Adam Day

*TL;DR: The current version of the Papermill Alarm detects signals in
98.9% of the Hindawi retractions conducted over the last 12 months.\*\*
It's never nice to see the harms caused by papermills, but it is good to
see independent verification of the Papermill Alarm's predictions.*

Here's a question: Why are there 2 multi-squillion-euro detectors at
CERN's Large Hadron Collider? Why not save some money and just build 1?
[It's because you need one detector to independently verify what the
other is
detecting](https://home.cern/science/experiments#:~:text=Having%20two%20independently%20designed%20detectors,for%20focussing%20on%20specific%20phenomena).
Independent verification (or *reproducibility*) is a cornerstone
of science.

Peer-review is a relatively new invention. It's a system for checking
the scientific content of research. It is imperfect and it is
fundamentally based on trust.

Peer-review *isn't* a system for checking for scientific fraud. If
someone makes up their results, then peer-review isn't supposed to catch
that. *It might*, but it's not realistic to expect someone to reproduce
the results in peer-review because it might not be possible (e.g. in
cases where you would need a squillion euros to rebuild the experiment).
That means that, to some extent, reviewers have to take authors' claims
on faith.

It's because publishers run the peer-review process that they have
become the de facto gatekeepers of science. Now that we have clear
widespread scientific misconduct, publishers are thrust into the role of
research-cops tasked with policing, detective work, and even mopping up
the crime-scenes.

Publishers are part-responsible for research integrity, so that's fair
to an extent. It just means building capabilities outside of traditional
peer-review. But I think it's also fair to say that, compared with other
responsible parties: authors, institutions, funders, editors etc
publishers seem to cop a disproportionate share of the flak when things
go wrong. Again --- there's a big difference between peer-reviewing
science and detecting fraud.

I'm not the first to say it, but the solution to improving research
integrity is for all of these parties to work together and find ways to
adapt to the problem *together*. The problem isn't research integrity;
*there is no research in a fake paper.* On the other hand, the solution
*is* integrity. Honesty, accuracy and reproducibility could all go a
long way.

Hindawi's work to clean up the papermill problem in their journals is
laudable. They showed integrity in admitting mistakes and rectifying
them. They have made significant steps so far. It can't have been fun or
motivating work, but it was the responsible thing to do and that should
be recognised.

At Clear Skies, what we want is to catch as many cases as possible prior
to peer-review so that publishers can focus on quality service instead
of having to constantly hunt for fraud. Hindawi's retractions give us
some interesting insights there.

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/1*LtEmJ4_ZLe_k6F7trN4LWQ.png" />
</figure>

The current version of the Papermill Alarm detects signals in **98.9%**
of the Hindawi retractions conducted over the last 12 months.\*\* It's
never nice to see the harms caused by papermills, but it is good to see
independent verification of the Papermill Alarm's predictions.

If you are interested in using the Papermill Alarm, either through your
peer-review platform, our API, or our webapp, [please get
in touch](https://clear-skies.co.uk/contact/).

#### Appendix: Data caveats

\*\* In the interest of responsible science, it's worth making a few
nerdy points about the data.

- The Papermill Alarm raises the alarm when the probability of a paper
  being connected to papermilling is significant. [There's a balancing
  act here which was described in a previous blog
  post](https://clearskiesadam.medium.com/where-to-draw-the-line-precision-and-recall-3c39a131c14b).
  So, while we find specific signals in 98.9% of Hindawi's retractions,
  we would trigger alerts in 90.2% of them. At that level, given that
  most of the retractions are concentrated in special issues containing
  multiple papers, I expect that we would catch all of those issues.
- The Papermill Alarm actually learns from retractions automatically. So
  in order to get the above figures, I had to remove Hindawi's
  retractions from the inputs to stop the Papermill Alarm from
  predicting something it had already seen.
- These figures come from a version of the tool that has not yet been
  released to the API, or webapp, so your mileage may vary if you are
  using those tools until the updates are released.
- When we talk about 'prediction' here, we are saying that the Papermill
  Alarm could find something that it hadn't seen before, but I should be
  clear that the retractions happened before the current version of the
  Papermill Alarm was created.

