Enforcing illegal holiday rentals

Municipality of Amsterdam

There is limited housing and accommodation in Amsterdam, both for Amsterdam residents and visitors. Renting out a house or houseboat to tourists must meet certain conditions.

Last change on 26th of November 2024, at 16:03 (CET) | Publication Standard 1.0

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

General information

Theme

• Economy
• Space and Infrastructure

Begin date

Field not filled in.

Contact information

[e-mail]

Responsible use

Goal and impact

For example, it is allowed for a maximum of 30 nights a year and to 4 people at a time. It must be reported to the municipality. Not everyone abides by these conditions. The municipality sometimes receives reports from neighbours who suspect that a property has been unjustly let. Staff from Supervision & Enforcement investigate them. The algorithm helps prioritise the reports so that limited enforcement capacity can be used efficiently and effectively. By analysing data, the algorithm estimates the risk that the property at the reported address is illegally let. To do so, the algorithm uses data from illegal holiday rentals from the past five years. This includes reports, data collected by the municipality during the work process, personal data (Basic Registration of Persons) and data on buildings (Basic Registration of Addresses and Buildings).

Considerations

Field not filled in.

Human intervention

There is no automated decision-making. When we investigate an address for suspected illegal holiday rentals, it comes from a report from, for example, a local resident or a rental platform. The algorithm helps Supervision & Enforcement staff determine which reports are most likely to be illegal rentals, so those reports can be investigated first. The employee is shown an overview, which shows based on which criteria the algorithm estimates the risk of illegal holiday rentals to be high or low. In this way, we provide insight into what the algorithm has based its risk assessment on. Whether illegal holiday lets are actually present is determined by the supervisor responsible and the project enforcer. A preliminary and field investigation is carried out for this purpose. Subsequently, the file is intensively discussed in a debriefing with the employees who can proceed to decision-making. The algorithm has a substantial influence on the order in which we deal with reports, but does not independently make decisions on whether or not illegal holiday rentals are involved. To prevent employees from putting too much faith in the algorithm, someone has been appointed to monitor the algorithm and its operation. In addition, work instructions have been drafted for the employees who will work with this tool. In doing so, employees attend workshop on the opportunities and risks of using algorithms.

Risk management

The system obviously affects the potential offender. Indeed, a report may be given more or less priority than without the algorithm. We have taken several measures to ensure that all risk estimates of the algorithm are not based on chance. One important measure is that we extensively and continuously evaluate this algorithm in a pilot phase for e.g. reliability, before implementing it in operations.

Operations

Data

Identity and housing data (BRP) A limited set of data from the Basic Registration of Persons (BRP) on the identity and housing situation of the occupants, namely: name of registered occupants; date of birth; Amsterdam departure date; address departure date; address settlement date; date of death. Building data (BAG) A limited set of data from the Basic Registration of Addresses and Buildings (BAG) about the building, namely: address, street code, postal code; id address; description of building; Amsterdam BAG code, national BAG code; type of dwelling (rented, social rented / free sector, sale); addressable BAG ID number of rooms; floor area; floor level of front door flat; number of floors; description floor residential property. Data from illegal holiday rental case Data from the report and any related illegal holiday rental case, namely: case id; date of start of investigation/report; stage of handling of the report; stage number, description and id; (report) code; code violation; code (handling) employee; anonymous reporter/not anonymous reporter; date of report; situation sketch; findings of investigation; user who created report (with date), or adjusted (with date adjustment); handling code (type of case, classification in team); result; date when case was closed; reason why case was closed.

Technical design

Architecture of the model Gemeente Amsterdam has developed an algorithm that can find connections and patterns in large amounts of information about illegal holiday rentals. The algorithm calculates which information is more likely to be associated with illegal holiday rentals and which information is not. The algorithm does this by performing mathematical calculations according to the probability tree principle. That is, the algorithm takes an average based on a large number of probabilities. This average is used to generate the mathematical expectation of illegal holiday rentals at an address. The algorithm calculates this expectation only when we receive a new report (e.g. from a local resident or rental platform) of possible illegal holiday rentals. This algorithm is called a 'random forest regression algorithm'. To make the trade-offs made by the algorithm understandable to people, we apply the "SHAP" methodology (SHapley Additive exPlanations; https://github.com/slundberg/shap). For each individual case, SHAP calculates which indicators contributed to that prediction and whether this caused the prediction to be higher or lower. This way, an employee can always see what the algorithm based the risk estimate on and make a considered decision. Performance The advantage of a 'random forest regression' is that it is a reasonably complex algorithm that can closely approximate reality. But there is a chance of overfitting. A 'tree' with many layers squeezes the data to deliver specific answers. It has been studied how many layers the model needs to remain generically deployable and thus not overfitting. In addition, data points are continuously categorised (grouped) so that the model has a manageable number of possibilities instead of an infinite number with continuous values. This improves the model's ability to reach a conclusion.