cancel
Showing results for 
Search instead for 
Did you mean: 

SAP InfiniteInsight Recommendations

0 Kudos

Hello colleagues,

Could you please help us with the several questions regarding recommendations model in IIS:

1. What is an influence on a Model of a field "Identify Best Sellers Above" in a description of transactions? Is there optimal value for it? Is it just for classification?

2. What do terms "Maximum Iterations" and "Minimum Support" mean? What is their influence on the results of the Model? Are there optimal values for them?

3. What do terms "Minimum Confidence" and "Minimum Predictive Power" mean? What is their influence on the results of the Model? In which cases it is necessary to tick off these items? Are there optimal values for them?

4. What is a "Target Event"? How and in which cases it is necessary to tick off and define it?

5. What does a term "Max Top Nodes" mean? How does it influence the results of the Model? In which cases it is necessary to tick off this item? Is there optimal value for it?

6. What is "Rules" in the description of the Model?

7. What data does the Model put on automatically created fields of the result: "sn_rec_rule_id", "sn_rec_kxReco", "sn_rec_source", "sn_rec_score"?

Thanks in advance,

Regards

Andrey

Accepted Solutions (1)

Accepted Solutions (1)

abdel_dadouche
Active Contributor
0 Kudos

Hi Andrey,

First of all, I wish you a happy 2015 with lot of interesting questions about Predictive!!

I don’t know if you managed to get an answer to your question, but here are some inputs:

1 - What is an influence on a Model of a field "Identify Best Sellers Above" in a description of transactions? Is there optimal value for it? Is it just for classification?

  • By default “Best sellers” are marked and excluded from the graph that is built for the Recommendation model as they creating more noise than information in the graph but they are kept in a separate list to be treated as needed for your business issue. Usually you don’t want to recommend a “Best Seller” item, this is why they are marked and skipped, unless there is nothing to recommend for example.
  • By default, this limit is set to 50,000. It means that products that have been bought more than 50,000 times will be considered as best-sellers.
  • There is no optimal value, nor magic formula as it all depends on the data used and the business issue being solved.
  • Usually we apply a value equal to 4 time the standard deviation of the number of transaction, but it consider the weight if used.                

2 - What do terms "Maximum Iterations" and "Minimum Support" mean? What is their influence on the results of the Model? Are there optimal values for them?

  • The "Minimum Support" allows you to indicate how many times a pair of products has been bought by the same user in the data set to be considered valuable as a recommendation. For example if the value is 3, then we must find 3 users that bought A & B so that A & B will be considered as a recommendation rule (actually there will be two rules: R(A)=B and R(B)=A where R is the recommendation function).
  • The "Maximum Iterations" allow you to study recommendation rules of a longer path than one, like R(A,B)=C if the value is 2. This is mostly use to implement Path/Sequence Analysis.

3 - What do terms "Minimum Confidence" and "Minimum Predictive Power" mean? What is their influence on the results of the Model? In which cases it is
necessary to tick off these items? Are there optimal values for them?

  • The “Minimum Confidence” allow you to indicate the confidence below which a rule will not be used as a recommendation. With the default
    of 0.5, a rule will be used as a recommendation only if product B has been bought together with product A more than half the time product A has been
    bought.
  • The “Minimum KI” allow you to indicate a quality threshold below which a rule will not be considered valuable as a recommendation.
  • These 2 options are filtering options applied to the computed model to skip “meaningless rules”.

4 - What is a "Target Event"? How and in which cases it is necessary to tick off and define it?

  • This only applies to “Social”, I think this is a bug that this panel is available here as it requires to define a target column and a target value. Then only transaction with the Target value will be potentially linked to form rules.

5 - What does a term "Max Top Nodes" mean? How does it influence the results of the Model? In which cases it is necessary to tick
off this item? Is there optimal value for it?

  • This is a final filtering opting that allows keeping only the top rules instead of all of them. This can dramatically reduce the rule set size.

6 - What is "Rules" in the description of the Model?

  • This is the size of the rule set, in other word how many product combination have been found in the model.

7 - What data does theModel put on automatically created fields of the result: "sn_rec_rule_id", "sn_rec_kxReco", "sn_rec_source", "sn_rec_score"?

  • "sn_rec_rule_id" : an increment number that allow you to sort the product rules where 1 is the first product to recommend
  • "sn_rec_kxReco" : the product id recommeneded "sn_rec_source" : the graph from where the recomendation was extracted (in Social you can build many graph and request the rule for each)
  • "sn_rec_score" : the rule score that was used to sort the recommendation rules


Hope this will answer your interrogations

@bdel

Answers (0)