![](https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=08aadd1f3a98){width="1"
height="1"}


<em>
 TL;DR: The current version of the Papermill Alarm detects signals in 98.9% of the Hindawi retractions conducted over the last 12 months.** It’s never nice to see the harms caused by papermills, but it is good to see independent verification of the Papermill Alarm’s predictions.
</em>
Here’s a question: Why are there 2 multi-squillion-euro detectors at CERN’s Large Hadron Collider?

Predicting retractions with the Papermill Alarm

TL;DR: We can detect individuals with a high probability of being
involved in milling papers. The question is: how should we respond?

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/1*VWPzClM15Rz8xz5lzfLaMg.png" />
<figcaption>Keystone connects any paper to the network of milled papers
and calculates that article’s level of connection to papermilling. This
image shows the mathematical ‘centre’ of the coauthorship part of the
network. I like to imagine there’s some sort of papermill-godfather
figure in there somewhere, or at least a <a
href="https://en.wikipedia.org/wiki/Del_Boy">Del boy</a>.</figcaption>
</figure>

A few years ago, my bike was stolen. It was an organised job. The
thieves arrived after dark, cut a hefty lock clean off, and my bike
disappeared silently forever. I imagined it dismantled and sold off into
some vast black market network. Poor bike. The police admitted that they
weren't going to do anything. I guess that if they catch someone
stealing a bike there isn't a lot they can do anyway. It isn't a
serious-enough crime to lock someone up.

If a crime is easy to do, pays well, and has no punishment, then that is
an endemic crime. It isn't going away because the people who do it have
incentives and don't have disincentives.

Organised research fraud fits a similar niche.

It's encouraging to see so much movement to change how we respond to
research misconduct in recent months with initiatives like United2Act
and working groups organised by the STM Association.

It's clear that historic practices for responding to misconduct are not
the right ones for today. (If they were, there wouldn't be so much of
it.) So discussion about the way forward is very welcome.

There's a problem with retractions. It's not that we shouldn't do them;
they help to clean up the literature. But if you mill fake papers for
people under a pseudonym and those fake papers get retracted, that
doesn't necessarily affect you. So retraction isn't a disincentive much
of the time.

[I spoke last year about the importance of KYA --- Know Your
Author](https://youtu.be/yOvpsQ7bDtA?si=SP8NEoT5OHDKJTa2&t=1281). If we
know who our authors are, we can build relationships and trust. That
gives us confidence in dealing with people we know and tells us when we
need to get to know people better.

Because it's all about relationships.

"Keystone" is a network analysis built on OpenAlex that I've been
running in one form or another since shortly after the Papermill Alarm
was released in 2022. We analyse the relationships between papers,
individuals, institutions and other signals to learn about the papermill
problem. Last year, Keystone went through some substantial development.
It is now integrated into the Papermill Alarm and is already in use by a
number of publishers.

I've spent a lot of time with Keystone and I'm seeing some interesting
things.

- We can tell which researchers are most likely to be running papermill
  operations.
- Some Keystone-authors' papers are real. This puts an interesting spin
  on investigations. That 'KYA' principle applies: an author only needs
  to have one verified fake in their history for me to be concerned
  about *all* of their output.

**So that begs the question: if papermill detection is 'done' to enough
of a degree that we can identify the current generation of bad actors,
how should we respond so that there isn't another one like this?**

I've said before that we don't prescribe how to act on alerts from Clear
Skies' services. But, if it was my journal, and I had cases to
investigate, I would start with Keystone. Doing so targets the source of
the problem and disincentivises practicing researchers from being
involved with milling. It's bang for buck.

Would you buy an authorship slot from someone with a dozen retractions
on their record?

![](https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3a48de927406){width="1"
height="1"}


TL;DR: We can detect individuals with a high probability of being involved in milling papers. The question is: how should we respond?  A few years ago, my bike was stolen. It was an organised job. The thieves arrived after dark, cut a hefty lock clean off, and my bike disappeared silently forever. I imagined it dismantled and sold off into some vast black market network. Poor bike. The police admitted that they weren’t going to do anything.

Keystone: papermill detection with network analysis

There's this phrase used by some technologists: '[epistemic
security](https://www.turing.ac.uk/sites/default/files/2020-10/epistemic-security-report_final.pdf)'.
Epistemic security has to do with things like the spreading of
misinformation on social media. I.e. if we spend enough time on social
media reading unreliable information, how does that influence us? That's
a matter of epistemic security. Epistemic security is interesting
because if you start a conversation about epistemic security at a party,
they don't invite you back.

