Given a flat structured data with features that can be considered hierarchical, where each feature is at a different level (e.g., Brand at the top level, Product, Color, and Size at different levels), I am trying to make predictions at each level while using the same target variable. However, if I start predicting from the lowest level, I will have all the features present in the data. If I start predicting at a higher level (e.g., Color), I will not have data for the lower level (e.g., Size). Therefore, I need to know: Should I use a single model for all levels or different models for each level? If I use a single model, what approach should I use?
Single model or multiple models for predicting at each level in a multi-level classification problem
1 Answer
$\begingroup$
$\endgroup$
9
if we consider your dataset as DB table, you denormalized your data and flattened it. So you don't have to think features with their levels. Of course they may have some relationship as you said, and it may have some multicollinearity. But you don't need to consider this situiation for model choice approach.
-
$\begingroup$ if i treat all features on same level, let us say i have two scenarios, 1) predict the target using all the four features brand, product, color and size 2) predict the target using brand, product, color (I don't have the size data available here). So my question is should i consider building two models for above two scenarios or just a single model would be sufficient to handle both the scenarios? $\endgroup$ Mar 20 at 13:52
-
$\begingroup$ you are actually doing some feature selection. if this is a classification model, you may check importance of features and then finalize your feature selection. and then fit into it a single model. this approach is more proper. $\endgroup$– AtacanMar 20 at 14:56
-
$\begingroup$ Nope, this is not feature selection, i am past the feature selection process and now at the prediction. It is a use case where i have to predict using all the features at the first and then i have to predict using all the features except the size feature. so in this case i'm confused to use single or multiple models? $\endgroup$ Mar 21 at 9:51
-
$\begingroup$ ok, let's assume you pass the feature selection step. so did you try to fit it with or without size feature? which one is better? and if you want to make multiple models for this same target, what is your advantage of it? $\endgroup$– AtacanMar 21 at 12:40
-
$\begingroup$ Yes in the first step i fit including the size feature and made predictions, in the next step i have make prediction without the size feature as i don't have the data for size at this step, here the question is to build another model and train it without the size feature or is there any model which can prediction including\excluding the size features? $\endgroup$ Mar 21 at 13:02