locked
What does Relative Impact = 100 indicate? RRS feed

  • Question

  • I ran the 'Analyze Key Influencers' analysis for age of my customers.  One of the findings is that age less than 35 favors 'not a customer' with a relative impact of 100.   How do I interpret this? Does this mean that out of all the age groups, this group is the least likely to be a customer?

     

    Tuesday, February 1, 2011 8:06 PM

Answers

  • Relative impacts, in analyze key influencers, are normalized odds ratio, calculated inside the GetAttributeDiscrimination stored procedure included in Analysis Services. The discrimination is generally computed between 2 states of the target attribute, but for the purpose of the data mining addins, it is computed between a state and the complement of that state (all other states)

    Here is how it works, I hope this helps explaining the meaning of the numbers

    When looking for influencers for Attribute O, state OS, the stored proc scans all the input attributes and their states. The Naive bayes mining model stores, in its content, all the significant conditional probabilities linking the input attributes and their states to the output attributes and their states.

    So, in essence, the stored procedure is scanning all the conditional probabilities having the form P( O=Os | I=IS)  --wher I1 is an input attribute, IS is a state of that input attribute. The equivalent conditional probability for the complement is also computed, i.e. P( O<>Os | I=IS).

    The odds ratio is computed as the ratio between these probabilities.

    All the odds ratios (for all the input attribute/input state combinations) are normalized between 0 (the minimum) and 100 (the maximum ratio). The normalized numbers are returned as the result of the procedure

    So: relative impact of 100 means that Age < 35 yields the highest odds ratio for the "Not a Customer" state of your target (or, the ratio between Not a Customer and all other states takes its highest value in the Age < 35 slice)

     


    bogdan crivat / http://www.bogdancrivat.net/dm
    • Marked as answer by ZackBostonMA Thursday, June 2, 2011 6:11 PM
    Tuesday, February 1, 2011 11:52 PM

All replies

  • Relative impacts, in analyze key influencers, are normalized odds ratio, calculated inside the GetAttributeDiscrimination stored procedure included in Analysis Services. The discrimination is generally computed between 2 states of the target attribute, but for the purpose of the data mining addins, it is computed between a state and the complement of that state (all other states)

    Here is how it works, I hope this helps explaining the meaning of the numbers

    When looking for influencers for Attribute O, state OS, the stored proc scans all the input attributes and their states. The Naive bayes mining model stores, in its content, all the significant conditional probabilities linking the input attributes and their states to the output attributes and their states.

    So, in essence, the stored procedure is scanning all the conditional probabilities having the form P( O=Os | I=IS)  --wher I1 is an input attribute, IS is a state of that input attribute. The equivalent conditional probability for the complement is also computed, i.e. P( O<>Os | I=IS).

    The odds ratio is computed as the ratio between these probabilities.

    All the odds ratios (for all the input attribute/input state combinations) are normalized between 0 (the minimum) and 100 (the maximum ratio). The normalized numbers are returned as the result of the procedure

    So: relative impact of 100 means that Age < 35 yields the highest odds ratio for the "Not a Customer" state of your target (or, the ratio between Not a Customer and all other states takes its highest value in the Age < 35 slice)

     


    bogdan crivat / http://www.bogdancrivat.net/dm
    • Marked as answer by ZackBostonMA Thursday, June 2, 2011 6:11 PM
    Tuesday, February 1, 2011 11:52 PM
  • Thanks Bogdan that explanation is very helpful.   I'm going to ask a folloup question in another post regarding what relationship  i can draw between a correlation analysis and the key influencers report.
    Thursday, June 2, 2011 6:11 PM