Publishing is a simple process in principle.

<figure>
<img
src="https://cdn-images-1.medium.com/max/399/1*LviW6ddetmF-1QOhl8d4kQ.png" />
</figure>

So peer-review is an epistemic security system. Peer-review stops bad
papers getting through --- and, therefore, stops bad information getting
into our brains. But, there's a problem. Peer-review isn't perfect and
it was never intended to be. [Security systems have holes... like
Swiss cheese](https://en.wikipedia.org/wiki/Swiss_cheese_model).

I used to edit journals and I always wanted to get the best referees on
every paper. But the reality is that sometimes you need to hold the best
reviewers in reserve so that they are available when needed. You also
need to give inexperienced reviewers the opportunity to learn and that
means accepting that the best standard of review isn't always possible.

<figure>
<img
src="https://cdn-images-1.medium.com/max/938/1*CXrDuBcDAvGGJsPT-NQYzw.png" />
</figure>

The Papermill Alarm is an early-warning system which alerts you to
papers with signs of papermilling. It helps to fill-in those holes. When
an editor sees an alert, they know this is one of those times when they
need to call on their best reviewers. It's an easy decision to make as
an editor and it turns out that it's an easy solution to the
problem, too.

We don't actually prescribe how editors use the output from the
Papermill Alarm, but one thing we've learned: peer-review is effective
against papermills. So applying the best standard of peer-review is an
effective control.

- I'm yet to see a situation where the rejection rate for papers which
  receive 'red' alerts on a journal isn't significantly higher than the
  rejection rate for other papers.
- We know that papermills [often try to manipulate the peer-review
  process](https://clearskiesadam.medium.com/papermills-in-peer-review-c507b12aa4ac)
  (or circumvent it entirely --- e.g. by abusing special issue
  programmes). They wouldn't do that if peer-review wasn't a problem
  for them.

That's nice, isn't it? It turns out that the thing we have always been
good at --- peer-review --- is already the best tool for the job. The
Papermill Alarm raises the alarm so that editors know when to take care.

So that's why we recommend using peer-review to control the problem. But
it's up to you how you use these insights. There might be a couple of
things to think about here...

### Detect and Respond

I like this quote from Bruce Schneier. It's about dealing with *novel*
computer hacks. But it applies to novel research fraud, too.

> "You can't prevent. You can't defend. All you can do is detect
> and respond"

It sounds defeatist at first, doesn't it? But I don't read it that way:
it makes the road forward clear and the nice thing is that we can
already do the detection part.

Data science is the process of extracting value from data. With the
Papermill Alarm, we've been very lucky to have not only high quality
public datasets to start off with like Crossref, PubMed and OpenAlex,
but also to have received data from publishers which describes the
problem uniquely from multiple perspectives. It's been said before, but
the key to solving this problem is cooperation and it's by working
together that we've been able to build this unique data resource.

When it comes to papermills, the 'hack' isn't novel any more. We can
detect the overwhelming majority of problem papers before peer-review.
Furthermore, using new methods, we can identify the individuals running
the mills as well.

So detection is done. The question now is, given this rare opportunity,
what is the best way to respond?

### Testing the tools

Recently, I made an estimate for the global rate of papermilling
[(reported in Nature News by Richard Van
Noorden](https://www.nature.com/articles/d41586-023-03464-x)).

<figure>
<img
src="https://cdn-images-1.medium.com/max/940/1*TUIJAHD5hkT5VMRaIWM5kQ.png" />
</figure>

I did the work back in April 2023 using one of our analytics pipelines.

I already had some good validation of the method at that time, [such as
the Chinese Academy of Sciences Early Warning
Lists](https://medium.com/@clearskiesadam/cast-a-wide-net-c083747cdb71).
The 'red' alerts are engineered to have zero (or near-zero) false
positives in testing. However, test data isn't the same as real-world
data, so we needed to confirm it was performing well in the wild.

What we are seeing here is a trend in which particular template titles
and abstracts appear in the literature. We *know* that the same
templates are used by papermills and how to characterise them.

