Skip to main content

Models

There can be differnt patterns in the data depending on things like which projects the registration was made on, in which department or work-category. To accomodate for this, we create a hierachy of AI models, each tailored to learn the patterns of that specific sub-groups. We call these models sub-models, and when you get a prediction, only one of our sub-models is resposible for that prediction.

If an employee has enough data, they get their own tailored employee-model. This is considered the most specific model level for a registration, so we will always use this model if it exists. If the employee doesn't have enough or recent enough data to fulfill the model requirements, the registration will move up the hierachy, and you'll recieve predictions from the work-category level, and so on. Below is an figure illustrating the model hierachy:

Model requirements​

To accurately capture the patterns in data, our model creation process necessitates a minimum number of registrations. The requirement for data points varies based on the complexity of the data patterns:

  • For simpler data patterns: Fewer data points are necessary.
  • For more complex data patterns: More data points are required.

These requirements are dynamically calculated through an analysis of the data, which allows for adjustments as new data is uploaded. This dynamic calculation means that the minimum thresholds may be lower than past static requirements.

However, in scenarios involving highly complex datasets, the thresholds revert to the highest, static limits previously set to ensure model accuracy.

These strictest limits are:

  • General Level: 200-500 registrations, spread over at least 30 unique days.
  • Department/Work Type/Project Levels: 100-200 registrations, spread over at least 30 unique days.
  • Employee Level: 50-100 registrations, spread over at least 30 unique days.
  • Employee Aggregated Level: 30-50 (aggregated) registrations, in the last 50 days.

All data utilized must be no older than 360 days to maintain data relevance and accuracy.