Combination of Active Learning and Semi-Supervised Learning under a Self-Training Scheme

Abstract

One of the major aspects affecting the performance of the classification algorithms is the amount of labeled data which is available during the training phase. It is widely accepted that the labeling procedure of vast amounts of data is both expensive and time-consuming since it requires the employment of human expertise. For a wide variety of scientific fields, unlabeled examples are easy to collect but hard to handle in a useful manner, thus improving the contained information for a subject dataset. In this context, a variety of learning methods have been studied in the literature aiming to efficiently utilize the vast amounts of unlabeled data during the learning process. The most common approaches tackle problems of this kind by individually applying active learning or semi-supervised learning methods. In this work, a combination of active learning and semi-supervised learning methods is proposed, under a common self-training scheme, in order to efficiently utilize the available unlabeled data. The effective and robust metrics of the entropy and the distribution of probabilities of the unlabeled set, to select the most sufficient unlabeled examples for the augmentation of the initial labeled set, are used. The superiority of the proposed scheme is validated by comparing it against the base approaches of supervised, semi-supervised, and active learning in the wide range of fifty-five benchmark datasets.

1. Introduction

The most common approach established in machine learning (ML) is supervised learning (SL). Under the SL schemes, classifiers are trained using purely labeled data. In contrast with the problem complexity, the performance of such schemes is directly analogous to the amount and the quality of labeled data which are used at the training phase. In a large variety of scientific domains, such as object detection [1], speech recognition [2], web page categorization [3], and computer-aided medical diagnosis [4,5,6] vast pools of unlabeled data are often available. Though, in most cases labeling data can be costly and time-consuming, as human effort and expertise are required to annotate the available data. Many research works [7] exist focusing on techniques with the aim of exploiting the available unlabeled data especially in favor of classification problems. The most common learning methods incorporating such techniques are active learning (AL) and semi-supervised learning (SSL) [8]. Both AL and SSL share an iterative learning nature, making them a perfect fit for constructing more complex combination learning schemes.

The primary goal of this paper is to put forward a new AL and SSL combination algorithm in order to efficiently exploit the plethora of available unlabeled data found in most of the ML datasets and provide an improved classification framework. The general flow of AL and SSL frameworks is presented in Figure 1. Both methods utilize an initial pool of labeled and unlabeled examples with the goal of efficiently augmenting the available knowledge. AL and SSL frameworks, in most cases, operate under an iterative logic aiming to predict the label in the most appropriate unlabeled examples. While the former method annotates the unlabeled instances by interactively querying a human expert based on a variety of querying strategies, the latter attempts to automatically produce the labels of unlabeled examples by exploiting the previously learned knowledge and a wide range of unlabeled instances selection criteria. After the successful augmentation of the initial labeled set, a final model is constructed in both cases with a view to the application on the unknown test cases.

Figure 1. The general frameworks of active learning and semi-supervised learning along with their shared elements.

As both methods share a lot of key characteristics, a major effort is now needed to combine the two learning approaches. The main contribution of the proposed algorithm is the employment of a self-training scheme for the combination of AL and SSL utilizing the fast and effective metrics of the entropy and the distribution of the prediction probabilities of the available unlabeled data. The plethora of experiments carried out, also play a major role in the validation of the proposed algorithm. The proposed method is examined through a number of different individual base learners, where the ensemble learning technique is also explored as the aggregated models tend to produce more accurate predictions and are commonly used in today’s applications [9,10].

Real-world case scenarios where AL and SSL combination methods can be applied include natural language processing (NLP) problems to which a lot of labeled examples are required to effectively train a model and also vast amounts of unlabeled data can be mined. Common applications on the NLP field are part of speech tagging, named entity recognition, sentiment analysis [11], fraud detection, and spam filtering. Especially, a number of AL [12], SSL, and combinations [13] of them have been proposed in the spam filtering domain. In Figure 2, an application on the Spambase [14] benchmark dataset briefly presents the accuracy improvement for the proposed scheme as the algorithm’s iterations progress. With regard to the base algorithm learner, the support vector machines (SVMs) [15] classifier was embedded. For comparison, in the same figure, the corresponding SSL part of the algorithm was fed with the same amount of unlabeled data to obtain only the semi-supervised accuracy.

Figure 2. Progression of accuracies in relation to the number of iterations executed for the proposed combination scheme and its semi-supervised counterpart, utilizing support vector machines (SVMs) as base learner, applied on the Spambase dataset using two different labeled ratios.

