Use this link to cite:
https://hdl.handle.net/2183/49672 Estudio de la aplicabilidad del algoritmo DoME en escenarios de Stream Learning
Loading...
Identifiers
Publication date
Authors
Cabaleiro Pintos, Laura
Other responsabilities
Universidade da Coruña. Facultade de Informática
Journal Title
Bibliographic citation
Type of academic work
Academic degree
Abstract
[Resumen] Dentro del ámbito del Aprendizaje Automático, la Regresión Simbólica destaca por su capacidad para obtener modelos matemáticos interpretables a partir de los datos, lo que la convierte en una alternativa de interés frente a otros enfoques. En este contexto, Development of Mathematical Expressions (DoME) se presenta como un método de Regresión Simbólica basado en la construcción iterativa de expresiones, capaz de generar modelos intrínsecamente interpretables en forma de ecuaciones matemáticas. DoME destaca como método de Aprendizaje Automático por su carácter determinista, el control sobre la complejidad de las expresiones generadas y la posibilidad de obtener modelos compactos, analizables y reutilizables sin depender de librerías específicas de aprendizaje automático. Sin embargo, la evaluación de DoME se ha centrado principalmente en escenarios de aprendizaje estático, por lo que su comportamiento en contextos secuenciales y de actualización continua sigue siendo una cuestión abierta. En este marco, el presente Trabajo Fin de Grado estudia la aplicabilidad de DoME en escenarios de stream learning. Para ello, se plantea una evaluación experimental organizada de forma incremental: en primer lugar, se analiza el comportamiento del método en un escenario batch para establecer una línea base; a continuación, se adapta DoME a un entorno secuencial mediante una estrategia de actualización basada en memoria deslizante; y, finalmente, se incorpora un mecanismo explícito de detección de deriva con el objetivo de estudiar su influencia en la adaptación del modelo. La experimentación se realiza sobre conjuntos de datos de distinta naturaleza, incluyendo problemas de regresión con estructura temporal, datos sintéticos con deriva y un problema de clasificación binaria. El objetivo del estudio es valorar si DoME puede generar modelos útiles e interpretables en este tipo de contextos, así como analizar cómo cambia su comportamientoal pasar de un entorno estático a otro de aprendizaje secuencial. Los resultados muestran un comportamiento diferenciado según el conjunto de datos y el escenario de evaluación. En particular, la adaptación a stream learning resulta especialmente favorable en Friedman-Drift, mientras que en Bike Sharing la línea base de aprendizaje batch mantiene un rendimiento más competitivo. La incorporación de Adaptive Windowing 2 (ADWIN2) también presenta un efecto heterogéneo, con una respuesta especialmente reactiva en UNSW-NB15.
[Abstract] Within the field of Machine Learning, Symbolic Regression stands out for its ability to obtain interpretable mathematical models from data, making it an interesting alternative to other approaches. In this context, DoME is presented as a Symbolic Regression method based on the iterative construction of expressions, capable of generating intrinsically interpretable models in the form of mathematical equations. DoME stands out as a Machine Learning method due to its deterministic nature, its control over the complexity of the generated expressions, and its ability to produce compact, analysable and reusable models without depending on specific Machine Learning libraries. However, the evaluation of DoME has mainly focused on static learning scenarios, so its behaviour in sequential contexts and continuous updating remains an open question. In this framework, this Bachelor’s Thesis studies the applicability of DoME in stream learning scenarios. To this end, an experimental evaluation is proposed and organised incrementally: first, the behaviour of the method is analysed in a batch scenario in order to establish a baseline; then, DoME is adapted to a sequential environment through an update strategy based on a sliding memory; and, finally, an explicit drift detection mechanism is incorporated with the aim of studying its influence on the adaptation of the model. The experimentation is carried out on datasets of different nature, including regressionproblems with temporal structure, synthetic data with drift, and a binary classification problem. The aim of the study is to assess whether DoME can generate useful and interpretable models in this type of context, as well as to analyse how its behaviour changes when moving from a static environment to a sequential learning one. The results show that the stream adaptation changes the behaviour of DoME with respect to the batch scenario. In particular, clear improvements are observed in Friedman-Drift, whereas in Bike Sharing the sequential adaptation does not preserve the performance of the static baseline. The incorporation of ADWIN2 also shows heterogeneous behaviour across datasets, with a particularly reactive response in UNSW-NB15.
[Abstract] Within the field of Machine Learning, Symbolic Regression stands out for its ability to obtain interpretable mathematical models from data, making it an interesting alternative to other approaches. In this context, DoME is presented as a Symbolic Regression method based on the iterative construction of expressions, capable of generating intrinsically interpretable models in the form of mathematical equations. DoME stands out as a Machine Learning method due to its deterministic nature, its control over the complexity of the generated expressions, and its ability to produce compact, analysable and reusable models without depending on specific Machine Learning libraries. However, the evaluation of DoME has mainly focused on static learning scenarios, so its behaviour in sequential contexts and continuous updating remains an open question. In this framework, this Bachelor’s Thesis studies the applicability of DoME in stream learning scenarios. To this end, an experimental evaluation is proposed and organised incrementally: first, the behaviour of the method is analysed in a batch scenario in order to establish a baseline; then, DoME is adapted to a sequential environment through an update strategy based on a sliding memory; and, finally, an explicit drift detection mechanism is incorporated with the aim of studying its influence on the adaptation of the model. The experimentation is carried out on datasets of different nature, including regressionproblems with temporal structure, synthetic data with drift, and a binary classification problem. The aim of the study is to assess whether DoME can generate useful and interpretable models in this type of context, as well as to analyse how its behaviour changes when moving from a static environment to a sequential learning one. The results show that the stream adaptation changes the behaviour of DoME with respect to the batch scenario. In particular, clear improvements are observed in Friedman-Drift, whereas in Bike Sharing the sequential adaptation does not preserve the performance of the static baseline. The incorporation of ADWIN2 also shows heterogeneous behaviour across datasets, with a particularly reactive response in UNSW-NB15.
Description
Editor version
Rights
Attribution-NonCommercial 4.0 International








