FINE TUNING PARAMETERS
Setting the "Let me fine-tune parameters" widget to True allows you to modify the parameters used to rank combinations, and can help return a more meaningful set of results.
Below the list of parameters with a short explanation.
Max number of nodes per result (0 = No max)
For simplicity's sake, you might only want to see results with up to a certain number of nodes/dimensions.
Number of desired results
Simply sets the number of rows in the returned result dataset. Defaults to 5 and can me brought up to 10. If the cumulative variance converges before the desired number of results, the tool will naturally will stop before.
Parameter aggregation algorithm
The app ranks the different combinations and selects the top results, giving appropriate "weights" to the different parameters based on the chosen scenario. This requires that the values of the different parameters be normalized, so they can be aggregated and ranked together.
There are different possible ways - different algorithms - to perform this normalization, and each algorithm might return a different ranking order and therefore different results. The app uses the simple min-max normalization as the default. If you are unsatisfied with the results you can try another normalization method. Some normalization method normalize the variance amount and percentage change differently (in fact giving them more weight) from the other parameters.
Max % of total variance per result
Sets the maximum "size" of each item in the returned result set. If one item "eats up" a large part of the variance and therefore results in a skewed set of returned rows, you can exclude that item from the accepted results.
Min % of total amount per result
Sets the minimum "size" of each item in the returned result set. Use to exclude non relevant items, for instance when the parameters are set to give weight to variance increase in percent.
Variance amount weight
If set to 1, the code will weigh the potential result items in proportion to the absolute value of their variance. Otherwise, weight will also be given to the percentage level of change of the item. Useful to highlight items that have changed a lot, even if the absolute value of their variance is modest.
Number of nodes weight
If set to 0, the code will not give weight to the number of nodes of a given potential result row. When set to a value greater than zero, more weight will be given to results with more nodes, even if their variance value is the same.
For instance the "volume variance pants Italy" combination has two nodes(country, product type) while the "volume variance red shirts Milan" combination has four nodes (country, city, product type, color) because Milan-city is a hierarchical child of Italy-country. If the parameter is set to more than zero the four-node result will be ranked higher than to the two-node result, even if the absolute value of the two variances is the same.
Useful to highlight items with high number of nodes.
Unique number of items weight
If set to 0, the code will not give weight to the number of total items contained in the columns of a given potential result row. Otherwise results "chosen" out of columns with more items will be weighted more.
The system counts the unique items in each dimension column, and calculates the total items of each combination. For instance, if we operate in a total of 5 cities, in 2 countries, where we sell a total of 10 different products in 4 different colors, the combination "volume variance pants Italy" will have a total items value of 12 (10 for the product dimension plus 2 for the country dimension).
The combination "volume variance red shirts Milan" will have a total items value of 21 (10 for the product dimension plus 4 for the color dimension, plus 2 for the country dimension, plus 5 for the city dimension). In practice, in this case, combinations that contain a specific product (of which there are 10) will be weighted more than combinations that contain a specific city (of which there are 5) because they are assumed to be more interesting.