Public space reports

Municipality of Amsterdam

If something needs to be fixed or cleaned up on the street or in a park, it can be reported to the municipality through SIA, the online reporting system.

Last change on 27th of November 2024, at 17:02 (CET) | Publication Standard 1.0

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

General information

Theme

Public Order and Safety

Begin date

Field not filled in.

Contact information

[e-mail]

Responsible use

Goal and impact

If something needs to be fixed or cleaned up on the street or in a park, it can be reported to the municipality through SIA, the online reporting system. A dangerous traffic situation or nuisance from people and catering establishments can also be reported.

Previously, people had to choose which category their report fit best (e.g. 'nuisance' or 'street furniture'), so that the report went to the right department in the municipality. But the municipality is a complex organisation and the list of categories is long. As a result, people did not always choose the right category. This sometimes caused delays in processing notifications.

Therefore, we now use an algorithm that recognises words, e.g. 'rubbish' and 'pavement'. On this basis, it determines which category the report best fits and which department should handle it.

So the caller no longer has to choose a category, and the call is handled faster because it goes to the right department.

Considerations

Field not filled in.

Human intervention

All reports categorised with less than 40% certainty are forwarded to the Action Service Centre. A human assessment then takes place and the report is still categorised. Reports that are wrongly forwarded to an incorrect category are also manually placed in the correct category by the responsible department (sometimes via the Action Service Centre).

Risk management

There are few risks in this algorithm. It places a notification in the right category and gets it to the attention of the right department faster. If the algorithm cannot place a notification in a category with sufficient certainty, it ends up in the 'Other' category. Staff at the Action Service Centre review those reports and manually put them in the right category. If the algorithm does not work properly, it therefore takes a little longer for the report to reach the right department. The reporter can add personal data if he or she wants to be kept informed. This data is stored securely and not used by the algorithm.

Operations

Data

The dataset consists of reports made previously (free text field). Initially, we used 300,000 notifications from past years to train the algorithm. It is regularly retrained with new reports and implemented corrections to existing reports. If the Action Service Centre or departments see an incorrect categorisation, (see Human Monitoring) they correct it manually in the notification system. These corrections are used in re-training. We are investigating whether retraining the algorithm can be automated in the future.

We cannot disclose this dataset in this register. Because the data comes from a free text field, it may contain personal data, although this is explicitly not requested.

Email address and telephone number for follow-up questions

Someone who makes a report can leave his or her telephone number and/or email address if he or she wishes. We will then keep the reporter informed of progress and have the option of calling him or her back. This information is not kept longer than necessary for this purpose and is therefore not used by the algorithm.

For other information on processing personal data, see: https://www.amsterdam.nl/privacy/specifieke/privacyverklaringen-wonen/meldingen-overlast-privacy

Technical design

The text of the notification is broken down into individual words. Each word from a notification is analysed to see how unique it is to that notification, compared to the total collection of words ('TF-IDF' or 'term frequency-inverse document frequency'). As a result, a word like 'the' or 'thank you' gets a low weight and a word like 'rubbish' gets a higher weight.

From that combination of words, logistic regression (a machine-learning technique) is then used to determine which category the report belongs to and thus which department within the municipality the report is most likely to fit.

This algorithm can very accurately detect in which category a combination of words belongs; the algorithm has a score of 0.88 (macro-weighted F1 score). Other methods have also been tried (W2V, CNN+LSTM, BERT) but were found to perform less well. More information: https://medium.com/maarten-sukel/how-to-use-machine-learning-for-the-classification-of-citizen-service-requests-b71159a85f36