Classification with Strategically Withheld Data

Machine learning techniques can be useful in applications such as credit approval and college admission. However, to classified more favorably contexts, an agent may decide strategically withhold some of her features, bad test scores. This is a missing data problem with twist: which depends on the chosen classifier, because specific classifier what create incentive certain feature values. We address training classifiers that are robust this behavior. design three classification methods: MINCUT, Hill-Climbing (HC) Incentive-Compatible Logistic Regression (IC-LR). show MINCUT optimal when true distribution fully known. it produce complex decision boundaries, hence prone overfitting cases. Based characterization truthful (i.e., those give no hide features), we devise simpler alternative called HC consists hierarchical ensemble out-of-the-box classifiers, trained using specialized hill-climbing procedure convergent. For several reasons, not effective utilizing large number complementarily informative features. To end, present IC-LR, modification removes drop also our algorithms perform well experiments real-world sets, insights into their relative performance different settings.