average-categorize websites

Now for a more interesting example. Recall work on pattern recognition of websites. In that case, we constructed the averages manually. We manually took the average of abc {1,2,...,10}, adelaidenow {1,2,...,10} etc. Then applied simm to webpage-11. This time, we get average-categorize to find our sample classes (again, webpage-11's are filtered out). Here is the BKO:

sa: load improved-fragment-webpages.sw
sa: load label-training-data-for-website-fragments.sw
sa: average-categorize[training-hash-4B,0.7,phi,average-cat]
sa: load create-cat-ave-pat-rec-matrix.sw

Here is the result:

sa: matrix[result]
[ phi: 1 ] = [  91.69  28.73  25.76  37.77  29.45  24.33  ] [ abc 11         ]
[ phi: 2 ]   [  28.77  78.04  26.72  29.86  25.25  28.19  ] [ adelaidenow 11 ]
[ phi: 3 ]   [  25.76  26.88  79.05  28.28  26.86  23.21  ] [ slashdot 11    ]
[ phi: 4 ]   [  37.8   29.75  28.16  85.54  32.06  24.95  ] [ smh 11         ]
[ phi: 5 ]   [  29.71  25.25  26.91  31.87  85.17  22.09  ] [ wikipedia 11   ]
[ phi: 6 ]   [  24.32  28.18  23.47  24.92  21.94  82.02  ] [ youtube 11     ]

Which counts as a success! This matrix is essentially identical to this one. Except this time, we didn't specify the classes, average-categorize did that for us, resulting in phi {1,2,3,4,5,6}.

So, a nice example of unlabeled learning. Though admittedly, webpage superpositions are rather well behaved. Certainly better behaved than the adult wage prediction example! In the sense that different classes have distinct superpositions. The wage prediction example completely failed that test.

Now we have distinct classes we can give them names. Pretty simply actually:

M |phi: 1> => |abc>
M |phi: 2> => |adelaidenow>
M |phi: 3> => |slashdot>
M |phi: 4> => |smh>
M |phi: 5> => |wikipedia>
M |phi: 6> => |youtube>

Now, take a look:

sa: matrix[M,result]
[ abc         ] = [  1  0  0  0  0  0  ] [  91.69  28.73  25.76  37.77  29.45  24.33  ] [ abc 11         ]
[ adelaidenow ]   [  0  1  0  0  0  0  ] [  28.77  78.04  26.72  29.86  25.25  28.19  ] [ adelaidenow 11 ]
[ slashdot    ]   [  0  0  1  0  0  0  ] [  25.76  26.88  79.05  28.28  26.86  23.21  ] [ slashdot 11    ]
[ smh         ]   [  0  0  0  1  0  0  ] [  37.8   29.75  28.16  85.54  32.06  24.95  ] [ smh 11         ]
[ wikipedia   ]   [  0  0  0  0  1  0  ] [  29.71  25.25  26.91  31.87  85.17  22.09  ] [ wikipedia 11   ]
[ youtube     ]   [  0  0  0  0  0  1  ] [  24.32  28.18  23.47  24.92  21.94  82.02  ] [ youtube 11     ]

Or, as a merged matrix:

sa: merged-matrix[M,result]
[ abc         ] = [  91.69  28.73  25.76  37.77  29.45  24.33  ] [ abc 11         ]
[ adelaidenow ]   [  28.77  78.04  26.72  29.86  25.25  28.19  ] [ adelaidenow 11 ]
[ slashdot    ]   [  25.76  26.88  79.05  28.28  26.86  23.21  ] [ slashdot 11    ]
[ smh         ]   [  37.8   29.75  28.16  85.54  32.06  24.95  ] [ smh 11         ]
[ wikipedia   ]   [  29.71  25.25  26.91  31.87  85.17  22.09  ] [ wikipedia 11   ]
[ youtube     ]   [  24.32  28.18  23.47  24.92  21.94  82.02  ] [ youtube 11     ]

Presumably humans/children do something similar. They build up and recognize classes of objects before they even have a name for them. Only later after the classes are well specified they discover a label (in the above case, the M learn rules).