Sequential Monte Carlo Learning for Time Series Structure Discovery

Authors: Feras Saad, Brian Patton, Matthew Douglas Hoffman, Rif A. Saurous, Vikash Mansinghka

ICML 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Empirical measurements on real-world time series show that our method can deliver 10x 100x runtime speedups over previous MCMC and greedy-search structure learning algorithms targeting the same model family. We use our method to perform the first large-scale evaluation of Gaussian process time series structure learning on a prominent benchmark of 1,428 econometric datasets.
Researcher Affiliation Collaboration Feras A. Saad 1 2 Brian J. Patton 2 Matthew D. Hoffmann 2 Rif A. Saurous 2 Vikash K. Mansinghka 3 2 1Carnegie Mellon University 2Google Research 3Massachusetts Institute of Technology.
Pseudocode Yes Algorithm 1 SMC Structure Learning via Data Annealing
Open Source Code Yes Available online at https://github.com/fsaad/Auto GP.jl
Open Datasets Yes We further evaluate our method on 1,428 monthly datasets from M3 (Makridakis & Hibon, 2000)
Dataset Splits No The paper uses the M3 dataset for forecasting, which implies a training and test split based on time horizons ('18 forecast horizons'). However, it does not explicitly state numerical percentages or sample counts for training, validation, and test sets, nor does it cite the specific predefined split methodology for M3 in the main text.
Hardware Specification Yes All the experiments were conducted on a Google Cloud n2d-standard-48 instance (server specs: AMD EPYC 7B12 48v CPU @2.25GHz, 192 GB RAM).
Software Dependencies No The paper mentions 'Gen probabilistic programming system' and 'statsforecast and neuralforecast packages', but does not provide specific version numbers for these software dependencies.
Experiment Setup Yes For Auto GP, Algorithm 1 was run using a linear annealing schedule with 5% of the data introduced at each step; M = 48 particles; adaptive resampling with ESS = M/2 = 24; and Nrejuv = 100 MCMC rejuvenation steps. The training time points t and (demeaned) values y are linearly transformed to [0, 1].