I have a large dataset with around 200 features, consisting mostly of timeseries and categorical data, with some continuous. The dataset is extracted from/by a postal service. Small example:

Random (scrambled) entries:

  shipment        delivery          cost        location                weight_kg

 2020-04-22      2020-04-23         77.31       UK:66c54f531....           0.5
 2020-04-23      2020-04-25         44.14       DK:22c54f531....           2.23
 2020-04-24      2020-04-27         53.84       UK:66c54f531....           1.57 
 2020-04-25      2020-04-26         22.09       UK:66c54f531....            

My first inclination was to make a demand-forecast model on shipment/count_monthly(shipment), but considering the amount of features, a multivariate case seemed more relevant. I am just not sure which additional features to add - and without this project becoming to generic (linear regression). Mine initial EDA depicted variables with low correlation, or removed otherwise to avoid multicollinearity.

Then, instead I considered a clustering approach, to gather and depict relations between the features in more detail. Just not sure how to approach it with such a data size and with timeseries, never really worked with that dtype, especially in combination with categorical dtypes. Any advice would be appreciated.

Edit: the various date columns (like shipment and delivery) are not chronological, and their values appear numerous times, thus cannot be timeseries either. This begs another question: does it even make sense to convert the columns in question to a datetime object?

  • $\begingroup$ Did you progress on this project? If yes what was the final approach you took? $\endgroup$ May 22, 2022 at 4:55

1 Answer 1


Most clustering algorithms are designed to work with numeric features.

Those date-related columns can be converted to numeric features. Shattering is one way to convert them to numeric features, create separate features for year, month, and day.

Additional time-based features can be created such as length of shipment.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.

Not the answer you're looking for? Browse other questions tagged or ask your own question.