aRxiv tutorial

arXiv is a repository of electronic preprints for computer science, mathematics, physics, quantitative biology, quantitative finance, and statistics. The aRxiv package provides an R interface to the arXiv API.

Installation

You can install the aRxiv package via CRAN:

install.packages("aRxiv")

Or use devtools::install_github() to get the (possibly more recent) version at GitHub:

install.packages("devtools")
library(devtools)
install_github("ropensci/aRxiv")

Basic use

Use arxiv_search() to search arXiv, arxiv_count() to get a simple count of manuscripts matching a query, and arxiv_open() to open the abstract pages for a set of results from arxiv_search().

We’ll get to the details in a moment. For now, let’s look at a few examples.

Suppose we wanted to identify all arXiv manuscripts with “Peter Hall” as an author. It is best to first get a count, so that we have a sense of how many records the search will return. (Peter Hall was “among the world’s most prolific and highly cited authors in both probability and statistics.”) We first use library() to load the aRxiv package and then arxiv_count() to get the count.

library(aRxiv)
arxiv_count('au:"Peter Hall"')

## [1] 64

The au: part indicates to search the author field; we use double quotes to search for a phrase.

To obtain the actual records matching the query, use arxiv_search().

rec <- arxiv_search('au:"Peter Hall"')
nrow(rec)

## [1] 10

The default is to grab no more than 10 records; this limit can be changed with the limit argument. But note that the arXiv API will not let you download more than 50,000 or so records, and even in that case it’s best to do so in batches; more on this below.

Also note that the result of arxiv_search() has an attribute "total_results" containing the total count of search results; this is the same as what arxiv_count() provides.

attr(rec, "total_results")

## [1] 64

The following will get us all 64 records.

rec <- arxiv_search('au:"Peter Hall"', limit=100)
nrow(rec)

## [1] 64

arxiv_search() returns a data frame with each row being a single manuscript. The columns are the different fields (e.g., authors, title, abstract, etc.). Fields like authors that contain multiple items will be a single character string with the multiple items separated by a vertical bar (|).

We might be interested in a more restrictive search, such as for Peter Hall’s arXiv manuscripts that have deconvolution in the title. We use ti: to search the title field, and combine the two with AND.

deconv <- arxiv_search('au:"Peter Hall" AND ti:deconvolution')
nrow(deconv)

## [1] 4

Let’s display just the authors and title for the results.

deconv[, c('title', 'authors')]

##                                                                             title
## 1                                     A ridge-parameter approach to deconvolution
## 2                                     On deconvolution with repeated measurements
## 3 Estimation of distributions, moments and quantiles in deconvolution\n  problems
## 4   Kernel methods and minimum contrast estimators for empirical\n  deconvolution
##                                        authors
## 1                 Peter Hall|Alexander Meister
## 2 Aurore Delaigle|Peter Hall|Alexander Meister
## 3               Peter Hall|Soumendra N. Lahiri
## 4                   Aurore Delaigle|Peter Hall

We can open the abstract pages for these 4 manuscripts using arxiv_open(). It takes, as input, the output of arxiv_search().

arxiv_open(deconv)

Forming queries

The two basic arguments to arxiv_count() and arxiv_search() are query, a character string representing the search, and id_list, a list of arXiv manuscript identifiers.

If only query is provided, manuscripts matching that query are returned.
If only id_list is provided, manuscripts in the list are returned.
If both are provided, manuscripts in id_list that match query will be returned.

query may be a single character string or a vector of character strings. If it is a vector, the elements are pasted together with AND.

id_list may be a vector of character strings or a single comma-separated character string.

Search terms

Generally, one would ignore id_list and focus on forming the query argument. The aRxiv package includes a dataset query_terms that lists the terms (like au) that you can use.

query_terms

##               term                                      description
## 1               ti                                            Title
## 2               au                                           Author
## 3              abs                                         Abstract
## 4               co                                          Comment
## 5               jr                                Journal Reference
## 6              cat                                 Subject Category
## 7               rn                                    Report Number
## 8              all                                 All of the above
## 9    submittedDate Date/time of initial submission, as YYYYMMDDHHMM
## 10 lastUpdatedDate        Date/time of last update, as YYYYMMDDHHMM

Use a colon (:) to separate the query term from the actual query. Multiple queries can be combined with AND, OR, and ANDNOT. The default is OR.

arxiv_count('au:Peter au:Hall')

## [1] 31498

arxiv_count('au:Peter OR au:Hall')

## [1] 31498

arxiv_count('au:Peter AND au:Hall')

## [1] 133

arxiv_count('au:Hall ANDNOT au:Peter')

## [1] 2394

It appears that in the author field (and many other fields) you must search full words, and that wild cards are not allowed.

arxiv_count('au:P* AND au:Hall')

## [1] 0

arxiv_count('au:P AND au:Hall')

## [1] 1116

arxiv_count('au:"P Hall"')

## [1] 43

Subject classifications

arXiv has a set of 127 subject classifications, searchable with the prefix cat:. The aRxiv package contains a dataset arxiv_cats containing the abbreviations and descriptions. Here are the statistics categories.

arxiv_cats[grep('^stat', arxiv_cats$abbreviation),]

##   abbreviation                   description
## 1      stat.AP     Statistics - Applications
## 2      stat.CO      Statistics - Computation
## 3      stat.ML Statistics - Machine Learning
## 4      stat.ME      Statistics - Methodology
## 5      stat.TH           Statistics - Theory

To search these categories, you need to include either the full term or use the * wildcard.

arxiv_count('cat:stat')

## [1] 0

arxiv_count('cat:stat.AP')

## [1] 13367

arxiv_count('cat:stat*')

## [1] 92530

Dates and ranges of dates

The terms submittedDate (date/time of first submission) and lastUpdatedDate (date/time of last revision) are particularly useful for limiting a search with many results, so that you may combine multiple searches together, each within some window of time, to get the full results.