The rest of this research work is organized as follows: In Section 2, the related work on similar classification methods is discussed. Following in Section 3, the proposed method is presented along with the exact algorithm implemented. An attempt to evaluate the efficacy of the combination scheme is made in Section 4, where extensive experimentation results can be found. Moreover, in this section, the average accuracies of the classifiers applied on the combination scheme are also briefly compared. In Section 5, a modification of the scheme is explored. The research conclusions are conferred in Section 6, where a number of areas to be explored as future work are mentioned. Finally, a software implementation of the wrapper algorithm is found in the Appendix A through the accompanying link.

2. Related Work

AL can be considered one of the most promising approaches for improving the performance of a prediction model in real-world scenarios where large amount of data exists, but their labeling is costly or infeasible [7]. AL assumes that human experts will be available to provide ground-truth labels for the unlabeled instances. Therefore, the philosophy of AL is to minimize the number of queries with the explicit goal to focus the labeling effort in the most profitable or informative instances, in other words, to minimize the training cost of the model [16]. Finally, these manually annotated samples are merged with the training dataset to get the highest classification accuracy. Several query strategies [7,17] have been proposed to measure the informativeness or the representativeness of the data. Informativeness-based strategies measure the contribution of an unlabeled instance on the uncertainty reduction of a statistical model, while representativeness-based strategies measure the instance contribution on representing the underlying structure of input patterns. The most commonly used query strategies can be considered certainty-based sampling, query-by-committee, and expected error reduction. In the first type of strategy, a single model is trained and the human expert (annotator) is queried to label the least confidence unlabeled instances based on the pre-trained model. The query-by-committee strategy involves more than one active learner (classification models) to be trained for the classification task. The unlabeled instances about which these models disagree the most are selected for human annotation. The third strategy is a decision-theoretic approach aiming to estimate the potential of the model’s generalization error reduction. In other words, a model is trained and used to estimate the expected future error of the unlabeled samples. Then, the instances with the minimal future error (risk) are selected and delivered for manual labeling. The effectiveness of AL and various query strategies has been shown in typical classification tasks, such as text classification [18], speech recognition [19], speech emotion classification [20], audio retrieval [21] to name a few.

In contrast to AL, SSL aims to automatically exploit unlabeled data in addition to labeled data to improve learning performance, without human intervention. In SSL, two basic assumptions about the data distribution are considered. The first assumes that data are inherently clustered, meaning that instances belonging to the same cluster have the same label. The other one assumes that data lie on a manifold, meaning that nearby samples have similar predictions. The idea behind both is that similar data points should have similar outputs and the unlabeled instances can expose similarities between these data points. Many different SSL methods have been designed in machine learning, including mainly transductive support vector machines [22], graph-based methods [23,24,25], co-training [26], self-training [1]. In the self-training scheme, the classification model is used to predict the labels of a portion of the unlabeled instances and, consequently, the most confident ones are added to the initial training dataset repeatedly until convergence. Rather than just relying on a unique model, in co-training [26] ensemble method is employed. For each model, separate feature sets (or views) of the same labeled data are used for training. Then, like self-training, the most confident predictions of each classifier on the unlabeled data are used to iteratively construct additional labeled training data. The co-training paradigm relies on three assumptions about the views, i.e., sufficiency, compatibility, and conditional independence [26]. On the other hand, graph-based methods treat all the samples (both labeled and unlabeled) as connected vertices (nodes) in a graph, aiming to connect these nodes, in other words, to weight these node-to-node pairwise edges by similarities between the corresponding sample pairs. Finally, minimum energy optimization is used to propagate labeling from the labeled to the unlabeled nodes.

