../../_images/badge-colab.svg ../../_images/badge-github-custom.svg

Extracting researchers based on affiliations and publications history

The purpose of this notebook is to demonstrate how to extract researchers data using the Dimensions API.

Specifically, we want to find out about all researchers based on these two criteria:

  1. Affiliation. That is, whether they are / have been affiliated to a specific GRID organization

  2. Publications. Whether they have published within a chosen time frame.

Prerequisites

This notebook assumes you have installed the Dimcli library and are familiar with the Getting Started tutorial.

[1]:

!pip install dimcli -U --quiet

import dimcli
from dimcli.utils import *

import os, sys, time, json
import pandas as pd
from IPython.display import Image
from IPython.core.display import HTML

print("==\nLogging in..")
# https://digital-science.github.io/dimcli/getting-started.html#authentication
ENDPOINT = "https://app.dimensions.ai"
if 'google.colab' in sys.modules:
  import getpass
  KEY = getpass.getpass(prompt='API Key: ')
  dimcli.login(key=KEY, endpoint=ENDPOINT)
else:
  KEY = ""
  dimcli.login(key=KEY, endpoint=ENDPOINT)
dsl = dimcli.Dsl()

==
Logging in..
Dimcli - Dimensions API Client (v0.9.1)
Connected to: https://app.dimensions.ai - DSL v1.31
Method: dsl.ini file

Select an organization ID

For the purpose of this exercise, we will are going to use grid.258806.1 and the time frame 2013-2018. Feel free though to change the parameters below as you want, eg by choosing another GRID organization.

[2]:

# sample org: grid.258806.1
GRIDID = "grid.258806.1" #@param {type:"string"}
START_YEAR = 2013 #@param {type:"slider", min:1900, max:2020, step: 1}
END_YEAR = 2018 #@param {type:"slider", min:1900, max:2020, step: 1}

if START_YEAR < END_YEAR: START_YEAR = END_YEAR
YEARS = f"[{START_YEAR}:{END_YEAR}]"

Background: understanding the data model

In order to process researchers affiliation data in the context of publications, we should first take the time to understand how this data is structured in Dimensions.

The JSON results of any query with shape search publications where .... return publications are composed by a list of publications. If we open up one single publication record we will immediately see that authors are stored in a nested object authors that contains a list of dictionaries. Each element in this dictionary represents one single publication author and includes other information e.g. name, surname, ID, the organizations he/she is affiliated with etc..

For example, in order to extract the second author of the tenth publication from our results we would do the following: results.publications[10]['authors'][1]:

# author info
...
    {'first_name': 'Noboru',
     'last_name': 'Sebe',
     'orcid': '',
     'current_organization_id': 'grid.258806.1',
     'researcher_id': 'ur.010647607673.28',
     'affiliations': [{'id': 'grid.258806.1',
       'name': 'Kyushu Institute of Technology',
       'city': 'Kitakyushu',
       'city_id': 1859307,
       'country': 'Japan',
       'country_code': 'JP',
       'state': None,
       'state_code': None}]}
 ...

Here’s a object model diagram summing up how data is structured.

[3]:

Image(url= "http://api-sample-data.dimensions.ai/diagrams/data_model_researchers_publications1.v2.jpg", width=800)

[3]:

There are a few important things to keep in mind:

  • Publication Authors VS Researchers. In Dimensions, publication authors don’t necessarily have a researcher ID (eg because they haven’t been disambiguated yet). So a publication may have N authors (stored in JSON within the authors key), but only a subset of these include a researcher_id link. PS see also the searching for researchers section of the API docs for more info on this topic.

  • Time of the affiliation. Researchers can be affiliated to a specific GRID organization either at the time of speaking (now), or at the time of writing (i.e. when the article was published). The DSL uses different properties to express this fact: current_research_org or simply research_orgs.

  • Denormalized fields. Both the Publication and Researcher sources include a research_orgs field. In both cases that’s a denormalized field ie a ‘shortcut version’ of data primarily found in the authors field of publications. However, although they share the same name, the two research_orgs fields don’t have the same meaning.

    • In publications, research_orgs contains the set of all authors’ affiliations for a single article.

    • In researchers, research_orgs contains the set of all research organizations a single individual has been affiliated to throughout his/her career (as far as Dimensions knows, of course!).

