advertisement: compare things at compare-stuff.com! |
An alternative experiment involves performing `predictions' for only
barrel, sandwich and ribbon architectures. These are not particularly useful predictions
because one cannot be certain in a blind prediction that the novel sequence
has one of these architectures. However, the results are interesting and
are also relevant to Section 3.5. The 94 domains (see
Table 3.3 for the numbers of each type) are predicted with 69%
accuracy (quartiles: 83%, 75%, 74% and 46%, random accuracy 43%). The improvement (compared with 50% for 12-architecture prediction) is to be expected since the task is easier. The Matthews coefficients are also higher:
,
and
.