What you have here is called Multi-Instance Learning. From Wikipedia
In machine learning, multiple-instance learning (MIL) is a type of supervised learning. Instead of receiving a set of instances which are individually labeled, the learner receives a set of labeled bags, each containing many instances.
The approach you take in this case is different. You need to bring the Multi-Instance Learning problem into a Single-Instance Learning one. One way you can do this is:
- Perform for each bag of instances a K-Means clustering
- Calculate the Hausdorff distance of each instance from each of the cluster
- Use those distances as features and keep the label from the y_train set
Then apply SMOTE on the new dataset (where you have one row for features and label) and any kind of model of Single-Instance Learning.
You can find details in this Review of Multi-Instance Learning and Its applications