NAME Statistics::Contingency - Calculate precision, recall, F1, accuracy, etc. SYNOPSIS use Statistics::Contingency; my \$s = new Statistics::Contingency(categories => \@all_categories); while (...something...) { ... \$s->add_result(\$assigned_categories, \$correct_categories); } print "Micro F1: ", \$s->micro_F1, "\n"; # Access a single statistic print \$s->stats_table; # Show several stats in table form DESCRIPTION The "Statistics::Contingency" class helps you calculate several useful statistical measures based on 2x2 "contingency tables". I use these measures to help judge the results of automatic text categorization experiments, but they are useful in other situations as well. The general usage flow is to tally a whole bunch of results in the "Statistics::Contingency" object, then query that object to obtain the measures you are interested in. When all results have been collected, you can get a report on accuracy, precision, recall, F1, and so on, with both macro-averaging and micro-averaging over categories. Macro vs. Micro Statistics All of the statistics offered by this module can be calculated for each category and then averaged, or can be calculated over all decisions and then averaged. The former is called macro-averaging (specifically, macro-averaging with respect to category), and the latter is called micro-averaging. The two procedures bias the results differently - micro-averaging tends to over-emphasize the performance on the largest categories, while macro-averaging over-emphasizes the performance on the smallest. It's often best to look at both of them to get a good idea of how your data distributes across categories. Statistics available All of the statistics are calculated based on a so-called "contingency table", which looks like this: Correct=Y Correct=N +-----------+-----------+ Assigned=Y | a | b | +-----------+-----------+ Assigned=N | c | d | +-----------+-----------+ a, b, c, and d are counts that reflect how the assigned categories matched the correct categories. Depending on whether a macro-statistic or a micro-statistic is being calculated, these numbers will be tallied per-category or for the entire result set. The following statistics are available: * accuracy This measures the portion of all decisions that were correct decisions. It is defined as "(a+d)/(a+b+c+d)". It falls in the range from 0 to 1, with 1 being the best score. Note that macro-accuracy and micro-accuracy will always give the same number. * error This measures the portion of all decisions that were incorrect decisions. It is defined as "(b+c)/(a+b+c+d)". It falls in the range from 0 to 1, with 0 being the best score. Note that macro-error and micro-error will always give the same number. * precision This measures the portion of the assigned categories that were correct. It is defined as "a/(a+b)". It falls in the range from 0 to 1, with 1 being the best score. * recall This measures the portion of the correct categories that were assigned. It is defined as "a/(a+c)". It falls in the range from 0 to 1, with 1 being the best score. * F1 This measures an even combination of precision and recall. It is defined as "2*p*r/(p+r)". In terms of a, b, and c, it may be expressed as "2a/(2a+b+c)". It falls in the range from 0 to 1, with 1 being the best score. The F1 measure is often the only simple measure that is worth trying to maximize on its own - consider the fact that you can get a perfect precision score by always assigning zero categories, or a perfect recall score by always assigning every category. A truly smart system will assign the correct categories and only the correct categories, maximizing precision and recall at the same time, and therefore maximizing the F1 score. Sometimes it's worth trying to maximize the accuracy score, but accuracy (and its counterpart error) are considered fairly crude scores that don't give much information about the performance of a categorizer. METHODS The general execution flow when using this class is to create a "Statistics::Contingency" object, add a bunch of results to it, and then report on the results. * \$e = Statistics::Contingency->new() Returns a new "Statistics::Contingency" object. Expects a "categories" parameter specifying the entire set of categories that may be assigned during this experiment. Also accepts a "verbose" parameter - if true, some diagnostic status information will be displayed when certain actions are performed. * \$e->add_result(\$assigned_categories, \$correct_categories, \$name) Adds a new result to the experiment. The lists of assigned and correct categories can be given as an array of category names (strings), as a hash whose keys are the category names and whose values are anything logically true, or as a single string if there is only one category. If you've already got the lists in hash form, this will be the fastest way to pass them. Otherwise, the current implementation will convert them to hash form internally in order to make its calculations efficient. The "\$name" parameter is an optional name for this result. It will only be used in error messages or debugging/progress output. In the current implementation, we only store the contingency tables per category, as well as a table for the entire result set. This means that you can't recover information about any particular single result from the "Statistics::Contingency" object. * \$e->micro_accuracy Returns the micro-averaged accuracy for the data set. * \$e->micro_error Returns the micro-averaged error for the data set. * \$e->micro_precision Returns the micro-averaged precision for the data set. * \$e->micro_recall Returns the micro-averaged recall for the data set. * \$e->micro_F1 Returns the micro-averaged F1 for the data set. * \$e->macro_accuracy Returns the macro-averaged accuracy for the data set. * \$e->macro_error Returns the macro-averaged error for the data set. * \$e->macro_precision Returns the macro-averaged precision for the data set. * \$e->macro_recall Returns the macro-averaged recall for the data set. * \$e->macro_F1 Returns the macro-averaged F1 for the data set. * \$e->stats_table Returns a string combining several statistics in one graphic table. Since accuracy is 1 minus error, we only report error since it takes less space to print. * \$e->category_stats Returns a hash reference whose keys are the names of each category, and whose values contain the various statistical measures (accuracy, error, precision, recall, or F1) about each category as a hash reference. For example, to print a single statistic: print \$e->category_stats->{sports}{recall}, "\n"; Or to print certain statistics for all categtories: my \$stats = \$e->category_stats; while (my (\$cat, \$value) = each %\$stats) { print "Category '\$cat': \n"; print " Accuracy: \$value->{accuracy}\n"; print " Precision: \$value->{precision}\n"; print " F1: \$value->{F1}\n"; } AUTHOR Ken Williams COPYRIGHT Copyright 2002 Ken Williams. All rights reserved. This distribution is free software; you can redistribute it and/or modify it under the same terms as Perl itself.