A key aspect of the machine learning models of the kind employed by Aible is that each record in the training data assumed to be independent.
This is not the case for autocorrelated data such as time series data where the value in a record at a given timestamp has a relationship the value in the preceding record(s).
For instance, if we’re measuring the outside temperature every 5 minutes the value at 10:25 is closely related to the value at 10:20 - it is not independent of the other values in the dataset:
This relationship information is lost as the model has no understanding of a sequence in the data. By default Aible Sense will find patterns relating to the day or week or the month of the sample (or indeed any date part extracted via a transformation) but we can also help the AI by restructuring the data slightly to include the time sequence relationship.
The best way to do this is to add a column (or columns) containing the values for the previous period(s).
For example. here is a time series dataset with a daily period:
To help the model to include the relationship we can add the value for the preceding day as an extra column:
We can go further and add more days of history to each row:
With this change the model can learn the relationships between the values on a given day and the values leading up to that day.
Obviously we’ll want additional descriptive dimension columns in the data to find additional patterns and relationships.
If you would like an example python script to make this transformation to your own data get in touch with success@aible.com.