Robustness-enhanced Uplift Modeling with Adversarial Feature Desensitization

TL;DR

The paper tackles robustness gaps in uplift modeling by identifying a feature sensitivity problem where perturbing key features can invert uplift predictions. It introduces RUAD, a model-agnostic framework that combines a feature selection module with a joint multi-label objective and an adversarial feature desensitization module to reduce sensitivity to these features during training. Through extensive experiments on IHDP and Production, RUAD yields superior uplift ranking metrics and demonstrates robustness to feature perturbations, while remaining compatible with diverse base uplift models. These contributions advance practical uplift modeling by enhancing stability and performance in real-world online marketing settings.

Abstract

Uplift modeling has shown very promising results in online marketing. However, most existing works are prone to the robustness challenge in some practical applications. In this paper, we first present a possible explanation for the above phenomenon. We verify that there is a feature sensitivity problem in online marketing using different real-world datasets, where the perturbation of some key features will seriously affect the performance of the uplift model and even cause the opposite trend. To solve the above problem, we propose a novel robustness-enhanced uplift modeling framework with adversarial feature desensitization (RUAD). Specifically, our RUAD can more effectively alleviate the feature sensitivity of the uplift model through two customized modules, including a feature selection module with joint multi-label modeling to identify a key subset from the input features and an adversarial feature desensitization module using adversarial training and soft interpolation operations to enhance the robustness of the model against this selected subset of features. Finally, we conduct extensive experiments on a public dataset and a real product dataset to verify the effectiveness of our RUAD in online marketing. In addition, we also demonstrate the robustness of our RUAD to the feature sensitivity, as well as the compatibility with different uplift models.

Figures (4)

• Figure 1: Bar graphs of predicted uplift with 5 bins, w.r.t the origin dataset (i.e., (a)) and three kinds of varieties (i.e., (b)-(d)). For each dataset, we randomly select 30% of all continuous-valued features and apply a Gaussian noise with $\eta \sim \mathcal{N}(0,0.05^2)$ as perturbation while constraining $\|{\eta}\|_{\infty}<0.1$. Note that a good uplift model will usually have a bar graph sorted in descending order.
• Figure 2: The architecture of our RUAD. The propensity network $\pi(x)$ is pre-trained to calculate the transformed response $y^*$. The left is the feature selection module (FS), which leverages a masker to select key sensitive features for jointly modeling transformed response $y^*$ and user response $y$. The right is an adversarial feature desensitization module (AFD) to reduce the sensitivity of the base uplift model to these key features. Specially, $\mathcal{L}_o$ and $\mathcal{L}_r$ are used for FS, while $\mathcal{L}_a$ is used for AFD. The detailed form of the loss function is presented in Eq.\ref{['equ:1']}.
• Figure 3: Bar graphs of predicted uplift with 5 bins, w.r.t the origin dataset (i.e., (a)) and three kinds of varieties (i.e., (b)-(d)). We present the results of our RUAD with S-NN as the base uplift model.
• Figure 4: Performance of our RUAD with three typical base uplift models on the Production dataset, i.e. S-NN, T-NN and Dragonnet. We evaluate the results by using the Qini coefficient and Kendall uplift rank correlation with 5 bins.