Analysis of the impact of Dissimilarity Space within the Concept Drift Problem
This post shares insights and download for my undergraduate thesis (TCC). My work focused on concept drift, where we explored the use of dissimilarity-based classifiers to tackle challenges in dynamically changing data environments. We implemented various approaches for selecting the reference set (R), which is crucial for constructing effective dissimilarity spaces.
Abstract
Concept Drift is a known problem mostly in the context of supervised models in which it affects negatively many models capabilities due to the change of the original concept and thus generating wrong target values. Due to time or other phenomena, i.e., changes over time during many possible phases in such a way that it deviates from generating the original intended target because of those changes for the model to capture. Researchers are continuously trying to find solutions that may be able to identify concept drift and also to quickly adapt predictive models so that its impact is reduced or entirely mitigated. In this paper, our goal is to apply classification techniques based on dissimilarity space representation based on various literature in hopes that it may solve most of the divergence caused by the problem and thus making a robust model in which can handle the problem at hand and thus diminishing it’s impact in such that the model targets maintain the correct classification even due to changes overtime. The results indicate that the transformation of data into the dissimilarity space did not yield significant benefits. One of the classifiers utilizing Instance Hardness metrics demonstrated performance comparable to that of traditional classifiers.
Index Terms — Machine Learning, Dissimilarity spaces, Concept drift, Data streams, Data Science
Dedicatory
This work is dedicated to Gabriel Antonio Gomes de Farias, a brilliant friend and collaborator who played a fundamental role in this project. His ideas, dedication, and support were invaluable throughout the journey. Gabriel’s passion for knowledge and kindness will always be remembered, and this work stands as a testament to his impact on both my life and this project. Thank you, Gabriel—you are deeply missed.