locked
Tutorial 4 and BPM from Command Line RRS feed

  • Question

  • Hi,

    I've been taking a looking at Tutorial 4 (Bayes Point Machine), as well as running Learners from the command line. One thing I've tried to do is to test the BinaryBayesPointMachine Classifier with the gender prediction data from Tutorial 4.  However, I don't seem to get the same results.

    When I use a file with the following training data:

    true income: 63 age: 38 x:1
    false income: 16 age: 23 x:1
    true income: 28 age: 40 x:1
    true income: 55 age: 27 x:1
    false income: 22 age: 18 x:1
    false income: 20 age: 40 x:1

    And then run Prediction with the following test data:

    income: 58 age: 36 x:1
    income: 18 age: 24 x:1
    income: 22 age: 37 x:1

    I get the following output:

    true=0.500000000000018 false=0.499999999999982
    true=0.330975886872321 false=0.669024113127679
    true=0.330975886872321 false=0.669024113127679

    Should this not return similar results to that shown in Tutorial 4?

    i.e. something similar to

    classifier output=
    [0] Bernoulli(0.9555)
    [1] Bernoulli(0.1565)
    [2] Bernoulli(0.287)

    Many thanks in advance.

    Thursday, May 5, 2016 11:36 AM

All replies

  • In the BPM tutorial the features are modelled as correlated - you can see that the weights are derived from a multivariate Gaussian. By contrast, in the BPM learner they're factorized. I'd expect the tutorial model to be slower, but potentially more accurate than the learner.

    -Y-

    Friday, May 6, 2016 10:01 AM