[Paper Review] Rethinking Streaming Machine Learning Evaluation

Shreya Shankar, Bernease Herman+1 more|arXiv (Cornell University)|May 23, 2022

Data Stream Mining Techniques4 citations

TL;DR

This paper argues that streaming machine learning evaluation must move beyond accuracy alone, proposing a suite of metrics—including sliding window tracking, class representation monitoring, importance-weighted accuracy estimates, and loss percentiles—to detect performance degradation due to concept drift, label delays, and data distribution shifts. The key contribution is a framework for actionable, real-time model monitoring that incorporates recency bias and ground-truth feedback.

ABSTRACT

While most work on evaluating machine learning (ML) models focuses on computing accuracy on batches of data, tracking accuracy alone in a streaming setting (i.e., unbounded, timestamp-ordered datasets) fails to appropriately identify when models are performing unexpectedly. In this position paper, we discuss how the nature of streaming ML problems introduces new real-world challenges (e.g., delayed arrival of labels) and recommend additional metrics to assess streaming ML performance.

Motivation & Objective

• To identify the limitations of relying solely on accuracy for streaming machine learning evaluation.
• To address real-world challenges in streaming ML, including delayed labels, shifting class representations, and concept drift.
• To propose a set of actionable, recency-sensitive metrics that reflect organizational and task-specific priorities.
• To enable practitioners to detect unexpected model failures earlier by tracking subgroup performance and loss distributions.
• To advocate for a shift from batch-based to streaming-aware evaluation frameworks in production ML systems.

Proposed method

• Track accuracy over multiple sliding window sizes (e.g., 7-day, daily) to assess sensitivity to window choice.
• Monitor the fraction of positive examples in each sliding window to detect changes in class representation.
• Apply importance weighting (IW) to estimate accuracy per subgroup (e.g., pickup location) using training-set performance and live data distribution.
• Compute the difference between importance-weighted estimates and actual accuracy to detect concept shift.
• Track loss percentiles (e.g., 70th, 90th) over sliding windows to identify rising error in high-loss segments.
• Use divergence between IW estimates and real accuracy as an early signal for model retraining.

Experimental results

Research questions

• RQ1 How does window size selection affect the reliability of accuracy as a streaming ML evaluation metric?
• RQ2 To what extent do changes in class representation over time distort model performance evaluation?
• RQ3 How do delayed or incomplete labels impact the interpretability and reliability of accuracy metrics in streaming settings?
• RQ4 Can importance-weighted accuracy estimates detect concept drift earlier than standard accuracy tracking?
• RQ5 How can loss distribution metrics help identify high-risk data segments for retraining?

Key findings

• Accuracy computed over cumulative windows fails to detect performance drops, such as the significant decline in tip prediction accuracy during the early COVID-19 pandemic.
• A 7-day sliding window smooths accuracy fluctuations but still masks abrupt performance drops, especially when class imbalance increases.
• Label delays of up to 7 days (exponentially distributed) cause accuracy to appear artificially low or inconsistent, leading to misleading evaluations.
• Importance-weighted accuracy estimates based on pickup location show a large divergence from real accuracy around March 2020, signaling concept shift due to pandemic-related changes in tipping behavior.
• The 70th percentile loss increases significantly before the accuracy drop, indicating rising error in moderately high-loss predictions, which can guide targeted retraining.
• The combination of IW estimate differences and loss percentiles provides early, actionable signals for model retraining, improving detection of performance degradation.