Statistiline triivi avastamise meetod
Files
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Masinõppemudelid eeldavad, et andmed pärinevad statsionaarsest jaotusest.Praktikas on tihti vaja mudelitega tõlgendada andmeid, mis pärinevad kiiresti dünaamiliselt muutuvast andmevoost. Seda muutust õppe- ja testvalimis nimetatakse kontseptuaalseks triiviks (ingl k concept drift). Kontseptuaalse triivi olemasolu võib kahjustada mudelennustuste täpsust ja usaldusväärsust. Seetõttu on kontseptuaalse triivi arvestamine väga oluline, et vähendada selle negatiivset mõju tulemustele. Kontseptuaalse triivi arvestamiseks tuleb see kõigepealt tuvastada. Selle tuvastamiseks kasutatakse triivi detektoreid. Reaktiivsed kontseptuaalse triivi detektorid püüavad tuvastada triivi niipe kui see ilmneb, jälgides aluseks oleva masinõppe mudeli toimimist. Tõlgendatavus on masinõppes tähtis ja meetod võib osutuda kasulikuks mitte ainult triivi olemasolu tuvastamiseks andmekogumis, vaid ka triivi põhjuste tuvastamisel ja analüüsimisel.Käesolevas töös rõhutatakse tõlgendatavuse tähtsust triivi tuvastamisel ja esitatakse statistilise triivi tuvastamise meetod (SDDM), mis tuvastab triivi kiiresti arenevates andmevoogudes, kusjuures võrdluses kaasaegsete meetoditega esineb vähem valepositiivseid ja valenegatiivsed tulemusi. Meetod annab ka kontseptuaalse triivi põhjuste tõlgenduse. Töös näidatakse meetodi tõhusust, rakendades seda nii sünteetilistele kui ka reaalsetele andmekogumitele.
Machine learning models assume that data is drawn from a stationary distribution. However, in practice, challenges are imposed on models that need to make sense of fast-evolving data streams, where the content of data is changing and evolving dynamically over time. This change between the underlying distributions of the training and test datasets is called concept drift. The presence of concept drift may compromise the accuracy and reliability of prospective computational predictions. Therefore, handling concept drift is of great importance in the direction of diminishing its negative effects on a model's performance. In order to handle concept drift, one has to detect it first. Concept drift detectors have been used to accomplish this - reactive concept drift detectors try to detect drift as soon as it occurs by monitoring the performance of the underlying machine learning model. However, the importance of interpretability in machine learning indicates that it may prove useful to not only detect that drift is occurring in the data, but to also identify and analyze the causes of the drift. In this thesis, the importance of interpretability in drift detection is highlighted and the Statistical Drift Detection Method (SDDM) is presented, which detects drifts in fast-evolving data streams with a smaller number of false positives and false negatives when compared to the state-of-the-art, and has the ability to interpret the cause of the concept drift. The effectiveness of the method is demonstrated by applying it on both synthetic and real-world datasets.
Machine learning models assume that data is drawn from a stationary distribution. However, in practice, challenges are imposed on models that need to make sense of fast-evolving data streams, where the content of data is changing and evolving dynamically over time. This change between the underlying distributions of the training and test datasets is called concept drift. The presence of concept drift may compromise the accuracy and reliability of prospective computational predictions. Therefore, handling concept drift is of great importance in the direction of diminishing its negative effects on a model's performance. In order to handle concept drift, one has to detect it first. Concept drift detectors have been used to accomplish this - reactive concept drift detectors try to detect drift as soon as it occurs by monitoring the performance of the underlying machine learning model. However, the importance of interpretability in machine learning indicates that it may prove useful to not only detect that drift is occurring in the data, but to also identify and analyze the causes of the drift. In this thesis, the importance of interpretability in drift detection is highlighted and the Statistical Drift Detection Method (SDDM) is presented, which detects drifts in fast-evolving data streams with a smaller number of false positives and false negatives when compared to the state-of-the-art, and has the ability to interpret the cause of the concept drift. The effectiveness of the method is demonstrated by applying it on both synthetic and real-world datasets.