PubMed IDs and scores of the relevant training examples.
Started at
2009/04/29 15:21:05 GMT
Time at which query was started
Finished at
2009/04/29 15:24:10 GMT
Time at which this file was written.
Feature score method
scores_laplace_split
Name of the method used to calculate feature scores.
Docstring for the method: For feature probabilities we use a Laplace prior, of 1 success and 1
failure in total, split between the classes according to size. This
avoids problems with class skew.
Min Information Gain
2e-05
We exclude features with less than this value of
Information Gain.
Base score
-60.1620415286
The log likelihood ratio of an empty article (one in
which every feature failed to occur).
Prior score
-13.8780135322
The log of the prior probability ratio for
an article being relevant versus irrelevant (added to log likelihood ratio
to obtain the final score). Equals the logit of the estimated
prevalence of relevant articles in Medline (which may be estimated
from the input size or specified separately).
Limit
100
The maximum number of results to include.
Threshold
-30.0
Default Naive Bayes classification threshold is
zero. This threshold is the minimum log probability ratio for
predicting an article to be relevant.
Feature Statistics
Quantity
Relevant Docs
Irrelevant Docs
Number of documents
15
17031961
Number of selected, occurring features
673
29075
Total occurrences of selected features
1167
785932470
Selected features per Medline record
77.800
46.145
Of the considered feature types, 29075 features are selected out
of 3703762 occurring at least once in training data. The aggressivity of
selection is 127.386.
The complete database lists 3703762 potential features.
Features with high TF.IDF
Features with TF.IDF above 0.2 or 0.3 could make good keywords. TF.IDF is term frequency times
inverse document frequency, where we treat the set of input citations as a
single document