So, in the real world we often have scenarios like the following one:

[4]:

Image(url= "http://api-sample-data.dimensions.ai/diagrams/data_model_researchers_publications2.v2.jpg", width=1000)

[4]:

Methodology: two options available

It turns out that there are two possible ways to extract these data, depending on whether we start our queries from Publications or from Researchers.

  1. Starting from the Publication source, we can first filter publications based on our constraints (eg year range [2013-2018] and research_orgs="grid.258806.1" - but it could be any other query parameters); second, we would loop over all of these publications so to extract all relevant reasearchers using the affiliation data.

  2. Starting from the Researcher source, we would first filter researchers based on our constraints (eg with research_orgs="grid.258806.1"); second, we would search for publications linked to these researchers, which have been published in the time frame [2013-2018]; lastly, we extract all relevant reasearchers using the affiliation data.

As we will find out later, both approaches are perfectly valid and return the same results.

The first apporach is generally quicker as it has only two steps (while the second one has three).

In real-world situations though, deciding which approach is best depends on the specific query filters being used and on the impact these filters have on the overall performance/speed of the data extraction. There is no fixed rule and a bit of trial & error can go a long way in helping you optimize your data extraction algorithm!

Approach 1. From publications to researchers

Starting from the publications source, the steps are as follows:

  1. Filtering publications for year range [2013-2018] and research_orgs=“grid.258806.1”

    • i.e. search publications where year in [2013:2018] and research_orgs="grid.258806.1" return publications

  2. Looping over publications’ authors and extracting relevant researchers

    • if ['affiliation']['id'] == "grid.258806.1"

      • => that gives us the affiliations at the time of publishing

    • if ['current_organization_id'] == "grid.258806.1"

      • => that gives us the affiliations at the time of speaking

First off, we can to get all publications based on our search criteria by using a loop query.

[17]:

pubs = dsl.query_iterative(f"""search publications
                                where year in {YEARS} and research_orgs="{GRIDID}"
                            return publications[id+title+doi+year+type+authors+journal+issue+volume]""")

Starting iteration with limit=1000 skip=0 ...
0-693 / 693 (1.23s)
===
Records extracted: 693

First we want to know how many researchers linked to these publications were affiliated to the GRID organization when the publication was created (note: they may be at at a different institution right now).

The affiliation data in publications represent exactly that: we can thus loop over them (for each publication/researcher) and keep only the ones matching our GRID ID.

TIP affiliations can be extracted easily thanks one of the ‘dataframe’ transformation methods in Dimcli: as_dataframe_authors_affiliations

[18]:

# extract affiliations from a publications list
affiliations = pubs.as_dataframe_authors_affiliations()
# select only affiliations for GRIDID
authors_historical = affiliations[affiliations['aff_id'] == GRIDID].copy()
# remove duplicates by eliminating publication-specific data
authors_historical.drop(columns=['pub_id'], inplace=True)
authors_historical.drop_duplicates('researcher_id', inplace=True)
print(f"===\nResearchers with affiliation to {GRIDID} at time of writing:", authors_historical.researcher_id.nunique(), "\n===")
# preview the data
authors_historical

===
Researchers with affiliation to grid.258806.1 at time of writing: 546
===

[18]:
aff_id aff_name aff_city aff_city_id aff_country aff_country_code aff_state aff_state_code researcher_id first_name last_name
0 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP Siewteng Sim
1 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.01116323260.31 Yoshito Andou
6 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.010362311021.55 Kubra Eksiler
7 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.011724474721.84 Satoshi Iikubo
12 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.016307426665.31 R. Okamoto
... ... ... ... ... ... ... ... ... ... ... ...
4025 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.011222056445.46 Harald KLEINE
4033 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.010000541204.77 Ryuji TAKAHASHI
4062 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.010305013750.53 Satoshi Hiai
4072 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.014011430535.10 Kei Yamamura
4073 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.014512005745.07 Aoi Honda

546 rows × 11 columns

Note: * The first ‘publication-affiliations’ dataframe we get may contain duplicate records - eg if an author has more than one publication it’ll be listed twice. That’s why we have an extra step where we drop the pub_id column and simply count unique researchers, based on their researcher ID.

This can be achieved simply by taking into consideration a different field, called current_organization_id, available at the outer level of the JSON author structure (see the data model section above) - outside the affiliations list.

