Deep Factors for Forecasting

Authors: Yuyang Wang, Alex Smola, Danielle Maddix, Jan Gasthaus, Dean Foster, Tim Januschowski

ICML 2019 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We conduct experiments with synthetic and real-world data to provide evidence for the practical effectiveness of our approach.
Researcher Affiliation Industry Yuyang Wang 1 Alex Smola 1 Danielle C. Maddix 1 Jan Gasthaus 1 Dean Foster 1 Tim Januschowski 1 1Amazon Research.
Pseudocode Yes Algorithm 1 Training Procedure for Deep Factor Models with Random Effects.
Open Source Code No The paper states that algorithms are implemented in MXNet Gluon but does not provide a link or explicit statement that the source code for their specific method is publicly available.
Open Datasets Yes We conduct experiments with synthetic and real-world data to provide evidence for the practical effectiveness of our approach... electricity (E) and traffic (T) from the UCI data set (Dheeru & Karra Taniskidou, 2017; Yu et al., 2016), nyc taxi (N) (Taxi & Commission, 2015) and uber (U) (Flowers, 2015) (cf. Appendix B.2).
Dataset Splits No The paper specifies the training length and mentions using a "test set" but does not explicitly detail a separate validation set split or how data was partitioned into training, validation, and test sets with percentages or counts.
Hardware Specification Yes We use a p3.8xlarge Sage Maker instance in all our experiments.
Software Dependencies No The paper states: "Our algorithms are implemented in MXNet Gluon (Chen et al., 2015)". While MXNet Gluon is named, a specific version number is not provided, nor are other software dependencies with their versions.
Experiment Setup Yes The Deep Factor model has 10 global factors with a LSTM cell of 1-layer and 50 hidden units. The noise LSTM has 1-layer and 5 hidden units. For a fair comparison with Deep AR, we use a comparable number of model parameters, that is, an embedding size of 10 with 1-layer and 50 hidden LSTM units. The student-t likelihood in Deep AR is chosen for its robust performance. The same model structure is chosen for MQ-RNN, and the decoder MLP has a single hidden layer of 20 units. We use the adam optimization method with the default parameters in Gluon to train the DF-RNN and MQ-RNN. We use the default training parameters for Deep AR.