Otomatik makine öğrenimi yöntemlerini kullanarak anomali tespit çerçevesi oluşturma

In terms of both research academic and general purposes, the importance of machine learning techniques is undeniable. However, these methods rely heavily on manual intervention for data refinement and parameter optimization, posing a challenge for researchers of varying levels of expertise. This reliance creates significant barriers for users with varying expertise levels, as existing frameworks struggle to manage diverse data types—structured, unstructured, and semi-structured—thereby compromising effective anomaly detection and limiting comprehensive data analysis. This thesis introduces an Automated Machine Learning (AutoML) framework designed for robust outlier detection across multiple data structures. Unlike traditional AutoML systems that necessitate some pre-classification of data or manual algorithm selection, this framework autonomously classifies datasets, discerns their characteristics, and directs them through a tailored pipeline to apply suitable outlier detection algorithms. This end-to-end automation is rare in the realm of outlier detection and represents a significant advancement. The framework's core functionality includes an automatic classification of datasets into categories like spatial, time-series, and dimensional statistical, based on intrinsic characteristics. This classification informs tailored preprocessing techniques that enhance data quality for effective outlier detection. For example, spatial datasets undergo geospatial transformations, while time-series data are adjusted for seasonality. The framework was tested across 100 diverse datasets from fields including finance, healthcare, and social media. It demonstrated a reduction in processing time and anincrease in precision, with a maintained recall rate exceeding 85 percent across majority of datasets. Additionally, the framework dynamically selects outlier detection algorithms based on the data type, enhanced by a meta-learning component that utilizes historical data to optimize algorithm selection. An early stopping mechanism conserves resources by halting processing once performance thresholds are met, ensuring efficiency and scalability. Overall, this paper contributes a novel and versatile framework that streamlines the outlier detection process, offering a systematic approach similar to an AutoML system. It paves the way for enhanced efficiency and effectiveness in outlier detection tasks across different domains and datasets. Furthermore, the integration of outlier detection mechanisms within this enhanced data environment is anticipated to yield faster identification of irregularities without compromising accuracy or reliability. The framework aims to accelerate the detection process and increasing the efficiency of outlier identification tasks. The comprehensive evaluation across diverse datasets seeks to showcase the framework's adaptability and consistent performance, validating its ability to automate outlier detection across varying data types and scenarios. The successful demonstration of this framework would mark a substantial step in simplifying and refining outlier detection processes in data analysis across multiple domains and datasets.