an important difference in the results.

We ran the following steps:

1. Because one of the features we use is the text of the abstracts, we start by getting uni-, bi-, and tri-grams from the abstracts. This is done with the `00_ngram_extraction` python file: 

        python3 code/prediction/00_ngram_extraction.py

> This will create an `ngram_table.csv` file in the `processed_data` directory.

2. Next, we build control variables:

        Rscript code/prediction/01-build_control_variables.R 
        
> This will build a set of descriptive variables, saved to `paper/data/prediction_descriptives.RData`

3. Then, we create textual features using the ngram table we created:

        Rscript code/prediction/02-build_textual_features.R

> This step filters the ngrams to those which occur across 30 subject areas and saves them as a matrix at `processed_data/top.ngram.matrix.RData`.

4. Finally, the data that we've collected is put into models, and statistics are collected:

        Rscript code/prediction/03-prediction_analysis.R

>This step will save the output of the models to `paper/data/prediction.RData`. 

Each of these final two steps will take a long time to run.

### 7. Building the Paper


Now we should have all of the data that we need sitting in the `paper/data` directory. We use a really neat package called [knitr](https://yihui.name/knitr/) to load and manipulate data directly in the document. Nearly all of the tables and figures in the document are created using code that is in the document, and knitr is the magic behind the scenes that makes that possible.

To build the paper, we will be using a [Makefile](https://www.gnu.org/software/make/manual/make.html) which requires a set of utilities to run. This is the part that might get a little tricky for Windows users.

You can install the packages you'll need from Debian or Ubuntu with the following command: 

            apt install rubber latexmk texlive-latex-recommended texlive-latex-extra texlive-fonts-extra texlive-fonts-recommended texlive-bibtex-extra moreutils gawk

After this, you should change to the `paper` directory and simply run the command `make`.

This will produce a quickly scrolling output to standard out, and if everything has worked, then in the end it will produce a bunch of files in the `paper` directory, one of which will be the final PDF file!

### Errors, Improvements, and Updates

Although we have tried to make this document as clear as possible and although we have tested it carefully ourselves, there are many ways it might fail. You might have a different software environment or the software dependencies we have used might have [bit rot](https://en.wikipedia.org/wiki/Software_rot) over time in a way that leads to things breaking.

You are welcome to get in touch with us if you have questions and we have provided our webpages and emails here for that purpose:

* [Jeremy Foote](http://jeremydfoote.com/) <<jdfoote1@gmail.com>>
* [Aaron Shaw](http://aaronshaw.org/) <<aaron.d.shaw@gmail.com>>
* [Benjamin Mako Hill](https://mako.cc/academic/) <<makohill@uw.edu>>

If you can fix issues you run into, find ways to clarify our instructions, or make fixes to our code, please tell us!

In addition to the [archival version in the Dataverse](https://dx.doi.org/10.7910/DVN/W31PH5), we have hosted our code in a [git revision control management](https://git-scm.com/) repository here: <https://code.communitydata.cc/social-media-chapter.git>
Even the page you are reading is included in our repository. If you notice any typos or errors, please fix it and send us your fix!

Please feel to follow the [instructions we have posted](https://code.communitydata.cc/) for sending updated versions of our code or documentation so that others trying to replicate and learn from our work can benefit from your work and improvements!