Concept Drift Detection Through Resampling

Authors: Maayan Harel, Shie Mannor, Ran El-Yaniv, Koby Crammer

ICML 2014 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experimental results show that the method has high recall and precision, and performs well in the presence of noise.
Researcher Affiliation Academia Maayan Harel MAAYANGA@TX.TECHNION.AC.IL Koby Crammer KOBY@EE.TECHNION.AC.IL Ran El-Yaniv RANI@CS.TECHNION.AC.IL Shie Mannor SHIE@EE.TECHNION.AC.IL Technion Israel Institute of Technology, Haifa, Israel.
Pseudocode Yes Algorithm 1: Concept Drift Detection Scheme; Algorithm 2: Slow Gradual Drift Detection Scheme.; Algorithm 3: TEST( ˆRord, n ˆRS i(ASi) o P i=1 , , δ) Procedure for detection scheme
Open Source Code No The paper mentions using 'scikit-learn: Machine Learning in Python toolbox' but does not provide a link or explicit statement about releasing their own implementation code.
Open Datasets Yes We compare detection performance on a user preference prediction task defined using the 20-news groups text dataset6, consisting of 18, 846 documents and over 75, 000 features. 6http://qwone.com/ jason/20Newsgroups
Dataset Splits Yes In the first step of our basic detection scheme the observed sub-sequence Zn is divided into a training window Sord = Zk 1 , 1 < k < n, and a test window S ord = Zn k+1.
Hardware Specification No The paper does not provide any specific hardware details such as GPU models, CPU types, or memory amounts used for running the experiments.
Software Dependencies No We used scikit-learn: Machine Learning in Python toolbox. Cross-validation on a single random concept showed low sensitivity to the choice of C on this dataset and therefore the default value C = 1 was chosen. The paper mentions 'scikit-learn' but does not specify its version number, nor does it specify version numbers for other key software components.
Experiment Setup Yes We set the sensitivity level of PERM and grad-PERM to δ = 0.01, = 0, the warning and detection thresholds of STEPD to w = 0.05, d = 0.01, and the parameters of EDDM to α = 0.95 and β = 0.90. The base algorithm was K-Nearest Neighbors (k = 3), each stream was randomly repeated 100 times, and P = 100 reshuffling splits were used in PERM. We use P = 500, and SVM and SVR with linear kernel as the learning algorithms. Cross-validation on a single random concept showed low sensitivity to the choice of C on this dataset and therefore the default value C = 1 was chosen.