The study explores the tradeoff between accuracy and energy efficiency in seven common concept drift detection methods: ADWIN, DDM, EDDM, HDDM-A, HDDM-W, KSWIN, and PageHinkley. It uses five synthetic datasets with both abrupt and gradual drift, and six different base classifiers to train the ML models.
The key findings are:
The drift detectors can be categorized into three types:
a) Detectors that sacrifice energy efficiency for high detection accuracy (KSWIN)
b) Balanced detectors that consume low to medium energy with good accuracy (HDDM-W, ADWIN)
c) Detectors that consume very little energy but are unusable due to very poor accuracy (HDDM-A, PageHinkley, DDM, EDDM)
The type of drift (abrupt vs. gradual) influences the energy consumption, with gradual drift tending to consume more energy than abrupt drift for KSWIN, EDDM, and ADWIN. The type of drift also affects the detection accuracy, with gradual drift generally performing better.
The choice of base classifier (ML model) does not significantly impact the energy consumption or accuracy of the drift detectors.
The findings provide valuable insights to ML practitioners on selecting the most suitable drift detection method for their ML-enabled systems, balancing the tradeoffs between accuracy and energy efficiency.
To Another Language
from source content
arxiv.org
Key Insights Distilled From
by Rafiullah Om... at arxiv.org 05-01-2024
https://arxiv.org/pdf/2404.19452.pdfDeeper Inquiries