Guan-t7/myTLAE

Community Implementation of *Temporal Latent Auto-Encoder* as described in [Temporal Latent Auto-Encoder: A Method for Probabilistic Multivariate Time Series Forecasting](https://arxiv.org/abs/2101.10460)

My impl. of Temporal Latent Auto-Encoder (TLAE)

TLAE decomposes high dimensional multivariate time series into a few basis series using MLPs, producing global predictions.

./model

• This is not an official implementation.
• TLAE*.py contain the two models as described in Temporal Latent Auto-Encoder: A Method for Probabilistic Multivariate Time Series Forecasting.
• Probabilistic version is not fully evaluated. It seems that the prediction should be somehow normalized before calculating CRPS.

./data

Get electricity, traffic dataset from DeepGLO. For example, ./data/electricity.npy contains an ndarray of the shape (370, 25000+). Construct the subset electricity-small from the period between 2014-01-01 and 2014-09-01.

Results reproduced

WAPE/MAPE/SMAPE	electricity	traffic
TCN-MF	0.106/0.525/0.188	0.226/0.284/0.247
TLAE reported	0.080/0.152/0.120	0.117/0.137/0.108
TLAE reproduced	0.090/0.200/0.128	0.135/0.156/0.127

CRPS / MSE	electricity-small	traffic
GP-Copula	0.056 / 2.4e5	0.133 / 6.9e-4
TLAE reported	0.058 / 2.0e5	0.097 / 4.4e-4
TLAE reproduced	0.069 / 3.7e6	0.121 / 4.5e-4

In short, metrics reproduced is better than previous SOTA (such as TCN-MF in DeepGLO), but not so good as reported in TLAE paper.

• RNNs are slow to train. Serious applications should replace GRU within the TLAE with TCN.
• The above metrics are calculated using rescaled (unnormalized) data. Final metrics vary using different data scaling schemes.
• Also, you might want to sequentially slide the training windows rather than randomly sampling.

Tested on a GeForce RTX 2080 Ti, using these packeges:

Python 3.9.9
torch 1.10.0
pytorch-lightning 1.5.10