Enc and EncY were trained to predict the class by passing in the multivariate time series input (x) and its corresponding ground truth one-hot encodings (y) to minimize categorical cross entropy loss [equation (5) ].
The loss weights (λrecon, λadv, and λclass) reflect the relative importance of the different functions of the denoising semi-supervised adversarial autoencoder and were determined by experimenting with different values. We found setting λrecon = 1, λadv = 0.01, and λclass = 0.01 enabled the model to strike a desirable compromise between high quality generated samples, accurate predictions, and a dense latent space approximating a standard multivariate Gaussian distribution. Experimentally, we found that losses began stabilizing after training for 10–20 epochs; therefore, we trained all models for 30 epochs to increase the likelihood of converging on a desirable solution for the generative model. Models were implemented in Keras using the Tensorflow backend on a laptop computer (Intel Core i7–7700, 2.80 GHz, 16 GB RAM) with GPU (4GB Nvidia GeForce GTX 1050) running Windows. Mode-specific generative models could be trained on the order of minutes to hours depending on the number of empirical training examples.
The loss weights (λrecon, λadv, and λclass) reflect the relative importance of the different functions of the denoising semi-supervised adversarial autoencoder and were determined by experimenting with different values. We found setting λrecon = 1, λadv = 0.01, and λclass = 0.01 enabled the model to strike a desirable compromise between high quality generated samples, accurate predictions, and a dense latent space approximating a standard multivariate Gaussian distribution. Experimentally, we found that losses began stabilizing after training for 10–20 epochs; therefore, we trained all models for 30 epochs to increase the likelihood of converging on a desirable solution for the generative model. Models were implemented in Keras using the Tensorflow backend on a laptop computer (Intel Core i7–7700, 2.80 GHz, 16 GB RAM) with GPU (4GB Nvidia GeForce GTX 1050) running Windows. Mode-specific generative models could be trained on the order of minutes to hours depending on the number of empirical training examples.