*2.3. Lack of Effective Metric to Assess the Performance of Logic Mining*

Effective metric in logic mining is crucial to ensure the actual performance of the induced logic in doing clustering and classification. According to the previous studies, the point of assessment and type of metric are still shallow and do not represent the performance of the logic mining. For instance, the work of [21] reported the error analysis learning phase of HNN but a failure to provide metrics that are related to the contingency table. As a result, the actual performance of the induced logic is still not well understood. Similar limitation reported in [14] where only metric of global minima ratio is used to demonstrate the connection between neurons. The local minimum solution signifies the induced logic rule does not correspond to the learned logic which contribute to the lack of generalization capability. In this case, if the measurement is solely based on the energy metric, then quantifying each element, in terms of confusion metric, is necessary so that the induced logic can carry out the classification task. In addition, the building block that leads to intermediate logics is solely based on the obtained synaptic weight. In this context, without synaptic weight analysis, the connection of the induced logic is poorly understood. For instance, logic mining [20] does not report the result of the strength of connection between variables in the induced logic. As a result, there is no method to assess the logical pattern stored in the content addressable memory (CAM). In this paper, comprehensive analysis, such as error analysis, synaptic weight analysis, and statistical analysis will be employed to get an overall view on the actual performance of all the logic mining models.
