• 1 reply
  • Latest Post - ‏2012-01-19T15:36:15Z by SystemAdmin
203 Posts

Pinned topic Specifying weight information in association rules mining

‏2012-01-11T16:37:44Z |

I have several of questions regarding the meaning and use of weight information in association rules mining.

1. How weight information is calculated in model creation? The Reedbook Dynamic Warehousing: Data Mining Made Easy says the following:

The aggregated weight of an association rule can be defined as the sum of all item weights multiplied by their multiplicities. For example, if those transactions that support the association rule '{milk ∧ baby food} ==> {diapers
}' contain on average 2.3 cartons of milk at $1 each and 4.7 portions of baby food at $0.80 each and 1.2 packets of diapers at $9.90, the average weight of the association rule is $15.56.

However, as far as I know, the algorithm does not take care of the number of units of each product in a market basket analysis, so I don't understand how the rule weight is calculated

2. In which situations the weight standard deviation, minimum and maximum are calculated? I only get the mean value (this is not the case for the TA weight, where all these metrics appears in the model)

3. Finally, when the weight of the items is not fixed but can change from one transaction to the other, a separate table for weight information cannot be used, but a weight column must be present in the main transaction table. When creating a rule model through the SQL Mining API, how this column can be specified? Having a look at the SQL code generated in Design Studio, I can see that DM_addWeight method is used, pointing to the same transaction table with a reference to the item and weight fields. Then, the DM_setFldWeight method links the previous weight with the item field. However, at the end, this seems to implement a separated-table-weight approach. If this is the case, I don't understand how weights can be retrieved and assigned to each item, since the item field column is not unique

Many thanks in advanced
Updated on 2012-01-19T15:36:15Z at 2012-01-19T15:36:15Z by SystemAdmin
  • SystemAdmin
    203 Posts

    Re: Specifying weight information in association rules mining


    At least, I think I have already an answer for the third of my previous questions.

    Seems that the algorithm is aware of the case where the weight table points to the same transaction table, allowing for different weights for the same item. In other words, if you specify a separate weight table, then the item column in this table must be unique. If this is not the case, a warning is generated (IDM3519: For the item field value %3, the weight values %1 and %2 are defined. The weight value %2 is ignored).

    Any suggestion for the rest of the questions?

    Many thanks in advanced.