Comparing deep learning models in terms of prediction and explainability in time series forecasting
In the realm of time-series forecasting, deriving meaningful predictions from historical trends and patterns presents a significant challenge. Traditional methods, while showcasing certain capabilities, often encounter limitations when faced with the non-linear and dynamic nature of many time-series...
| Autor: | |
|---|---|
| Tipo de recurso: | tesis de maestría |
| Fecha de publicación: | 2023 |
| País: | España |
| Institución: | Universitat Politècnica de Catalunya (UPC) |
| Repositorio: | UPCommons. Portal del coneixement obert de la UPC |
| Idioma: | inglés |
| OAI Identifier: | oai:upcommons.upc.edu:2117/401828 |
| Acceso en línea: | https://hdl.handle.net/2117/401828 |
| Access Level: | acceso abierto |
| Palabra clave: | Deep learning Time-series analysis explainability time-series forecasting Autoregressive Integrated Moving Averag deep learning ARIMA Aprenentatge profund Sèries temporals -- Anàlisi Àrees temàtiques de la UPC::Informàtica::Intel·ligència artificial |
| Sumario: | In the realm of time-series forecasting, deriving meaningful predictions from historical trends and patterns presents a significant challenge. Traditional methods, while showcasing certain capabilities, often encounter limitations when faced with the non-linear and dynamic nature of many time-series datasets. Deep learning, with its ability to decipher intricate data patterns, has emerged as a promising contender in this domain. Given the expanding landscape of deep learning architectures, it becomes essential to discern their comparative strengths and weaknesses. Thus, the primary objective of this study is to understand the progression of deep learning methodologies over time and to validate whether subsequent models genuinely enhance their predecessors' performance in terms of prediction, complexity, and explainability. This research offers a comprehensive comparison of three prominent deep learning models: Multi-Layer Perceptrons (MLP), Recurrent Neural Networks (RNN), and Long Short-Term Memory networks (LSTM). The experiments encompassed different window sizes: 7, 30, 90, 180, and 360 days, to ascertain the robustness and adaptability of each model across varying temporal contexts. The results pointed to a consistent trend of improvement in prediction from MLP to RNN, and subsequently to LSTM, as demonstrated by the reduction in error metrics. LSTM outperformed the other tested methods. In addition to predictive performance, the study uses the SHapley Additive exPlanations (SHAP) method to delve into the explainability of these models. This provided insights into feature importance across different model architectures. The differences on how each of the tested DL architectures handles sequential data are reflected in the used explainability method and this gives better understanding about the architectures used and the data used for the modeling. In conclusion, this thesis not only acknowledges the advancements in deep learning models for time-series forecasting but also emphasizes the importance of comprehending their foundational mechanisms. This synthesis of performance and explainability delivers a comprehensive perspective, positioning practitioners to make informed decisions in real-world forecasting scenarios. |
|---|