Hüseyin Kaya

The goal is to create a reinforcement learning-based auto-scaling software suitable for cloud applications.

The end-product is designed as a Kubernetes operator to be listed in Kubernetes market places.

A bioinformatics project to detect mutations in DNA chromatograms, which are 4-channel—A, C, G, and T—time series.

At the core of the project, there is a time series classification algorithm running that takes a query time series and returns the locations of the possible mutations by first aligning the query with the reference time series in the database.

The MATLAB implementation is wrapped and used as a back end for PHP and later provided as a free service in the company's customer portal.

An in-house project I prepared and implemented for Iontek Inc. The project aimed to improve the primer, a short nucleotide sequence, design process. Iontek constantly needs to develop its real-time PCR diagnostic tests.

The end product takes multiple DNA sequences—all strings of A, C, G, and T—and returns a primer that matches only a subset of sequences. The implementation is done on the NVIDIA CUDA platform to gain more performance on speed.

I integrated the core CUDA implementation with a Java Swing GUI. The product is used in-house by product developers.

A MATLAB implementation of various time series alignment methods, including Dynamic Time Warping (DTW), constrained DTW, and my own implementation called Signal Alignment via Genetic Algorithm (SAGA) published in "H Kaya, S Gunduz-Oguducu, Information Sciences 228, 113-130".

The utility combines the alignment methods with classification methods such as K-NN, linear and nonlinear SVM. I primarily used this utility in Iontek Inc. to classify time series originating from DNA sequencing.

A joint research project to investigate the use of time series forecasting methods in predicting the number of transactions passing through the payment gateway managed by Payten Turkey. We tested various methods, including the traditional seasonal naive approach, auto-regressive moving average (ARIMA), exponential smoothing (ETS), and modern ones such as Facebook Prophet and CNNQR provided in the Amazon Forecast service. The time-series data is curated from our payment gateway transactions covering six years.

A computer vision research to develop a feature selection algorithm to be used to classify hyperspectral remote sensing images. The features correspond to the spectral bands in the images.

The selection procedure is based on measuring the effect of each band in classification performance. And this is achieved by using a global sensitivity analysis method called High Dimensional Model Representation.

A reactive user interface and a computing back end for earthquake, flood, and landslide damage assessment simulation prepared for "Tomorrow's Earthquake-resilient Cities," an ongoing exhibition in Istanbul. The interface is prepared for a touch-screen 4K TV. The UI is based on solara.dev, a React-like Python framework.

A computer vision project to classify post-earthquake building images based on ConvNeXt models. I designed and conducted all experiments in the manuscript, prepared the visualizations, and built a user interface for inference. The manuscript is submitted to a journal.