bayesCPD

Preliminaries

BayesCPD is a Python library for performing efficient changepoint detection in a flexible class of Bayesian models, including conjugate exponential family models and a number of latent variable models. Currently, this only supports online changepoint detection.

Dependencies:

• Python 2.7
• Numpy
• Scipy
• Cython
• Matplotlib
• Pandas
• IPython and IPython Notebook

Building:

After cloning the git repository, you'll need to add this directory to your PYTHONPATH. Then to build, run the script

$ ./compile.sh

Performing Changepoint Detection

Overview

To create a changepoint model in general, create a class which inherits from BaseModel and implement the abstract methods. The main components involved are implementing the predictive distribution and updating the model's parameters.

There are a few modes of inference currently supported:

• Time-independent inference
• Full changepoint detection
• Efficient changepoint detection approximation

In general, to run inference for a specific model, you create a dictionary containing the values of the prior parameters with a string naming the variable and a value.

# Normal-Normal model
from bayesCPD.models.normal import NN

# constant Hazard function
from bayesCPD.utils.Modelutils import constant_hazard

prior_params = {'mu', 0., 'tau', 0.}

Then feed those parameters to the model, specifying which mode to use.

model = NN(prior, ivar=1, mode='all')

You can either generate data from the prior (or some new set of values)

# generate 1000 datapoints from the prior
model.generate_data(T=1000, hazard=constant_hazard, rate=250)
model.plot_data()

Or load your own dataset:

model.load_data(my_dataset)

Then run inference and plot results:

# run online changepoint detection
model.inference(rate=250)
model.plot_results()

In this particular model, it is implemented such that the predictive means are plotted on the data in the top plot. The blue solid line is the predictive mean when using the changepoint model, and the orange dotted line is the predictive mean under time-independent online Bayesian updating.

We have implemented several models you can currently run in the models directory, including univariate and multivariate versions of conjugate Gaussian models, Gaussian and other mixture models, and latent Dirichet allocation.

See notebooks/normal_example.ipynb for this example for details. More examples for other models are also in the notebooks directory.

Implementing a model

(Coming soon)