Data science consultancy in Groningen
We build models that predict, classify or optimise and keep learning from new data, like the recommendation models for See Tickets.
For See Tickets, we developed recommendation models that keep learning
See Tickets is an international ticketing platform. We brought ticket and behavioural data together on a scalable Azure platform and use AI to predict which events will appeal to each visitor. See Tickets is not based in Groningen; this case study shows how we build a model that keeps learning.
Which decision in our organisation could a model predict or help prepare for in advance?
Healthcare and educational institutions and public authorities in Groningen and the northern Netherlands have been recording data on registrations, appointments and applications for years. Yet planning often still follows a fixed schedule. A model based on that history predicts what’s coming; your planner decides.
Predict, classify, optimise or compare: the type of model depends on the decision you need to make
Predicting what someone will choose
For See Tickets, recommendation models predict which events suit each visitor. We tested the recommendations in practice and adjusted them.
Classifying signals automatically
For ProRail, we used machine learning and fibre-optic cables that measure sound waves to automate the detection of trains, railway components and wheel defects.
Optimising timing
For Rijkswaterstaat, we used a congestion model and LSTM to investigate whether data could help optimise the operation of the Botlekbrug and Spijkenisserbrug bridges.
Tracking and comparing candidates
Working with Van Amersfoort Racing, we developed a smart tool that makes it easy for the team to track and compare young drivers.
First the decision and the metric, then the algorithm
-
Decision and metric
We define which decision the model supports and how you’ll tell whether that decision improves.
-
Testing against historical data
The model predicts past periods without knowing the outcomes; then we compare its predictions with what actually happened.
-
Learning in use
Once the model is in use, we monitor the outcomes and retrain it when real-world conditions change.
Do we have enough historical data for our own model?
A good fit when
Your organisation makes the same type of decision repeatedly, such as planning capacity or categorising requests. The relevant data has been kept over a longer period, and someone is ready to use the results.
Not a good fit when
The decision rarely comes up, the recorded data says little about the outcome, or a fixed rule in your system already works. In that case, we’ll advise against building a model and explain what would help instead.
What determines the scope
How many sources we connect, how clean the historical data is, how often the model needs to run, where the results need to appear, and whether your team or ours will manage the model afterwards.
Will our people understand why the model says what it does?
They need to. For each result, we show which factors carry the most weight, and we choose a simpler model if it predicts almost as well. The people using it must be able to check its reasoning.
What kind of model fits your question: a prediction, a classification or an optimisation?
A prediction
Choose this if you want to know what to expect, such as registrations, busy periods or dropouts. The model provides a forecast with a range, and your planner makes the decision.
A classification
Choose this if you need to sort many cases into groups, such as reports, applications or signals. Our advice: start with the groups your staff already use.
An optimisation
Choose this if you’re looking for the best option within fixed constraints, such as a time or a sequence. The model calculates the consequences of each option, as it does for bridge openings.
We’ve been building models since 2016, for organisations well beyond Amsterdam too
Data Science Lab is a Databricks and Microsoft partner based on Westerdoksdijk in Amsterdam. For clients in Groningen and the north of the Netherlands, we work at your premises, at ours or online. We share what we learn at our Techdays.

In our first conversation, we’ll decide together which decision warrants a model
Stef Stoelwinder
Business DevelopmentStef is your first point of contact. In our first conversation, we’ll discuss which decision needs improving and what historical data you already have to support it.
E-mail info@datasciencelab.nl
Phone +31 (0)20 3636 724
What would you like a model to predict or classify for your team?
Tell us which decision you’re making and what data you store. We’ll reply by email or arrange an initial conversation.
Message sent
Thanks for your message. We’ll be in touch as soon as possible.
Frequently asked questions
It builds a model that predicts, classifies or optimises using your own data, tests it against historical data and puts it into use. In Groningen and the north of the Netherlands, this often involves healthcare, education and government.
If you’re looking for data science support in Groningen or elsewhere in the north of the Netherlands, we provide the team to build the model and hand it over. You can find all our applications under AI applications.
We don’t have an office there: we’re based on Westerdoksdijk in Amsterdam. We don’t have a published case study from Groningen either. The case studies on this page are from elsewhere and show the kinds of models we build.
If you first want to work out where AI can add value in your organisation, our AI consultant in Groningen can help you make that choice.
We compare the model with your current approach, such as using last year’s average or the planner’s rule of thumb. We test both against past periods where the outcome is known to see which came closer.
If the model doesn’t clearly perform better, we’ll say so. A rule of thumb that works doesn’t need replacing.
By continuing to monitor the inputs and outcomes, and retraining the model when behaviour changes. For See Tickets, we developed recommendation models that continuously learn and adjusted them based on actual results.
You can read how monitoring and retraining models in production works under MLOps.
You need a record of the outcome you want to predict over an extended period, along with the characteristics you already know when making the decision. For enrolments, for example, that could mean the enrolment date, the course and data from previous years.
If the data concerns people, we include only what is needed for the purpose. You can find out what we provide for a privacy assessment under DPIA for AI applications.
A calculation rule is suitable when the relationship is known and stable, such as a fixed standard per student or per bed. Machine learning is worthwhile when many factors come into play at once and the relationship changes over time.
We often start with the rule as a point of comparison. You can read how we build and explain models under machine learning.
Yes. In Utrecht, the focus is on predictive models for healthcare and services; see data science agency Utrecht. Around Haarlem, we build models that assess; see data science agency Haarlem.
The sequence is the same everywhere: first the decision, then the historical data, then the model.
When the decision is rarely made, there’s too little historical data to learn from, or no one will use the outcome in their work. A model also needs maintaining, and that takes attention.
Sometimes a dashboard showing previous years is enough to plan better. If so, we’ll say that too.