Modeling Multivariate Time Series
Explore how to model multivariate time series by reshaping data into image-like tensors, enabling the use of Conv2D networks. Understand the equivalence between Conv1D and Conv2D approaches, and learn how neighborhood models capture local temporal and spatial dependencies efficiently for improved flexibility and performance.
We'll cover the following...
In the previous lesson, we constructed convolutional neural networks following the conventional approach of placing the temporal axis along a spatial axis and the multivariate features as channels.
Another approach is placing the features along a second spatial axis. This can be done by reshaping the input samples from (timesteps, features) to (timesteps, features, 1) tensors. The reshaping is explained in the illustration below. The reshaped time series appears in the shape of an image with a single channel. Therefore, this approach is termed as modeling a multivariate time series as an image.
In the reshaped time series sample, the timesteps and features are the spatial axes with one channel. Due to two spatial axes, the convolutional network is constructed with Conv2D.
First, we construct a Conv2D network equivalent to the baseline Conv1D network to show their interchangeability. Next, we construct another network with Conv2D, which is intended to learn the “local” temporal and spatial dependencies in and ...