Pig welfare compliance model

Netherlands Food and Consumer Product Safety Authority (NVWA)

The NVWA must check whether pig farmers take adequate care of their animals. This algorithm helps companies find where the risks are highest.

Last change on 22nd of April 2025, at 10:01 (CET) | Publication Standard 1.0

• Publication category: Impactful algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

Economy

Begin date

2022-03

Contact information

https://www.nvwa.nl/over-de-nvwa/contact

Link to publication website

https://www.nvwa.nl/onderwerpen/dierenwelzijn/inspectieresultaten/inspectieresultaten-2023/inspectieresultaten-dierenwelzijn-varkens-2023

Responsible use

Goal and impact

This algorithm is designed to make the most useful use of our inspectors' time. Some inspections we do at farms chosen by chance. This way, we keep a good eye on how pig farms in general are doing. We also do inspections because we have received a report about something that may not be allowed by law. We then want to sort that out quickly and, if necessary, deal with it. Furthermore, we mainly try to inspect farms where it seems more likely that the rules are not being followed. This algorithm helps find such companies. This is followed by an inspection as we always do. It leads us to do more inspections at companies where something is wrong, and less at companies where everything is going well. Inspecting where the risks seem greatest is required by law; with this algorithm, we can do it better.

Considerations

This algorithm helps us find farms where pigs are more likely not to be kept according to the law. Then we can intervene and make sure the situation improves. That is a big advantage. The algorithm also ensures that we can be sure that we calculate risks in the same way for all farms, with information that we are allowed to use for that purpose. That is also an advantage.

A possible disadvantage is that the algorithm learns from previous inspections. This could lead to the algorithm suggesting the same companies over and over again, or skipping them. Another possible disadvantage is that inspectors might see more problems if they know they are at a company because an algorithm has suggested it. This could lead to some companies getting more fines because the algorithm is better at finding them, not because they are worse at complying with the law. But because we know these risks are there, we can also take measures to keep them small. For example, we ensure that inspectors do not know for sure whether a company has been selected by the algorithm. That way, their judgement is not affected.

We believe that all in all, the advantages outweigh the disadvantages.

Human intervention

This algorithm predicts the probability of non-compliance for each pig farm. From the farms with the highest probabilities, we make a list of farms to inspect. The number of farms on the list is determined by humans, and the list itself is also checked by hand. If we then discover possible errors or improvements, we adjust the algorithm. Then each company on the list is visited by an inspector. So the algorithm only gives advice on which pig farms to inspect. The algorithm does not make decisions and does not advise on the inspection itself.

Risk management

As we wrote earlier, we see two main risks in using this algorithm. The first is that we might inspect some pig farms too often or too little. To prevent this, we are taking several measures. First, we always keep doing inspections at pig farms that are selected by chance. This way, we check whether the algorithm really helps detect more problems than we would otherwise. Second, we compare the list of pig farms suggested by the algorithm with the list of all pig farms out there. If we find that the algorithm no longer predicts well enough where there are problems, or always picks or skips the same kind of farms, we improve it, or don't use it. We also ask inspectors to share their experiences based on the inspections performed. Their feedback helps us to further improve the algorithm.

The second risk is that the algorithm could influence what inspectors think of the companies they inspect. To make sure this happens as little as possible, we arrange it so that an inspector is never sure whether a company has been suggested by the algorithm. In addition, we properly explain that the algorithm only calculates a probability. The fact that a pig farm has been suggested by the algorithm really does not mean there is anything wrong with it; only that it looks like pig farms where we have found a problem before.

In general, we think the risks of the algorithm are small. There is a lot of human control; we also do inspections that the algorithm has nothing to do with; and the total number of farms inspected on the advice of the algorithm is not that large.

Legal basis

• The Animals Act
• Animal Husbandry Decree

Links to legal bases

• De Wet Dieren: https://wetten.overheid.nl/BWBR0030250/2024-07-01
• Besluit houders van dieren : https://wetten.overheid.nl/BWBR0035217/2024-07-01

Operations

Data

The algorithm uses data from four sources:

1. Our own inspection results,
2. the system in which pig farmers must report the movement or death of their animals ('I&R'),
3. the 'agricultural census', in which all farmers must report the characteristics of their farms annually,
4. the 'basic administration of addresses and buildings', which contains data on all Dutch buildings ('BAG').

Links to data sources

• I&R: https://www.rvo.nl/onderwerpen/identificatie-en-registratie-dieren/varkens-melden/varkens-verplaatsen
• landbouwtelling: https://www.rvo.nl/onderwerpen/gecombineerde-opgave
• BAG: https://www.kadaster.nl/zakelijk/registraties/basisregistraties/bag

Technical design

So far, the algorithm has been revamped with every use. This involves comparing different supervised machine learning techniques and choosing the best predictive one. We use these techniques to automatically learn the relationship between business characteristics and inspection results. To do so, we divide all available inspections into a training set and a test set. The inspections in the training set are used to learn the correlation; the inspections in the test set are used to test whether the calculated correlation is also predictive of inspections that the algorithm has not seen before.

External provider

The algorithm was created by the NWVA itself.