Skip to content
/ pohmm Public

Python implementation of the partially observable hidden Markov model

License

Notifications You must be signed in to change notification settings

vmonaco/pohmm

Repository files navigation




Build Status pip version

pohmm

pohmm is an implementation of the partially observable hidden Markov model, a generalization of the hidden Markov model in which the underlying system state is partially observable through event metadata at each time step.

An application that motivates usage of such a model is keystroke biometrics where the user can be in either a passive or active hidden state at each time step, and the time between key presses depends on the hidden state. In addition, the hidden state depends on the key that was pressed; thus the keys are observed symbols that partially reveal the hidden state of the user.

Much of this code is based on the hmmlearn python package.

Install

To install the latest stable version:

$ pip install pohmm

To install the development version:

$ pip install git+git://github.com/vmonaco/pohmm.git

Example

Given an observation sequence,

TRAIN_OBS = [4, 3, 4, 2, 3, 1]

With partially observed states,

TRAIN_PSTATES = ['b', 'a', 'b', 'a', 'c', 'a']

Create and train a model:

model = pohmm.Pohmm(n_hidden_states=2,
                    init_spread=2,
                    emissions=['lognormal'],
                    init_method='obs',
                    smoothing='freq',
                    random_state=1234)

model.fit([np.c_[TRAIN_OBS]], [TRAIN_PSTATES])

To view the fitted model parameters,

np.set_printoptions(precision=3)
print(model)

# POHMM
# H-states: 2
# P-states: (3) ['a', 'b', 'c']
# Emission: ['lognormal']
# --------------------------------------------------------------------------------
# Transition probas
# ________________________________________
# . -> .
# [[ 0.5    0.5  ]
#  [ 0.333  0.667]]
# ________________________________________
# a -> .
# [[  5.000e-01   5.000e-01]
#  [  1.000e-05   1.000e+00]]
# ________________________________________
# . -> a
# [[ 0.5  0.5]
#  [ 0.5  0.5]]
# ________________________________________
# b -> .
# [[  5.000e-01   5.000e-01]
#  [  5.000e-06   1.000e+00]]
# ________________________________________
# . -> b
# [[  5.000e-01   5.000e-01]
#  [  1.000e-05   1.000e+00]]
# ________________________________________
# c -> .
# [[  5.000e-01   5.000e-01]
#  [  1.000e+00   1.000e-05]]
# ________________________________________
# . -> c
# [[  5.000e-01   5.000e-01]
#  [  1.000e-05   1.000e+00]]
# ________________________________________
# a -> a
# [[ 0.5    0.5  ]
#  [ 0.375  0.625]]
# ________________________________________
# a -> b
# [[  5.000e-01   5.000e-01]
#  [  1.000e-05   1.000e+00]]
# ________________________________________
# a -> c
# [[  5.000e-01   5.000e-01]
#  [  1.000e-05   1.000e+00]]
# ________________________________________
# b -> a
# [[ 0.5  0.5]
#  [ 0.1  0.9]]
# ________________________________________
# b -> b
# [[ 0.5    0.5  ]
#  [ 0.083  0.917]]
# ________________________________________
# b -> c
# [[ 0.5    0.5  ]
#  [ 0.083  0.917]]
# ________________________________________
# c -> a
# [[ 0.5    0.5  ]
#  [ 0.833  0.167]]
# ________________________________________
# c -> b
# [[ 0.5  0.5]
#  [ 0.5  0.5]]
# ________________________________________
# c -> c
# [[ 0.5  0.5]
#  [ 0.5  0.5]]
# --------------------------------------------------------------------------------
# Starting probas
# .: [  1.000e-05   1.000e+00]
# a: [  1.000e-05   1.000e+00]
# b: [  1.000e-05   1.000e+00]
# c: [  1.000e-05   1.000e+00]
# --------------------------------------------------------------------------------
# Steady probas
# .: [ 0.375  0.625]
# a: [ 0.5  0.5]
# b: [ 0.25  0.75]
# c: [ 0.25  0.75]
# --------------------------------------------------------------------------------
# Emissions
# ________________________________________
# Feature 0: lognormal
# .: logmu = [  1.116e-05   1.093e+00]
# a: logmu = [  4.367e-06   9.452e-01]
# b: logmu = [  2.151e-05   1.289e+00]
# c: logmu = [  9.213e-06   1.096e+00]
# .: logsigma = [ 0.004  0.262]
# a: logsigma = [ 0.002  0.218]
# b: logsigma = [ 0.005  0.087]
# c: logsigma = [ 0.003  0.131]
# --------------------------------------------------------------------------------

And obtain the loglikelihood of a test sequence:

TEST_OBS = [2, 1, 6]
TEST_PSTATES = ['c', 'a', 'd']
print(model.score(np.c_[TEST_OBS], TEST_PSTATES))

# -4.81378034867

Finally, use the model to generate data:

sample_obs, sample_pstates, sample_hstates = model.sample(n_obs=10, random_state=1234)
print('Observations:\n', sample_obs)
print('partial states:\n', sample_pstates)
print('hidden states:\n', sample_hstates)

# Observations:
# [[ 4.321]
#  [ 2.579]
#  [ 3.674]
#  [ 1.006]
#  [ 2.875]
#  [ 1.002]
#  [ 1.003]
#  [ 2.473]
#  [ 2.744]
#  [ 1.002]]
# p-states: 
#  ['b' 'a' 'b' 'a' 'c' 'a' 'b' 'a' 'c' 'a']
# h-states: 
#  [1 1 1 0 1 0 0 1 1 0]

Other examples can be found in the examples/ directory. To obtain results for che CMU Keystroke dataset, run

$ python examples/keystroke.py

How to cite

@article{Monaco2017,
title = "The Partially Observable Hidden Markov Model and its Application to Keystroke Dynamics ",
journal = "Pattern Recognition ",
year = "2017",
issn = "0031-3203",
doi = "https://doi.org/10.1016/j.patcog.2017.11.021",
url = "https://www.sciencedirect.com/science/article/pii/S0031320317304752",
author = "John V. Monaco and Charles C. Tappert",
}

About

Python implementation of the partially observable hidden Markov model

Resources

License

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published