Bilkent University · BilIR

Machine learning for data that never stops changing.

I am a PhD researcher at Bilkent University, working with Prof. Fazlı Can in the Bilkent Information Retrieval Group. My research is about models that keep learning after deployment — detecting concept drift, adapting on the fly, and holding accuracy as the underlying distribution moves.

Alongside the PhD I am a data scientist at Invent.ai, where the same problem shows up with money attached: demand forecasts for roughly a million SKU-store pairs, retrained and re-evaluated against a world that shifts every week.

Publications Research CV

Pouya Ghahramanian, at the LLM-OFA poster session — At the LLM-OFA poster

01 — Research

Three strands, one question

Concept drift

Detecting when a data stream's underlying distribution has moved — including when no labels are available to tell you.

LACE

Online & continual learning

Single-pass architectures that update as data arrives, instead of retraining from scratch on a schedule.

AdaNEN · BELS

Adapting foundation models

Keeping large language and time-series foundation models current under temporal drift and delayed supervision.

LLM-OFA

02 — Selected publications

Peer-reviewed work

All publications

2026LACE

LACE: Unsupervised Concept Drift Detection in Multi-Label Data Streams Through Label Cluster Evolution

Gofralilar, M. K., Ghahramanian, P., & Can, F.

ACM CIKM
An unsupervised concept-drift detector for multi-label streams that cuts detection delay by 63.5% against the previous best unsupervised method, at a 0% missed-detection rate.
DETECTION DELAY −63.5% · 0% MISSED · MULTI-LABEL STREAMS
2025LLM-OFA

LLM-OFA: On-the-Fly Adaptation of Large Language Models to Address Temporal Drift Across Two Decades of News

Ghahramanian, P., Bakhshi, S., & Can, F.

ACM CIKM
An On-the-Fly Adaptation framework and the Adaptimizer optimizer for continually adapting LLMs under temporal drift, evaluated across two decades of news.
2 DECADES OF NEWS · CONTINUAL LLM ADAPTATION
Paper Code
2024AdaNEN

A Novel Neural Ensemble Architecture for On-the-Fly Classification of Evolving Text Streams

Ghahramanian, P., Bakhshi, S., Bonab, H., & Can, F.

ACM TKDD
AdaNEN — a neural ensemble for evolving data streams that improves classification accuracy by up to 8.8% across 13 benchmark datasets.
ACCURACY +8.8% · 13 DATASETS · SINGLE-PASS
Paper Code Thesis

03 — Results

What the work moved

8.8%

accuracy gain

AdaNEN, across 13 benchmark datasets

63.5%

lower detection delay

LACE, at a 0% missed-detection rate

2 decades

of news, continually adapted

LLM-OFA, ACM CIKM 2025

04 — Currently

Two desks

Full experience

Currently

I split my time between two versions of the same question.

At Bilkent I am a named senior researcher on three TÜBİTAK-funded R&D programs (117E870, 120E103, 125E060), leading the machine-learning workstream on classification, anomaly detection and concept-drift adaptation over large-scale streaming data. I have been with the group since 2019, first for my M.Sc. and now the PhD.

At Invent.ai I own demand forecasting and replenishment models in production for three enterprise retail clients. Last year that work cut lost sales by 10% for one client — the same drift problem, but where being wrong shows up on a shelf.

I also review for ACM SIGIR, CIKM and SIGIR-AP, and teach Information Retrieval, Algorithms and Computer Organization at Bilkent.