6.1.1. Benchmark Datasets

In this experiment, 12 publicly available datasets are obtained from UCI repository https://archive.ics.uci.edu/mL/datasets.php (accessed on 10 December 2021). These datasets are widely used in the classification field and are representative of practical classification problem. The details of the datasets are summarized in Table 1.



To avoid possible field bias, the area of interest in the dataset varies from science to social datasets. The choice of datasets is based on two aspects. First, we only select a dataset that contains more than 100 instances to preserve the statistical property of a distribution. For example, we avoid choosing balloon datasets because the number of instances is statistically too small to assess the capability learning phase of the proposed model. Second, we only select a dataset that contains more than six attributes. The choice of having more than six attributes is to check the effectiveness of the proposed model in adapting the concept of an optimal attribute selection. In other words, this experiment is unable to assess the effectiveness of the proposed model using association analysis and permutation if the number of attributes is low. Note that the state of the data will be stored in neuron by using bipolar representation *Si* ∈ {−1, 1} and each state can represent the behaviour of the dataset with respect to *Qbest*. In terms of data normalization, *k*-mean clustering [34] will be used to normalize the continuous datasets into 1 and −1. For a dataset that contains categorical data, the proposed model and the existing model will randomly select *Q ki* <sup>2</sup>*SAT*. Since the number of missing values for all datasets is very small and negligible, we replaced the missing value with a random neuron state. The experiment employs a train-split method where 60% of the dataset will be trained and 40% of the dataset will be tested [31]. Note that multi-fold validation was not implemented in this paper because we wanted to ensure that *Qbest* learned by HNN has a similar starting point for all logic mining models. A multi-fold validation method will eliminate the original point of assessment during the training phase of logic mining. Hence, the comparison among logic mining is not possible.
