Machine Results Interpreting and Data Understanding

Machine learning solves a large number of tasks that a human could not cope with. It has huge potential and is applicable in many areas of human activity. However, often scientists in the field of machine learning, getting the results of the models, can not get an accurate forecast, explain, or interpret the data obtained. The stage of analysis and preparation of data is one of the main tasks. In machine learning, data play a critical role, machine learning model developers often ignore them or treat them incorrectly. As a result, hundreds of hours are spent on adjusting the model based on incorrect data, which may cause low accuracy of the model, and this will not be related with model optimization and adjusting. When analyzing datasets, knowing what needs to be discovered in the data is often just as important as learning its overall structure. The area of data analysis concerned with detecting rare occurrences is called anomaly detection. Anomaly detection has many applications in security, healthcare, finance, and many others. Because of the complexity of obtaining markedup data, machine learning algorithms with supervised learning are less attractive for the task of anomaly detection. And the lack of reliable and marked-up data makes it difficult to evaluate anomaly detection methods. The article provides a specific example of identifying errors that affect the quality of machine learning models. This study shows the identified errors in the SWaT 2015.