Apriori Node Expert Options

For those with detailed knowledge of Apriori's operation, the following expert options allow you to fine-tune the induction process. To access expert options, set the Mode to Expert on the Expert tab.

Evaluation measure. Apriori supports five methods of evaluating potential rules.

• Rule Confidence. The default method uses the confidence (or accuracy) of the rule to evaluate rules. For this measure, the Evaluation measure lower bound is disabled, since it is redundant with the Minimum rule confidence option on the Model tab. See the topic Apriori Node Model Options for more information.
• Confidence Difference. (Also called absolute confidence difference to prior.) This evaluation measure is the absolute difference between the rule's confidence and its prior confidence. This option prevents bias where the outcomes are not evenly distributed. This helps prevent "obvious" rules from being kept. Set the evaluation measure lower bound to the minimum difference in confidence for which you want rules to be kept.
• Confidence Ratio. (Also called difference of confidence quotient to 1.) This evaluation measure is the ratio of rule confidence to prior confidence (or, if the ratio is greater than one, its reciprocal) subtracted from 1. Like Confidence Difference, this method takes uneven distributions into account. It is especially good at finding rules that predict rare events. Set the evaluation measure lower bound to the difference for which you want rules to be kept.
• Information Difference. (Also called information difference to prior.) This measure is based on the information gain measure. If the probability of a particular consequent is considered as a logical value (a bit), then the information gain is the proportion of that bit that can be determined, based on the antecedents. The information difference is the difference between the information gain, given the antecedents, and the information gain, given only the prior confidence of the consequent. An important feature of this method is that it takes support into account so that rules that cover more records are preferred for a given level of confidence. Set the evaluation measure lower bound to the information difference for which you want rules to be kept.

• Normalized Chi-square. (Also called normalized chi-squared measure.) This measure is a statistical index of association between antecedents and consequents. The measure is normalized to take values between 0 and 1. This measure is even more strongly dependent on support than the information difference measure. Set the evaluation measure lower bound to the information difference for which you want rules to be kept.

Allow rules without antecedents. Select to allow rules that include only the consequent (item or item set). This is useful when you are interested in determining common items or item sets. For example, cannedveg is a single-item rule without an antecedent that indicates purchasing cannedveg is a common occurrence in the data. In some cases, you may want to include such rules if you are simply interested in the most confident predictions. This option is off by default. By convention, antecedent support for rules without antecedents is expressed as 100%, and rule support will be the same as confidence.