The date/time information is of the form YYYYMMDDHHMMSS, for example 20071018122534 for 2007-10-18 12:25:34. You can use * for a wildcard for the times. For example, to get all manuscripts with initial submission on 2007-10-18:

arxiv_count('submittedDate:20071018*')

## [1] 196

But you can’t use the wildcard within the dates.

arxiv_count('submittedDate:2007*')

## [1] 0

To get a count of all manuscripts with original submission in 2007, use a date range, like [from_date TO to_date]. (If you give a partial date, it’s treated as the earliest date/time that matches, and the range appears to be up to but not including the second date/time.)

arxiv_count('submittedDate:[2007 TO 2008]')

## [1] 55749

Search results

The output of arxiv_search() is a data frame with the following columns.

res <- arxiv_search('au:"Peter Hall"')
names(res)

##  [1] "id"               "submitted"        "updated"          "title"           
##  [5] "abstract"         "authors"          "affiliations"     "link_abstract"   
##  [9] "link_pdf"         "link_doi"         "comment"          "journal_ref"     
## [13] "doi"              "primary_category" "categories"

The columns are described in the help file for arxiv_search(). Try ?arxiv_search.

A few short notes:

Each field is a single character string. authors, link_doi, and categories may contain multiple items, separated by a vertical bar (|).
Missing entries will have an empty character string ("").
The categories column may contain not just the aRxiv categories (e.g., stat.AP) but also codes for the Mathematical Subject Classification (MSC) (e.g., 14J60) and the ACM Computing Classification System (e.g., F.2.2). These are not searchable with cat: but are searchable with a general search.

arxiv_count("cat:14J60")

## [1] 0

arxiv_count("14J60")

## [1] 704

Sorting results

The arxiv_search() function has two arguments for sorting the results, sort_by (taking values "submitted", "updated", or "relevance") and ascending (TRUE or FALSE). If id_list is provided, these sorting arguments are ignored and the results are presented according to the order in id_list.

Here’s an example, to sort the results by the date the manuscripts were last updated, in descending order.

res <- arxiv_search('au:"Peter Hall" AND ti:deconvolution',
                    sort_by="updated", ascending=FALSE)
res$updated

## [1] "2010-03-01 11:33:37" "2008-10-27 14:27:52" "2008-04-04 12:19:05"
## [4] "2007-10-18 12:25:34"

Technical details

Metadata limitations

The arXiv metadata has a number of limitations, the key issue being that it is author-supplied and so not necessarily consistent between records.

Authors’ names may vary between records (e.g., Peter Hall vs. Peter G. Hall vs. Peter Gavin Hall vs. P Hall). Further, arXiv provides no ability to distinguish multiple individuals with the same name (c.f., ORCID).

Authors’ institutional affiliations are mostly missing. The arXiv submission form does not include an affiliation field; affiliations are entered within the author field, in parentheses. The metadata instructions may not be widely read.

There are no key words; you are stuck with searching the free text in the titles and abstracts.

Subject classifications are provided by the authors and may be incomplete or inappropriate.

Limit time between search requests

Care should be taken to avoid multiple requests to the arXiv API in a short period of time. The arXiv API user manual states:

In cases where the API needs to be called multiple times in a row, we encourage you to play nice and incorporate a 3 second delay in your code.

The aRxiv package institutes a delay between requests, with the time period for the delay configurable with the R option "aRxiv_delay" (in seconds). The default is 3 seconds.

To reduce the delay to 1 second, use:

options(aRxiv_delay=1)

Don’t do searches in parallel (e.g., via the parallel package). You may be locked out from the arXiv API.

Limit number of items returned

The arXiv API returns only complete records (including the entire abstracts); searches returning large numbers of records can be very slow.

It’s best to use arxiv_count() before arxiv_search(), so that you have a sense of how many records you will receive. If the count is large, you may wish to refine your query.

arXiv has a hard limit of around 50,000 records; for a query that matches more than 50,000 manuscripts, there is no way to receive the full results. The simplest solution to this problem is to break the query into smaller pieces, for example using slices of time, with a range of dates for submittedDate or lastUpdatedDate.

The limit argument to arxiv_search() (with default limit=10) limits the number of records to be returned. If you wish to receive more than 10 records, you must specify a larger limit (e.g., limit=100).

To avoid accidental searches that may return a very large number of records, arxiv_search() uses an R option, aRxiv_toomany (with a default of 15,000), and refuses to attempt a search that will return results above that limit.

Make requests in batches

Even for searches that return a moderate number of records (say 2,000), it may be best to make the requests in batches: Use a smaller value for the limit argument (say 100), and make multiple requests with different offsets, indicated with the start argument, for the initial record to return.

This is done automatically with the batchsize argument to arxiv_search(). A search is split into multiple calls, with no more than batchsize records to be returned by each, and then the results are combined.