X hits on this document





37 / 63

or =100 ppm Ag) and those with lower grades. Each sample was assigned a label that

identified whether the concentration was above or below these thresholds. Of the 694

samples for which we had Au and Ag assay data, only about 6% of the observations were

found to be high grade (=1 ppm Au or =100 ppm Ag) for either gold or silver.

In order to build an unbiased statistical model, the number of observations was

balanced via random sampling to contain an approximately equal number of high and low

grade samples. If an equal number of high and low grade samples was not selected to

develop the model, the model would tend to over-predict the grade with a higher

frequency in the sample population. For Au, after balancing, 78 samples were selected to

build the models (36 had Au =1 ppm and 42>1 ppm). Similarly, for Ag, 77 samples were

selected to build the models - 38 samples had >100 ppm Ag and 39 <100 ppm Ag.

After balancing, the dataset was split into a training dataset (70% of the

observations) and testing dataset (30% of the observations). The software iteratively

builds and evaluates the predictive models and ranks the resulting models based on

accuracy, i.e., the proportion of correct predictions. The models that showed the highest

“profit” or greatest accuracy for both Au and Ag were the neural network, the C5

decision tree and Quest decision tree models, all of which predicted the correct results in

about 70-75% of the tests. For both Ag and Au, most other models showed significantly

lower accuracy.

The binary classifier model building and testing software also provides

information on the relative importance of each of the variables included in the model. For

both Au and Ag, the presence of colloform silica texture was the variable with the

greatest importance, i.e., colloform texture silica has the greatest predictive power to


Document info
Document views249
Page views249
Page last viewedTue Jan 24 03:12:07 UTC 2017