Performative Drift Resistant Classification Using Generative Domain Adversarial Networks

TL;DR

This work defines performative drift, a model-induced distribution shift, and proposes GDAN, a hybrid of Domain Adversarial Neural Networks and Generative Adversarial Networks, to learn domain-invariant representations and reverse drift without frequent retraining. GDAN optimizes a two-player game that aligns representations and reconstructs the pre-drift distribution through a generator, while a label classifier preserves predictive accuracy. Empirical results on two drift simulators show that GDAN can limit performance degradation over time and, in some cases, outperform retraining under constrained labeling, though its benefits diminish when drift direction changes rapidly. The study provides practical insights into drift understanding as a robust alternative to retraining in performative settings and outlines directions for real-world validation and drift-detection integration.

Abstract

Performative Drift is a special type of Concept Drift that occurs when a model's predictions influence the future instances the model will encounter. In these settings, retraining is not always feasible. In this work, we instead focus on drift understanding as a method for creating drift-resistant classifiers. To achieve this, we introduce the Generative Domain Adversarial Network (GDAN) which combines both Domain and Generative Adversarial Networks. Using GDAN, domain-invariant representations of incoming data are created and a generative network is used to reverse the effects of performative drift. Using semi-real and synthetic data generators, we empirically evaluate GDAN's ability to provide drift-resistant classification. Initial results are promising with GDAN limiting performance degradation over several timesteps. Additionally, GDAN's generative network can be used in tandem with other models to limit their performance degradation in the presence of performative drift. Lastly, we highlight the relationship between model retraining and the unpredictability of performative drift, providing deeper insights into the challenges faced when using traditional Concept Drift mitigation strategies in the performative setting.

Figures (4)

• Figure 1: GDAN architecture, composed of a domain adversarial network and a generative adversarial network. Single-sided arrows illustrate an input/output relationship, while double-sided arrows describe concatenation of data.
• Figure 2: Visualization of the flow of the simulation. For training of the GDAN architecture, the only necessary distributions are annotated baseline $t=0$ and the first distribution when drift is detected $t=1$. Iterations $1 \ldots n$ are used for evaluation purposes. After a model is trained, $t=1$ is regenerated and the testing starts on it. In each iteration, four models are evaluated. The diagram depicts a division, which is similar to a test/train split in a classic ML setup.
• Figure 3: Kernel Density Estimation of the Perdomo data generator's performative features. Using GDAN, the blue (drifted) data points are projected back to the red (original) data distribution. These projections are indicated by the green areas which show that GDAN is capable of reversing the effects of the drift, despite not being able to recreate the shape of the original distribution exactly.
• Figure 4: PCA visualization of the Izzo data generator on iterations where the logistic regressor was retrained. The figure illustrates the cyclical nature of the drift, as the blue (drifted) clusters oscillate, moving to and from the stationary red (original data distribution) clusters. This dynamic significantly affects GDAN's ability to project the drifted data points as shown by the green clusters.