Identifying Predictions That Influence the Future: Detecting Performative Concept Drift in Data Streams

TL;DR

The paper defines performative drift ($PD$) as distributional changes in data streams induced by deployed predictions and situates it within the broader concept-drift framework. It introduces CheckerBoard Performative Drift Detection (CB-PDD) and a model-agnostic synthetic data generator to evaluate PD detectors, then demonstrates CB-PDD’s effectiveness across synthetic and semi-synthetic datasets with favorable detection rates and low false positives, even in the presence of intrinsic drift. The results highlight CB-PDD’s ability to isolate PD from intrinsic drift and to compare favorably with traditional drift detectors under various conditions, while also revealing limitations related to class imbalance and parameter sensitivity. The work concludes with implications for practical PD detection and outlines future work on online variants and real-world PD data collection.

Abstract

Concept Drift has been extensively studied within the context of Stream Learning. However, it is often assumed that the deployed model's predictions play no role in the concept drift the system experiences. Closer inspection reveals that this is not always the case. Automated trading might be prone to self-fulfilling feedback loops. Likewise, malicious entities might adapt to evade detectors in the adversarial setting resulting in a self-negating feedback loop that requires the deployed models to constantly retrain. Such settings where a model may induce concept drift are called performative. In this work, we investigate this phenomenon. Our contributions are as follows: First, we define performative drift within a stream learning setting and distinguish it from other causes of drift. We introduce a novel type of drift detection task, aimed at identifying potential performative concept drift in data streams. We propose a first such performative drift detection approach, called CheckerBoard Performative Drift Detection (CB-PDD). We apply CB-PDD to both synthetic and semi-synthetic datasets that exhibit varying degrees of self-fulfilling feedback loops. Results are positive with CB-PDD showing high efficacy, low false detection rates, resilience to intrinsic drift, comparability to other drift detection techniques, and an ability to effectively detect performative drift in semi-synthetic datasets. Secondly, we highlight the role intrinsic (traditional) drift plays in obfuscating performative drift and discuss the implications of these findings as well as the limitations of CB-PDD.

Figures (17)

• Figure 1: Example of the Checkboard Pattern produced by the Checkerboard Detector for an arbitrary feature value in a Binary Classification task. In this Figure, the CheckerBoard Detector is parameterized with $f=0.5$ and $\tau = 2000$.
• Figure 2: shows the effect trial length ($\tau$) has on the detection rate of CB-PDD across various PD strengths ($\sigma$). The main finding is that as $\tau$ increases, so does the detection rate.
• Figure 3: shows the effect $f$ has on the detection rate of CB-PDD across various PD strengths ($\sigma$) in a low $\epsilon$ setting.
• Figure 4: shows the effect $f$ has on the detection rate of CB-PDD across various PD strengths ($\sigma$) in a high $\epsilon$ setting.
• Figure 5: shows the detection rate of CB-PDD when deployed in tandem with a predictive model. $m$ describes the portion of instances given to the predictive model. These results show that as $m$ increases, the detection rate decreases. The trend is more noticeable in the Threshold Classifier (TC) which induces its own PD in comparison to a Random Classifier (RC) which induces no PD.