Yakamoz ve Sunshine: İki yeni zaman serisi anomali tespit yöntemi

In modern society, availability and reliability of data have become crucial. Hence, one important task for a variety of fields is to detect abnormal instances in data. One of the objectives of identification of anomalies is to detect outlying data points, leading to new discoveries. In this case, anomalies themselves are of primary interest. The other objective is to detect anomalies in the preprocessing step in statistical analysis that may otherwise lead to model misspecification, biased parameter estimation and incorrect results. Obtaining forecasts using time series is a subject that has never lost its popularity, and it is becoming ever more important with the development of technology. The presence of even a few abnormal observations in the series, on the other hand, has a negative effect on these forecasts. In view of the problems, it is understood that there is a need for widely applicable and effective anomaly detection methods. Our main goal is to provide model-free, unsupervised, efficient and accurate detection of anomalous observations in time series data. For this reason, two novel time series anomaly detection methods have been proposed. These methods do not require anomalous samples in their training datasets. This skill has considerable practical value since the collection of labelled data can be difficult. In addition, the location of the anomaly does not affect the performance of the methods. Success of the proposed methods in finding anomaly/anomalies is investigated using real and simulated time series datasets. From the performance comparison results, it is seen that the proposed methods are very effective in finding anomaly/anomalies in these time series regardless of the variation in the data.