Testing Parameters

-b <k>

: Specifies a floating point smoothing parameter to be used when testing with naive Bayes. When a feature that was never encountered during training is encountered in a testing example, its strength is multiplied by the smoothing parameter (instead of a weight from the weight vector) and then the product is subtracted from the total activation of the example instead of being added to it. This parameter is not available in -interactive mode. The default is 15, which essentially makes features that were never encountered in the same example as target node contribute to 's activation as they would have if they were active in roughly of the examples in which was also active.

-i <+ | ->

: This parameter specifies whether to use incremental learning. If -i + is specified, then mistakes made during testing are used to update the network immediately after the example is tested. The network is written out after testing all examples in the input with ``.new'' appended to its original filename. If a feature was originally not written to the network due to discarding or eligibility, it can still be added in incremental learning during a different invocation of SNoW, due to the fact that only information on eligible features is written to the network file^5.2. This parameter is not available in -interactive, -evaluate, or -server modes. Note that in this mode the examples need to be labeled. Default -.

-L <k>

: Specifies a long value used to limit the amount of target IDs displayed during output with the various -o settings. More specifically, when an output mode that lists all targets and information about them is enabled and -L <k> is enabled as well, only the first targets will appear in the list. Note that these are the targets with the highest sigmoid activations.

-l <+ | ->

: This parameter specifies whether test examples are labeled or not. When set to +, SNoW checks every feature ID in every example in the input to determine if it's a target ID. When set to -, SNoW doesn't check to see if a feature is a target, which can make example parsing faster but also disables the output of evaluation statistics (just the information about how accurate SNoW's predictions were). This parameter is not available in -interactive or -evaluate mode. Default '+'.

-p <k>

: This parameter specifies a floating point prediction threshold which must be met in order for SNoW to make a prediction. This parameter can be used as a prediction confidence filtering for cases in which SNoW is not confident enough in its prediction. When testing, if the prediction confidences of the targets with the two highest prediction confidences differ by less than the prediction threshold, no prediction is made. A target ID of -1 is then output if the output mode is accuracy or winners. [Carlson et al., 2001] discusses this option and describes some experiments exhibiting its usefulness. Default 0.

-w <k>

: Specifies a smoothing value for Winnow and Perceptron learners. When a feature that was never encountered during training is encountered in a testing example, its strength is multiplied by the smoothing parameter (instead of a weight from the weight vector) and then the product is subtracted from the total activation of the example instead of being added to it. Default 0.

... file^5.2

See the -a parameter for a way to include pending features in the network