Luckily Dimcli includes another handy method for umpacking authors into a dataframe: as_dataframe_authors

[19]:

authors_current = pubs.as_dataframe_authors()
authors_current = authors_current[authors_current['current_organization_id'] == GRIDID].copy()
authors_current.drop(columns=['pub_id'], inplace=True)
authors_current.drop_duplicates('researcher_id', inplace=True)
print(f"===\nResearchers with affiliation to {GRIDID} at the time of speaking:", authors_current.researcher_id.nunique(), "\n===")
authors_current

===
Researchers with affiliation to grid.258806.1 at the time of speaking: 518
===

[19]:
first_name last_name corresponding orcid current_organization_id researcher_id affiliations
1 Yoshito Andou grid.258806.1 ur.01116323260.31 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
6 Kubra Eksiler grid.258806.1 ur.010362311021.55 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
7 Satoshi Iikubo ['0000-0002-5186-4058'] grid.258806.1 ur.011724474721.84 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
12 R. Okamoto grid.258806.1 ur.016307426665.31 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
22 Shuzi Hayase grid.258806.1 ur.01055753603.27 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
... ... ... ... ... ... ... ...
3455 Makoto TAKENAKA grid.258806.1 ur.015326226575.37 [{'name': 'Kagawa Prefectural Industrial Techn...
3475 Ryuji TAKAHASHI grid.258806.1 ur.010000541204.77 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
3504 Satoshi Hiai True grid.258806.1 ur.010305013750.53 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
3514 Kei Yamamura grid.258806.1 ur.014011430535.10 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
3515 Aoi Honda ['0000-0002-4485-1523'] grid.258806.1 ur.014512005745.07 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...

518 rows × 7 columns

Approach 2. From researchers to publications

Using this approach, we start our search from the ‘researchers’ database (instead of the ‘publications’ database).

There are 3 main steps:

  1. Filtering researchers with research_orgs=GRID-ID (note: this gives us affiliated researches at any point in time)

    • search researchers where research_orgs="grid.258806.1" return researchers

  2. Searching for publications linked to these researchers and linked to GRID-ID, which have been published in the time frame [2013-2018]

    • search publications where researchers.id in {LIST OF IDS} and year in [2013:2018] and research_orgs="grid.258806.1" return publications

    • NOTE: this a variation of the Approach-1 query above: we have just added the researchers IDs filter (thus reducing the search space)

  3. Extracting relevant researchers from publications, using the same exact steps as in approach 1 above.

    • if ['current_organization_id'] == "grid.258806.1"

      • => that gives us the affiliations at the time of speaking

    • or if ['affiliation']['id'] == "grid.258806.1"

      • => that gives us the affiliations at the time of publishing

[20]:

q = f"""search researchers where research_orgs="{GRIDID}"
        return researchers[basics]"""
researchers_json = dsl.query_iterative(q)
researchers = researchers_json.as_dataframe()
researchers.head()

Starting iteration with limit=1000 skip=0 ...
0-1000 / 8976 (0.86s)
1000-2000 / 8993 (1.41s)
2000-3000 / 8993 (1.08s)
3000-4000 / 8976 (2.34s)
4000-5000 / 8993 (1.16s)
5000-6000 / 8976 (0.90s)
6000-7000 / 8976 (1.27s)
7000-8000 / 8976 (0.97s)
8000-8993 / 8993 (1.06s)
===
Records extracted: 8993

[20]:
id first_name last_name research_orgs orcid_id
0 ur.012161120453.51 Seita Seki [{'id': 'grid.258806.1', 'types': ['Education'... NaN
1 ur.014746662301.94 Takeshi Yoshinaga [{'id': 'grid.258806.1', 'types': ['Education'... NaN
2 ur.016064142555.51 Charles Ronald Harahap [{'id': 'grid.258806.1', 'types': ['Education'... NaN
3 ur.016437530541.00 Yurie Sugimoto [{'id': 'grid.258806.1', 'types': ['Education'... NaN
4 ur.011113737576.59 Zhen Wang [{'id': 'grid.258806.1', 'types': ['Education'... [0000-0002-9136-3100]

Now we need to select only the researchers who have published in the time frame [2013:2018]. So for each researcher ID we must extract the full publication history in order to verify their relevance.

The most efficient way to do this is to use a query that extracts the publication history for several researchers at the same time (so to avoid overruning our API quota), then, as a second step, producing a clean list of relevant researchers from it.

