Fig 1.
Stages of building the global MLP network model based on dispersed data.
Fig 2.
Schematic diagram of MLP network structure for one local decision table.
Fig 3.
Schematic diagram of global model.
Table 1.
Basic characteristics of data sets.
Fig 4.
Imbalance of data—cardinality of decision classes in training and test sets.
Table 2.
Results of classification accuracy acc for the proposed approach: One hidden layer, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (1AO-1HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 3.
Results of classification accuracy acc for the proposed approach: One hidden layer, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (3AO-1HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 4.
Results of classification accuracy acc for the proposed approach: One hidden layer, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (1AO-1HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 5.
Results of classification accuracy acc for the proposed approach: One hidden layer, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (3AO-1HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 6.
Results of classification accuracy acc for the proposed approach and the Vehicle data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (1AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 7.
Results of classification accuracy acc for the proposed approach and the Dry Bean data sets: Two hidden layer, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (1AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 8.
Results of classification accuracy acc for the proposed approach and the Sensorless data sets: Two hidden layer, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (1AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 9.
Results of classification accuracy acc for the proposed approach and the Crowd Sourced data sets: Two hidden layer, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (1AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 10.
Results of classification accuracy acc for the proposed approach and the Vehicle data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (3AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 11.
Results of classification accuracy acc for the proposed approach and the Dry Bean data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (3AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 12.
Results of classification accuracy acc for the proposed approach and the Sensorless data sets: Two hidden layer, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (3AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 13.
Results of classification accuracy acc for the proposed approach and the Crowd Sourced data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using average of weights and various number of neurons in the hidden layer (3AO-2HL-AVG).
Designation I is used for the number of neurons in the input layer.
Table 14.
Results of classification accuracy acc for the proposed approach and the Vehicle data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (1AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 15.
Results of classification accuracy acc for the proposed approach and the Dry Bean data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (1AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 16.
Results of classification accuracy acc for the proposed approach and the Sensorless data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (1AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 17.
Results of classification accuracy acc for the proposed approach and the Crowd Sourced data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—One artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (1AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 18.
Results of classification accuracy acc for the proposed approach and the Vehicle data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (3AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 19.
Results of classification accuracy acc for the proposed approach and the Dry Bean data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (3AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 20.
Results of classification accuracy acc for the proposed approach and the Sensorless data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (3AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 21.
Results of classification accuracy acc for the proposed approach and the Crowd Sourced data sets: Two hidden layers, the method to substitute values of missing attributes in local tables—Three artificial objects generated based on one original object, MLP networks aggregation using sum of weights and various number of neurons in the hidden layer (3AO-2HL-SUM).
Designation I is used for the number of neurons in the input layer.
Table 22.
Comparison of classification accuracy acc obtained for different numbers of artificial objects.
Fig 5.
Comparison of classification accuracy acc obtained for different aggregation method.
Table 23.
Comparison of classification accuracy acc obtained for different aggregation method and different numbers of hidden layers.
Fig 6.
Comparison of classification accuracy acc obtained for different aggregation method.
Fig 7.
Comparison of classification accuracy acc obtained for different numbers of hidden layers.
Table 24.
Comparison of classification accuracy acc obtained for the proposed method and approaches: 1AO-1HL-AVG, 3AO-1HL-AVG, 1AO-1HL-SUM, 3AO-1HL-SUM, 1AO-2HL-AVG, 3AO-2HL-AVG, 1AO-2HL-SUM, 3AO-2HL-SUM.
Fig 8.
Critical value of difference (CD) and ranks for the Nemenyi test and accuracy values and methods: Proposed approach; Homogeneous ensemble MLP; Ensemble of classifiers (KNN, DT, NB).
Groups of methods that are not significantly different (with the level of significance at 0.05) are connected.
Table 25.
Comparison of classification accuracy acc obtained for the proposed method and other methods known from the literature.
Fig 9.
Box-plot chart with (Median, the first quartile—Q1, the third quartile—Q3) the value of classification accuracy acc for the proposed method and the other approaches.
Fig 10.
Critical value of difference (CD) and ranks for the Nemenyi test and balanced accuracy values and methods: Proposed approach; Homogeneous ensemble MLP; Ensemble of classifiers (KNN, DT, NB).
Groups of methods that are not significantly different (with the level of significance at 0.05) are connected.
Table 26.
Comparison of classification balanced accuracy bacc obtained for the proposed method and other methods known from the literature.
Fig 11.
Box-plot chart with (Median, the first quartile—Q1, the third quartile—Q3) the value of balanced accuracy bacc for the proposed method and the other approaches.
Fig 12.
Critical value of difference (CD) and ranks for the Nemenyi test and F1-score values and methods: Proposed approach; Homogeneous ensemble MLP; Ensemble of classifiers (KNN, DT, NB).
Groups of methods that are not significantly different (with the level of significance at 0.05) are connected.
Table 27.
Comparison of classification F1−score obtained for the proposed method and other methods known from the literature.
Fig 13.
Box-plot chart with (Median, the first quartile—Q1, the third quartile—Q3) the value of F1−score for the proposed method and the other approaches.
Fig 14.
Critical value of difference (CD) and ranks for the Nemenyi test and precision values and methods: Proposed approach; Homogeneous ensemble MLP; Ensemble of classifiers (KNN, DT, NB).
Groups of methods that are not significantly different (with the level of significance at 0.05) are connected.
Table 28.
Comparison of classification precision obtained for the proposed method and other methods known from the literature.
Fig 15.
Box-plot chart with (Median, the first quartile—Q1, the third quartile—Q3) the value of precision for the proposed method and the other approaches.
Fig 16.
AUCROC graph for Crowd Sourced imbalanced data sets and all versions of dispersion: a) 3 local tables, b) 5 local tables, c) 7 local tables, d) 9 local tables, e) 11 local tables and three different approaches: first row graph—homogeneous ensemble of MPL networks classifiers, second row graph—ensemble of classifiers (KNN, DT, NB), third row graph—proposed approach.
Fig 17.
AUCROC graph for Crowd Sourced balanced data sets and all versions of dispersion: a) 3 local tables, b) 5 local tables, c) 7 local tables, d) 9 local tables, e) 11 local tables and three different approaches: first row graph—homogeneous ensemble of MPL networks classifiers, second row graph—ensemble of classifiers (KNN, DT, NB), third row graph—proposed approach.