Research
Models are trained once and then asked to survive a world that keeps moving. My work is about the gap between those two facts.
The problem
Almost everything in production machine learning assumes the future resembles the training set. It rarely does. Vocabulary shifts, demand patterns break, sensors drift, a pandemic rewrites a category overnight. The model does not fail loudly — it degrades quietly, and by the time offline metrics catch up the damage is done.
Three questions follow from that, and they are the three strands of my work.
Detecting that something changed
Concept drift detection is the first problem, and the hard version is the unsupervised one: knowing the distribution has moved before the labels arrive to prove it. In most real deployments labels are delayed by weeks, or never arrive at all.
LACE (2026, ACM CIKM) attacks this for multi-label streams by watching how clusters of co-occurring labels evolve, rather than watching a performance metric that needs ground truth. Against the previous best unsupervised method it reduces detection delay by 63.5% while keeping a 0% missed-detection rate.
Learning continuously, in one pass
Once you know the world moved, you have to move with it — without retraining from scratch every time.
AdaNEN (2024, ACM TKDD) is a neural ensemble architecture for evolving text streams that learns on the fly, in a single pass, adapting its own capacity as drift arrives. It improves classification accuracy by up to 8.8% across 13 benchmark datasets, and grew out of my M.Sc. thesis. BELS (2023, IEEE Access) explores the same territory with broad rather than deep learning, where fast incremental updates matter more than depth.
Keeping foundation models current
The newest strand, and where most of my attention is now: foundation models are enormously expensive to retrain, which makes drift a much sharper problem for them than for a small online classifier.
LLM-OFA (2025, ACM CIKM) proposes an On-the-Fly Adaptation framework and an optimizer called Adaptimizer for continually adapting large language models under temporal drift, evaluated across two decades of news.
The same question is now open for time-series foundation models, where the supervision is delayed — a forecast’s true value only arrives much later, so the model has to decide how to adapt before it knows whether it was right. That is where most of my current work sits. More on it once it is through review.