[21]:

results = []
researchers_ids = list(researchers['id'])
# no of researchers IDs per query: so to ensure we never hit the 1000 records limit per query
CHUNKS_SIZE = 300

q = """search publications
                where researchers.id in {}
                and year in {}
                and research_orgs="{}"
            return publications[id+title+doi+year+type+authors+journal+issue+volume] limit 1000"""

from dimcli.shortcuts import chunks_of

for chunk in chunks_of(researchers_ids, size=CHUNKS_SIZE):
    data = dsl.query(q.format(json.dumps(chunk), YEARS, GRIDID))
    try:
        results += data.publications
    except:
        pass

print("---\nFound", len(results), "publications for the given criteria (including duplicates)")

# simulate a DSL payload using Dimcli
pubs_v2 = dimcli.DslDataset.from_publications_list(results)

# transform to a dataframe to remove duplicates quickly
pubs_v2_df = pubs_v2.as_dataframe()
pubs_v2_df.drop_duplicates("id", inplace=True)
print("Final result:", len(pubs_v2_df), "unique publications")

Returned Publications: 32 (total = 32)
Time: 0.78s
Returned Publications: 54 (total = 54)
Time: 1.34s
Returned Publications: 78 (total = 78)
Time: 1.27s
Returned Publications: 43 (total = 43)
Time: 0.87s
Returned Publications: 25 (total = 25)
Time: 0.86s
Returned Publications: 88 (total = 88)
Time: 1.00s
Returned Publications: 70 (total = 70)
Time: 0.92s
Returned Publications: 62 (total = 62)
Time: 0.76s
Returned Publications: 24 (total = 24)
Time: 0.66s
Returned Publications: 46 (total = 46)
Time: 0.66s
Returned Publications: 75 (total = 75)
Time: 0.77s
Returned Publications: 47 (total = 47)
Time: 0.66s
Returned Publications: 29 (total = 29)
Time: 0.97s
Returned Publications: 80 (total = 80)
Time: 0.78s
Returned Publications: 89 (total = 89)
Time: 0.79s
Returned Publications: 21 (total = 21)
Time: 0.64s
Returned Publications: 41 (total = 41)
Time: 0.63s
Returned Publications: 66 (total = 66)
Time: 0.76s
Returned Publications: 17 (total = 17)
Time: 0.67s
Returned Publications: 36 (total = 36)
Time: 0.63s
Returned Publications: 67 (total = 67)
Time: 0.98s
Returned Publications: 46 (total = 46)
Time: 0.94s
Returned Publications: 76 (total = 76)
Time: 0.94s
Returned Publications: 54 (total = 54)
Time: 0.94s
Returned Publications: 59 (total = 59)
Time: 0.77s
Returned Publications: 50 (total = 50)
Time: 0.70s
Returned Publications: 36 (total = 36)
Time: 0.70s
Returned Publications: 46 (total = 46)
Time: 0.79s
Returned Publications: 89 (total = 89)
Time: 0.99s
Returned Publications: 61 (total = 61)
Time: 0.78s
---
Found 1607 publications for the given criteria (including duplicates)
Final result: 680 unique publications

This step is basically the same as in approach 1 above.

[22]:

# extract affiliations from a publications list
affiliations_v2 = pubs_v2.as_dataframe_authors_affiliations()
# select only affiliations for GRIDID
authors_historical_v2 = affiliations_v2[affiliations_v2['aff_id'] == GRIDID].copy()
# remove duplicates by eliminating publication-specific data
authors_historical_v2.drop(columns=['pub_id'], inplace=True)
authors_historical_v2.drop_duplicates('researcher_id', inplace=True)
print(f"===\nResearchers with affiliation to {GRIDID} at time of writing:", authors_historical_v2.researcher_id.nunique(), "\n===")
# preview the data
authors_historical_v2

===
Researchers with affiliation to grid.258806.1 at time of writing: 546
===

[22]:
aff_id aff_name aff_city aff_city_id aff_country aff_country_code aff_state aff_state_code researcher_id first_name last_name
0 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.011113737576.59 Zhen Wang
1 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.011773542351.82 Muhammad Akmal Kamarudin
2 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP Ng Chi Huey
3 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.010311155152.66 Fu Yang
4 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.010416773662.27 Manish Pandey
... ... ... ... ... ... ... ... ... ... ... ...
9445 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.015771611513.28 Yoshiro Fukui
9885 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.011176635061.41 Takao Kodama
9908 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.014036341366.26 Kei Ohnishi
10098 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.0775361534.49 Ken'ichi Yokoyama
10122 grid.258806.1 Kyushu Institute of Technology Kitakyushu 1.85931e+06 Japan JP ur.01156670436.97 Hiroki Obata

546 rows × 11 columns

Also here, the procedure is exactly the same as in approach 1.

