Why order of evidence provided for creating Bayesian Networks node CPD alters inference probabilities

[saurabh1deshpande]

Hi, I was trying out 2. Bayesian Networks.ipynb and swapped the evidence while creating cpd for grade, ['D', 'I' ] instead of ['I', 'D'] and all the inference probabilities changed. Intuitively I feel the order should not change the inference probabilities.

l_dist = infer.query(['L'],evidence={'I':0, 'D':0}) print(l_dist.get('L'))

For ['I', 'D'] ╒═════╤══════════╕ │ L │ phi(L) │ ╞═════╪══════════╡ │ L_0 │ 0.6114 │ ├─────┼──────────┤ │ L_1 │ 0.3886 │ ╘═════╧══════════╛

For ['D', 'I' ] ╒═════╤══════════╕ │ L │ phi(L) │ ╞═════╪══════════╡ │ L_0 │ 0.3489 │ ├─────┼──────────┤ │ L_1 │ 0.6511 │ ╘═════╧══════════╛

May be this is a naive question but would like to know if it is correct behavior and what might be causing this ?

Thanks !

[ankurankan]

@saurabh1deshpande Did you also change the values accordingly? Because we use the order of evidence variables to infer which value corresponds to any specific state. For example, a CPD on G with the evidence [D, I] would expect a values array of the form:

                 D_0         |          D_1
----|------------------------|--------------------
    |   I_0     |     I_1    |     I_0     |     I_1
G_0 |           |            |             |
G_1 |           |            |             |
G_2 |           |            |             |

whereas the evidence [I, D] would expect the values array to be of the form:

                 I_0         |          I_1
----|------------------------|--------------------
    |   D_0     |     D_1    |     D_0     |     D_1
G_0 |           |            |             |
G_1 |           |            |             |
G_2 |           |            |             |

As you can see, depending on the order of the evidence, different values can be assigned to each state, which would result in different results.

[saurabh1deshpande]

ah !..no I didn't..thnx for the explanation !