One study, still running.
I am working on a reproducible study of training-time masking: whether deliberately hiding features during training makes classical tabular classifiers hold up when the data they meet later has gaps the training data did not.
It is not published and there are no results to report yet. What exists is a fixed protocol and a working pipeline, which is the part worth being careful about, because tabular benchmarks are unusually easy to accidentally cheat at.
1 study · nothing published · verified 2026-08-23
Status
Active
- Active work
- 1 study
- Published
- None yet
- Stage
- In progress, unpublished
Training-time masking for robust tabular classification
Undergraduate researcher
Thapar Institute of Engineering & Technology
External research project, alongside the degree at IIT Jodhpur
Ongoing
The question
If a model will eventually see records with missing fields, does simulating that damage during training buy real robustness, or just a worse model?
A controlled comparison of masking strategies for classical tabular classification under synthetic missingness shift. The point is the protocol as much as the finding: same preprocessing, same splits, same model, with the masking regime as the only thing that varies.
Design
- Three training regimes compared: clean training, uniform feature masking, and structured block masking.
- Logistic Regression as the classifier, held fixed so the masking regime is the only moving part.
- Evaluation across both clean and masked test conditions, so a regime cannot look good by simply degrading everything equally.
Keeping it honest
- Leakage-aware preprocessing: every transform is fitted on the training split only, never on the full dataset.
- Median imputation paired with explicit missingness indicators, so the model can distinguish an imputed value from an observed one.
- Standardization and one-hot encoding applied inside the same train-only discipline.
- Missingness is synthetic and controlled, which is what makes the shift measurable rather than anecdotal.
Datasets
- Bank Marketing
- Online Shoppers Purchasing Intention
No results published. There is no paper and no public repository yet, so there is nothing here to link and nothing to cite.
Less formal experiments and build logs live in the workbench.