Publicação
Plotting Time: Exploring Visual Representations for Time Series Classification
| Resumo: | Time series data is a collection of data points acquired in successive order over a period of time, allowing us to obtain temporal information and make time-based predictions through the combination of Machine Learning (ML) algorithms. Time series are prevalent in crucial sectors for society’s development, such as Economy, Health, Weather, and Astronomy, with the objective of improving the quality of life through the prediction of climate changes, economic variations, earthquakes, and other types of events. These sectors require models with good predictive abilities and capable of scaling as the volume of data gradually increases. We can address this issue by using Deep Learning (DL) models that can keep a good performance while increasing the amount of data. One example is the Convolutional Neural Network (CNN), which uses images as input in several activity sectors. There is not much time series-related work with deep learning models and image generation. As a result, our objective is to develop new methods for image generation and then train them with a simple CNN. We focus on time series data to create a new algorithm for converting non-image time series data into graphical images that contain either box plots or violin plots with statistical information. We hypothesize that CNNs can interpret and learn different elements of the plots, and by comparing two different approaches, we can verify this statement. Our results indicate that CNNs may not understand some elements of the box and violin plots, for example, the outliers and quartiles, and focus more on the density and distribution of the data. In the future, it would be interesting to study alternative image generation algorithms and explore graphical representations in multivariate datasets. |
|---|---|
| Autores principais: | Marques, Brian Manuel Monteiro |
| Assunto: | Series Temporais Geração de Imagens Redes Neuronais Convolucionais Representações Gráficas Teses de mestrado - 2023 |
| Ano: | 2023 |
| País: | Portugal |
| Tipo de documento: | dissertação de mestrado |
| Tipo de acesso: | acesso aberto |
| Instituição associada: | Universidade de Lisboa |
| Idioma: | inglês |
| Origem: | Repositório da Universidade de Lisboa |
| Resumo: | Time series data is a collection of data points acquired in successive order over a period of time, allowing us to obtain temporal information and make time-based predictions through the combination of Machine Learning (ML) algorithms. Time series are prevalent in crucial sectors for society’s development, such as Economy, Health, Weather, and Astronomy, with the objective of improving the quality of life through the prediction of climate changes, economic variations, earthquakes, and other types of events. These sectors require models with good predictive abilities and capable of scaling as the volume of data gradually increases. We can address this issue by using Deep Learning (DL) models that can keep a good performance while increasing the amount of data. One example is the Convolutional Neural Network (CNN), which uses images as input in several activity sectors. There is not much time series-related work with deep learning models and image generation. As a result, our objective is to develop new methods for image generation and then train them with a simple CNN. We focus on time series data to create a new algorithm for converting non-image time series data into graphical images that contain either box plots or violin plots with statistical information. We hypothesize that CNNs can interpret and learn different elements of the plots, and by comparing two different approaches, we can verify this statement. Our results indicate that CNNs may not understand some elements of the box and violin plots, for example, the outliers and quartiles, and focus more on the density and distribution of the data. In the future, it would be interesting to study alternative image generation algorithms and explore graphical representations in multivariate datasets. |
|---|