[23]:

authors_current_v2 = pubs_v2.as_dataframe_authors()
authors_current_v2 = authors_current_v2[authors_current_v2['current_organization_id'] == GRIDID].copy()
authors_current_v2.drop(columns=['pub_id'], inplace=True)
authors_current_v2.drop_duplicates('researcher_id', inplace=True)
print(f"===\nResearchers with affiliation to {GRIDID} at the time of speaking:", authors_current_v2.researcher_id.nunique(), "\n===")
authors_current_v2

===
Researchers with affiliation to grid.258806.1 at the time of speaking: 518
===

[23]:
first_name last_name corresponding orcid current_organization_id researcher_id affiliations
0 Zhen Wang ['0000-0002-9136-3100'] grid.258806.1 ur.011113737576.59 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
1 Muhammad Akmal Kamarudin ['0000-0002-2256-5948'] grid.258806.1 ur.011773542351.82 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
3 Fu Yang ['0000-0001-7673-8026'] grid.258806.1 ur.010311155152.66 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
4 Manish Pandey ['0000-0003-0963-8097'] grid.258806.1 ur.010416773662.27 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
5 Gaurav Kapil grid.258806.1 ur.07620457442.79 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
... ... ... ... ... ... ... ...
8268 Yoshiro Fukui grid.258806.1 ur.015771611513.28 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
8680 Takao Kodama True grid.258806.1 ur.011176635061.41 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
8701 Kei Ohnishi grid.258806.1 ur.014036341366.26 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
8868 Ken'ichi Yokoyama True grid.258806.1 ur.0775361534.49 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...
8891 Hiroki Obata grid.258806.1 ur.01156670436.97 [{'id': 'grid.258806.1', 'name': 'Kyushu Insti...

518 rows × 7 columns

Conclusions

As anticipated above, both approaches are equally valid and in fact they return the same (or very similar) number of results. Let’s compare them:

[24]:

# create summary table
data = [['1', len(authors_current), len(authors_historical),
         """search publications where year in [2013:2018] and research_orgs="grid.258806.1" return publication""",
         ],
        ['2', len(authors_current_v2), len(authors_historical_v2),
         """search researchers where research_orgs="grid.258806.1" return researchers --- then --- search publications where researchers.id in {IDS} and year in [2013:2018] and research_orgs={GRIDID} return publications""",
        ]]

pd.DataFrame(data, columns = ['Method', 'Authors (current)', 'Authors (historical)',  'Query'])

[24]:
Method Authors (current) Authors (historical) Query
0 1 518 546 search publications where year in [2013:2018] ...
1 2 518 546 search researchers where research_orgs="grid.2...

In some cases you might encounter small differences in the total number of records returned by the two approaches (eg one method returns 1-2 extra records than the other one).

This is usually due to a synchronization delay between Dimensions databases (e.g. publications and researchers). The differences are negligible in most cases, but in general it’s enough to run same extraction again after a day or two for the problem to disappear.

It depends on which/how many filters are being used in order to identify a suitable results set for your research question.

  • The first approach is generally quicker as it has only two steps, as opposed to the second method that has three.

  • However if your initial publications query returns lots of results (eg for a large institution or big time frame), it may be quicker to try out method 2 instead.

  • The second approach can be handy if one wants to pre-filter researchers using one of the ther available properties (e.g. last_grant_year)

So, in general, deciding which approach is best depends on the specific query filters being used and on the impact these filters have on the overall performance/speed of the data extraction.

There is no fixed rule and a bit of trial & error can go a long way in helping you optimize your data extraction algorithm!


Note

The Dimensions Analytics API allows to carry out sophisticated research data analytics tasks like the ones described on this website. Check out also the associated Github repository for examples, the source code of these tutorials and much more.

../../_images/badge-dimensions-api.svg