Efficient Online Inference for Nonparametric Mixture Time Series

This code corresponds to our UAI 2021 paper. In this paper, we provide an efficient streaming inference algorithm for infinite mixture models using a novel Bayesian recursion on the Chinese Restaurant Process.

To make streaming inference possible, we break the CRP's conditional distribution $$p(z_t|z_{<t}, \alpha)$$ on the complete history by replacing it with a marginal distribution $$p(z_t|\alpha)$$. The running sum of the previous marginal distributions $$\sum_{t' < t} p(z_{t'} = k)$$ (left) and the Chinese restaurant table distribution $$p(K_{t-1} = k)$$ (middle) compete to determine the next marginal distribution $$p(z_t = k)$$ (right). Note the logarithmic scaling.

Setup

After cloning the repository, create a virtual environment for Python 3:

python3 -m venv rcrp

Then activate the virtual environment:

source rcrp/bin/activate

Ensure pip is up to date:

pip install --upgrade pip

Then install the required packages:

pip install -r requirements.txt

We did not test Python2, but Python2 may work.

Running

Each experiment has its own directory, each containing a main.py that creates a plots subdirectory (e.g. exp_00_crp_prior/plots) and then reproduces the plots in the paper. Each main.py should be run from the repository directory e.g.:

python3 exp_00_crp_prior/main.py

Contact

Questions? Comments? Interested in collaborating? Open an issue or email Rylan Schaeffer at rylanschaeffer@gmail.com and cc Ila Fiete at fiete@mit.edu.