I spent much of the last year combing through the predictions from this
method. What I found was that I would see dozens of journals with no
alerts at all, and then I would see one journal with dozens or even
hundreds of alerts.

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/1*GGJz8tJPvxZufMWiNJmahw.png" />
</figure>

It appeared that these rare journals had been targeted and so I
spot-checked them by picking out individual papers and manually looking
for problems. Once the spot check had confirmed that the false-positive
rate was low, I would move on to the next journal.

While this was happening, I would report the findings to publishers.
That led to some very useful feedback. Sometimes the journals were
already under investigation and we even predicted a few mass-retractions
that way.

<figure>
<img
src="https://cdn-images-1.medium.com/max/1024/1*0cJe3dWHUeFXsRxVzF17PA.png" />
</figure>

*I'd also like to acknowledge valuable feedback gratefully received from
a community of sleuths throughout this process including Guillaume
Cabanac, Cyril Labbé, Alexander Magazinov, Elisabeth Bik, Smut Clyde,
Fidelia, Sholto David, Kaveh Bazargan, Cheshire.*

At the journal level, when the alert-rate is high, we always find
something.

There was another thing that was important here. Some publishers had
almost no signal at all. I'd talk to them and they would tell me that
they didn't get papermills submitting because they were tough on
peer-review and they thought that tough peer-review put the mills off.
(And, as I said above, our data analysis on peer-review seems to confirm
that they were right.)

So we know the false-positive rate is low. The chart above gives us
1.7--1.8%, but we have some potential for overestimation (e.g. there is
a small amount of missing data which we can't see). So I'm comfortable
saying that I think 1.5% is a good rough lower-bound for the rate of
papermilling in 2022.

Most of our methods are intended to be run at the article-level. So for
the estimate, I just used one method that I knew would work at scale.

At this point, there's certainly scope for someone to do a better
estimate. But I don't think it's possible that would yield a rate lower
than 1.5%.

![](https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=90b10fb8a6cc){width="1"
height="1"}


There’s this phrase used by some technologists: ‘epistemic security’. Epistemic security has to do with things like the spreading of misinformation on social media. I.e. if we spend enough time on social media reading unreliable information, how does that influence us? That’s a matter of epistemic security. Epistemic security is interesting because if you start a conversation about epistemic security at a party, they don’t invite you back.

The Papermill Alarm

The Papermill Alarm draws a line between papers which have the
characteristics of papermill-products and those that don't. Given data
like that shown below, where red dots represent papermill-products and
green dots represent normal papers, where should we draw that line?

<figure>
<img
src="https://cdn-images-1.medium.com/max/939/1*QtxNBVi_PwaEgOuuYW-87Q.png" />
<figcaption>Some dummy data to illustrate the point. Interesting aside:
we actually use dummy data like this all the time as we develop
processes at Clear Skies. This prevents the sensitive data that we work
with being used or seen outside of the context where it is
necessary.</figcaption>
</figure>

### Limit the spread

There's well-known (and fun!) thought experiment in statistics that goes
something like this:

There is a rare disease affecting one in ten thousand people. Someone
creates a diagnostic test for this disease with 99% accuracy. You take
the test and it comes up positive. What is the probability that you have
the disease?

- 99%
- 98%
- 50%
- 0.98%

The obvious answer seems to be 99%. The test is 99% accurate after all.
But let's think about it a different way.

If we have 1,000,000 people (and the disease affects 1 in 10,000 of
those people), then 100 of those people have the disease and 999,900
don't have the disease. Let's consider those groups separately.

The test is 99% accurate. This means:

- when we run the test on the 100 people who do have the disease, then
  99 of them will be correctly diagnosed and 1 person will be
  misdiagnosed as not having the disease: a *false-negative.*
- However, when we run the test on the other 999,900 patients, **who do
  not have the disease**, 99% of them are correctly diagnosed as *not*
  having the disease, but 1% of them (9,999 people) will be
  'false-positives'. I.e. they will be incorrectly diagnosed as having
  the disease.

So, of the 10,098 people who test positive, only 99 actually have the
disease. **That means that if you test positive, even with 99% accuracy,
the chance that you have this disease turns out to be less than 1%!**

Isn't that strange?

