By reducing the dimensionality of a model parameter space, this strategy allows the space to be explored in more detail. The other possible strategy consists of refining the whole by discarding models that use weak attributes. We hope that such refinement can improve the performance of BMA. To test the hypothesis formulated in section 2 and refine the sets of DT models obtained with BMA, we propose a new strategy aimed at eliminating DT models that use weak attributes. According to this strategy, the BMA technique described in Section 2 is first used to collect DT models. Then, the posterior probabilities of using the attributes in the set of DT models are estimated. These estimates give us a posteriori information about the importance of features. After obtaining a range of posterior probabilities, we then set a threshold value to exclude attributes whose probabilities are lower than this threshold – we define these attributes as low. In the next step, we find the DT models that use these weak attributes and finally discard these DT models from the set. Obviously, the higher the threshold value, the lower the number of attributes is defined, and therefore the greater part of the DT models is rejected. The effectiveness of this elimination technique is evaluated in terms of the accuracy of the refined DT set on the test data. The uncertainty in the overall results is evaluated in terms of entropy. Having a set of threshold probability values ​​obtained in a series of experiments, we can expect that there is an optimal threshold value at which performance becomes higher. We can also expect to find a threshold value at which the uncertainty becomes lower. In the next section, we test the proposed technique on the p...... middle of paper ...... the threshold is gradually increased from 0.0 to 0.005. At the same time, the uncertainty in decisions decreases from 478.4 to 469.0 in terms of entropy E of the ensemble. For comparison, we applied a technique of removing the same weak attributes and then re-running BMA on the dimensionally reduced data. From Table 1, we can see that the performance of BMA increased slightly from 27.4 to 29.0 when 23 weak attributes were removed. Removing 31 attributes resulted in a decrease in overall entropy from 478.3 to 463.6. Overall, both techniques are shown to provide comparable performance and overall entropy. However, the technique of removing attributes was found to tend to work with greater variation. As part of this technique, for each threshold value, it is necessary to retrain the DT set on the data of a new dimensionality..