I'm working on building a training set for a classification problem. I have a choice where I can either build a dataset where each of my subject (customer) can have multiple observations/rows or I can have one observation per subject. Multiple observations per subject are possible when observations are taken from different time windows.

My question: Does keeping multiple observations per subject in the training dataset violate independence of observations? Because let's say two measurements for the same feature for one subject can come from two different but overlapping time window.

  • $\begingroup$ You need to classify a subject into a class? Or you need to classify observations to a subject? $\endgroup$
    – liakoyras
    Nov 8, 2022 at 13:30
  • $\begingroup$ Depends on how we build the data; if one row per subject then classifying a subject; if multiple observations per subject, then classifying an observation. Since ours is not a time series problem, so I am not inclined to the latter, which might led us violating some of the assumptions. $\endgroup$ Nov 8, 2022 at 14:57
  • 1
    $\begingroup$ Defining what you need to classify/predict is the first part of tackling the problem. Find this, and then the rest can be probably answered by yourself $\endgroup$
    – liakoyras
    Nov 8, 2022 at 17:48
  • $\begingroup$ We are developing a customer at-risk/churn model. Off-the-shelf churn models typically classify subjects. But because a customer may be at risk of churn at different periods of the year (macro econmic or personal finance forces driving decline in customer activity), we considered modeling at the observational level. But there is a possibility that multiple observations of a subject measured over some overlapping time period might have some inherent dependence (not sure how to measure and validate it). Due to time constraints, we might only perform one row per subject. $\endgroup$ Nov 8, 2022 at 23:11

1 Answer 1


Yes - Keeping multiple observations per subject in the training dataset violates the independence of observations assumption.

Then the data should be treated as repeated measures or time series, and those types of analysis should be applied.


This site is temporarily in read-only mode and not accepting new answers.

Not the answer you're looking for? Browse other questions tagged .