The importance of input values is a relative measure of how significant each of the inputs is in the predictive model. Values range from 0 to 100. Higher values are associated with more important variables (inputs). If the importance of an input value is ever set to zero, then that input is useless and might as well be omitted. You can omit that variable in the future if you desire. In fact, if you are seeking to eliminate inputs from you model, it is probably a good bet that some of the lowest contributing variables can be safely not used in the future.
However, do not assume that if the importance value of input 1 is 10 and the importance value of input 2 is 5, that input 1 is twice as good as input 2. All we can really say with confidence is that input 1 is more important than input 2.
In fact, contribution factors are only a guide or estimation of the value of the input variables. They should never be considered highly accurate. Additionally, if you use too many inputs (more that 10), the accuracy of the contribution factors will decrease. If you are using contribution factors to decide amongst a large group of potential inputs, it is best to test them in groups of 10, taking a few of the best ones from each group of 10.
You may receive any of the following messages due to errors in your prediction, which will be explained as follows:
Insufficient Data – No data available for training set – No data is available for the specified input, you may want to verify that the data you specified exists over the period of time specified for the training set.
Constant Data – The input is constant across the training set – The specified input remains constant across all points in the training set. This provides no information that the neural network can use and thus the input is skipped.
Insufficient Data – Less than x days of data (Minimum Training Set Size) available for training set – Data exists for the specified input; however, the amount of data that exists is less than the Minimum Training Set Size that is specified in the Training Criteria. You may want to change the Minimum Training Set Size by modifying the Training Criteria or provide more data so that the neural network can use this input.
Topic of Interest: