Bilkent University
Department of Computer Engineering

---

M.S.THESIS PRESENTATION

 

LACE: UNSUPERVISED CONCEPT DRIFT DETECTION IN MULTI-LABEL DATA STREAMS THROUGH LABEL CLUSTER EVOLUTION

 

Mehmet Kadri Gofralılar
Master Student
(Supervisor:Prof.Dr.Fazlı Can)

Computer Engineering Department
Bilkent University

Abstract: Concept drift detection in multi-label data streams is challenging when true label vectors are delayed or unavailable. Most existing drift detectors monitor prediction errors and therefore assume immediate access to ground-truth labels, which is often unrealistic in stream settings. To address this problem, we propose LACE (Label Cluster Evolution), an unsupervised concept drift detector that uses changes in predicted label-dependency structure as evidence of drift. LACE maintains a sliding window of recent predictions, clusters predicted labels in consecutive subwindows, and compares the resulting label partitions to detect structural changes over time. The drift decision uses Adjusted Rand Index for regular partition comparisons, Normalized Variation of Information for degenerate cases, and a Monte Carlo permutation procedure for ambiguous cases. We evaluate LACE against several supervised error-based drift detectors and LD3, the only prior unsupervised drift detector for multi-label data streams to our knowledge. On real datasets, LACE achieves the best overall predictive rank, improving over LD3, the second-best method, by 4.18% and over the average baseline rank by 23.40%. On synthetic datasets with known drift points, LACE achieves 0% Missed Detection Rate and reduces Delay For Detection by 84.45% relative to LD3, the only other method with 0% Missed Detection Rate. This thesis also presents hyperparameter selection, ablation studies, and an efficiency evaluation based on runtime.

 

DATE: July 22, Wednesday @ 13:30

Place: EA 516