Contribution
- dd
Pseudo-feature method
Propose pseudo-feature method : eanable transfer learning when "Upstream and downstream feature sets differ"
- When upstream data is missing a column, (compared to downstream data)
- pre-train a model on the upstream data without that feature
- fine-tune the pre-trained model on downstream data to predict values in the column absend from the upstream data
- After assigning pseudo-values of this feature to the upstream samples, Re-do the pre-training and transfer the feature extractor to the downstream task.