### Precision and recall

The above is [a common example in Bayesian
statistics](https://www.youtube.com/watch?v=vZd5ZqWwTZI). What it means
for [Clear Skies](https://clear-skies.co.uk/) is that, if we want to
build detection algorithms for papermill-products, we have to control
for imbalance in the data in order to be able to make useful
predictions. We also have to be a bit more careful about what we mean
when we say 'accuracy'. There are 2 very handy metrics here.

#### Precision

What % of the detections that you make are accurate?

- High precision means that we can have a lot of confidence in
  predictions, but that we should expect to have a lot of *false
  negatives.*
- In the above example, precision is 0.98%

#### Recall

What % of the things you want to detect are you detecting?

- High recall means that we can expect to find a high percentage of what
  we are looking for, but the downside is that we will also see a lot of
  *false positives.*
- In the above example, recall is 99%

So, our imaginary diagnostic test has high recall and low precision. I
know that medical researchers sometimes read this blog. [Your words are
different to
ours](https://towardsdatascience.com/should-i-look-at-precision-recall-or-specificity-sensitivity-3946158aace1#:~:text=So%2C%20let's%20look%20at%20the,how%20many%20got%20negative%20results%3F),
but hopefully the concepts are familiar!

*(Also, I'm not a doctor, but I assume that high recall is a good choice
for medical diagnostics. If you get misdiagnosed as having a disease,
that's not good. But you might be able to run other tests for
confirmation. It seems far worse if you were misdiagnosed as* ***not***
*having a disease that you need to be treated for.)*

When it comes to papermill detection, high recall would mean a situation
like the one described above where we flag or 'diagnose' a very high
percentage of papers, including a lot of non-papermill products, in
order to ensure catching all of the papermill-products. The next stage
of the process is to run further tests (one reason why we have multiple
pipelines!) or for Research Integrity specialists to sift through those
flagged papers and apply their own checks. In principle, this is quite a
nice 2-stage process which can maximise the number of papermill-products
we detect correctly.

However, there's a downside here. We don't want to waste the Research
Integrity specialist's time by flagging lots of papers which are likely
to be ok.

But here's the cool thing: we actually get to choose whether we want
high precision or high recall. That's nice, isn't it? Which should
we pick?

<figure>
<img
src="https://cdn-images-1.medium.com/max/939/1*DwHPpOGZZkkS6zV-EyZoTA.png" />
<figcaption>In this example, we classify everything above the line as a
‘red’ alert. If we go for high precision, we miss a lot of papermills,
but if we go for high recall, we might flag too many
non-papermill-products.</figcaption>
</figure>

High precision would mean accepting that we miss a few
papermill-products, but that a very high percentage of detections are
what we are looking for.

We could try to find a balance in between the 2. But we can do better
than that.

We know that papermills will test-the-waters with journals all the time.
When they find one that has a high acceptance rate for their products,
they target it.

<figure>
<img
src="https://cdn-images-1.medium.com/max/678/1*N8EJnBNIco0bNULhdyGjQQ.png" />
<figcaption>The monthly output of one journal by Papermill Alarm
alert-rating (real data this time!). It looks like this journal had an
ongoing problem for a while, but something caused the alert-rate to
shoot up. I can’t say exactly what that thing was in this case, but I
can say that - while most of the world’s journals have no alerts - this
is a common pattern in the ones that do.</figcaption>
</figure>

So the effect of using a high-precision tool would be to reduce the
chances of being targeted with a minimal amount of work (compared with a
high-recall tool). When you think about it, we don't need 100% recall.
We just need to catch enough that it isn't worth the mill's time to keep
submitting.

So there's value in both. We want to have a high precision option and a
high recall option. That's why the Papermill Alarm issues red alerts and
orange alerts. Red alerts have high precision and orange alerts have
high recall.

So you get the best of both worlds.

[Contact us for more details](https://clear-skies.co.uk/contact/)

![](https://medium.com/_/stat?event=post.clientViewed&referrerSource=full_rss&postId=3c39a131c14b){width="1"
height="1"}


The Papermill Alarm draws a line between papers which have the characteristics of papermill-products and those that don’t. Given data like that shown below, where red dots represent papermill-products and green dots represent normal papers, where should we draw that line?

Rogue Scholar Posts

Predicting retractions with the Papermill Alarm

Keystone: papermill detection with network analysis

The Papermill Alarm

Where to draw the line? Precision and recall