Différences

Ci-dessous, les différences entre deux révisions de la page.

--- documentation:data:models [2024/04/09 12:22] – créée baudryj
+++ documentation:data:models [2024/04/09 12:30] (Version actuelle) – baudryj
@@ Ligne 2: / Ligne 2: @@
-<mermaid>
+Great summary of various time series analysis methods and models! Here's a brief overview of each method mentioned:
-graph TD
-    A[Time Series Models] --> B[Statistical Models]
-    A --> C[Machine Learning]
-    A --> D[Deep Learning]
-    A --> E[Other Approaches]
-    B --> ARIMA[ARIMA\n(Trend, Non-seasonal)]
-    B --> SARIMA[SARIMA\n(Trend, Seasonal)]
-    B --> HW[Exponential Smoothing\n(Trend, Seasonal)]
-    C --> Prophet[Prophet\n(Flexible, Daily observations)]
-    C --> GP[Gaussian Processes\n(Flexible regression)]
-    C --> HMM[Hidden Markov Models\n(Sequence of states)]
-    C --> DTW[Dynamic Time Warping\n(Sequence similarity)]
-    D --> LSTM[LSTM Networks\n(Long-term dependencies)]
-    D --> CNN[Convolutional Neural Networks\n(Patterns in segments)]
-    D --> Transformer[Transformer Models\n(Long-range dependencies)]
-    E --> Note1[Note: Gaussian Processes, Hidden Markov Models, and Dynamic Time Warping are classified under 'Other Approaches' due to their unique methodologies not fitting neatly into the conventional ML or DL categories. They are versatile and can be applied across various contexts.]
-    classDef statistical fill:#f9f,stroke:#333,stroke-width:2px;
-    classDef machineLearning fill:#ccf,stroke:#333,stroke-width:2px;
-    classDef deepLearning fill:#fcf,stroke:#333,stroke-width:2px;
-    classDef otherApproaches fill:#cfc,stroke:#333,stroke-width:4px,stroke-dasharray: 5, 5;
-    class B statistical;
-    class C machineLearning;
-    class D deepLearning;
-    class E otherApproaches;
-</mermaid>
+  * Autoregressive Integrated Moving Average (ARIMA): This model uses past observations to forecast future values, accounting for non-stationary time series data by differencing and integrating the original series. It is suitable for stationary or trending data without seasonality.
+  * Seasonal ARIMA (SARIMA): This extension of ARIMA incorporates seasonality, allowing for better modeling of cyclical patterns in data, making it ideal for time series with a regular pattern that repeats over specific intervals.
+  * Exponential Smoothing: This method uses exponentially decaying weights to combine previous observations and forecast errors, which is suitable for stationary data with trends or seasonality. It provides an easy way to smooth out time series data while preserving the underlying patterns.
+  * Prophet: Facebook's open-source tool designed for time series forecasting that automatically detects changepoints and seasonalities in data, making it ideal for complex real-world datasets with missing observations and multiple seasonalities.
+  * Long Short-Term Memory (LSTM) Networks: LSTMs are a type of recurrent neural network that can learn long-term dependencies in sequences, making them suitable for time series data with complex patterns and trends.
+  * Convolutional Neural Networks (CNNs): Originally developed for image processing, CNNs have been adapted to analyze time series data by treating each observation as a feature map. They excel at detecting local patterns and are particularly useful when analyzing high-dimensional data.
+  * Transformer Models: Initially designed for natural language processing tasks, transformers can be used in time series forecasting by considering the entire dataset as a sequence. They effectively capture long-term dependencies and provide excellent performance on complex time series tasks.
+  * Gaussian Processes: This probabilistic model offers a flexible approach to modeling nonlinear relationships between variables, making it suitable for handling noisy time series data with unknown underlying functions.
+  * Hidden Markov Models (HMMs): Suitable for analyzing hidden states or systems with changing dynamics over time that cannot be captured by simple models like ARIMA or exponential smoothing.
+  * Dynamic Time Warping (DTW): DTW is a distance metric that measures similarity between two temporal sequences while allowing them to vary in speed, making it suitable for detecting patterns and anomalies in time series data.
+When selecting an appropriate method for analyzing time series data, consider factors such as:
+  * Data characteristics (stationary or non-stationary, trending, seasonality)
+  *
+  * Complexity of the underlying patterns or trends
+  *
+  * Ability to handle missing observations
+  *
+  * Required level of interpretability and model complexity.
+Understanding these methods will help you choose the best approach for your time series analysis needs!