Although AL has led to the reduction of the human labeling burden, without sacrificing the model’s performance [7], it is still inefficient in some situations, e.g., the acquisition of a large amount of human annotations is impractical or not feasible at all. Thus, SSL comes in handy by minimizing the unlabeled data that will be fed to the human annotator. Specifically, human experts are required to label only those instances with the lowest certainty (as determined by the AL algorithm), while the remaining instances are automatically labeled by a machine annotator (by the SSL algorithm). Indeed, several studies have been proposed that combine AL and SSL under the same methodology. One of the first attempts [27] was in the text classification field, where expectation maximization was employed along with pool-based active learning. Later, Muslea et al. proposed the combination of co-testing and co-training showing improved classification accuracy in Web pages and pictures classification. During co-training, two classifiers are trained separately on two different views, and only the contention points, i.e., the unlabeled instances in which the classifier disagrees the most, were selected for human annotation. Finally, expectation maximization co-training (co-EM) was employed to automatically label instances that showed a low disagreement between the two classifiers. Other studies exploited certainty-based AL with self-training aiming to manual labeling with minimum human cost in spoken language understanding [28], natural language processing [29], sound classification [30], disease classification [31] and cell segmentation [32]. In another study [33], the authors addressed the problem of imbalanced training data in object detection. First, a simple object detection model was trained using a small portion of perfect samples instead of using the entire training dataset, while the imperfect samples were partitioned into several batches. Then, a batch-mode learning of AL and SSL combination was employed by integrating the uncertainty and diversity criteria from the concept of AL and the confidence criterion from that of SSL.

3. Proposed Method

The proposed method constitutes a combination of AL and SSL approaches, in order to leverage the advantages of both techniques. A mixed self-training method is employed utilizing the entropy of unlabeled instances, with the aim to identify the most confusing instances in the case of active round, while in the semi-supervised round the internal learner’s distribution of probabilities for all possible labels per each instance is exploited as a sorting mechanism for the selection of the most confident examples.

Algorithm 1: Combination Scheme

1: LOAD the dataset D and construct the labeled set L and the unlabeled set U 2: INITIALIZE the classifier CLS 3: CALCULATE the labeled ratio R = size(L)/size(L+U) 4: DEFINE the maximum number of iterations MaxIter 5: DEFINE the maximum percentage of unlabeled examples to be added in each iteration T in respect with R 6: SET maxUnlabPerIter = T * R * size(D) 7:   8: SET i = 0 9: WHILE i<MaxIter AND size(Ui)>0: /* where U0= U */ 10:  Train(CLS) on the current labeled set Li /* where L0=L */ 11:  IF i modulo 2 == 0: 12:   Classify(Ui) using CLS and construct matrix Mpr containing corresponding prediction probabilities along with the predicted labels 13:   SORT Mpr descending according to the prediction probabilities 14:   STORE the top maxUnlabPerIter instances of Mpr in a matrix Mfinal 15:    /* now containing the most confident instances along with their predictions */ 16:  ELSE: 17:   Calculate the distribution_of_probabilities(Ui) and return a matrix DistUi 18:   Calculate the entropy(DistUi) for each element and return a matrix EntrUi 19:   SORT EntrUi descending according to their entropies 20:   Label the top maxUnlabPerIter using human expertise 21:   STORE the top maxUnlabPerIter instances along with their labels in a matrix Mfinal 22:    /* now containing the most confusing instances along with their true labels */ 23:  END_IF 24:  Augment(Li) by adding Mfinal instances 25:  Clean(Ui) by removing Mfinal instances 26:  SET i = i + 1 27: END_WHILE 28:   29: Train(CLS) using Laugmented (☰ Llast iteration) 30: LOAD the unknown test cases as Testset 31: Classify(Testset) using CLS to produce the final predictions

Uncertainty-based metrics are widely deployed in the AL field as the literature suggests [34], mainly due to their strong performance in terms of calculation efficiency and effectiveness in the process of selecting the most confusing instances. On the other hand, in the SSL field research works exist [8,35] proving the effectiveness of probabilistic iterative schemes. As the nature of these types of metrics is similar, they can prove to be a robust combination for the construction of schemes such as the proposed. Moreover, it is also known [36] that the SSL self-training technique further helps to overpass the lack of exploration problems that occur during the AL entropy-based training process causing the algorithms to stuck at suboptimal solutions, continuously selecting instances which do not improve the current classifier.

The proposed algorithm can be characterized as a simple yet very effective wrapper algorithm that can utilize a wide range of learners, assuming that they can produce probability distributions for their predictions. A detailed presentation of the algorithm follows in the next paragraphs.

Let D denote the initial training set, consisting of a labeled set of examples L and an unlabeled set of examples U thus defining a labeled ratio R, as in the following equation:

$$LabeledRatio=\frac{size(L)}{size(L+U)}$$

(1)

where the $size\left(X\right)$ function returns the size of a set of instances.

Initially, a base learner (CLS) is selected and trained on L. Afterwards, a self-training scheme is employed with the aim to augment the L using the available unlabeled examples of D. The number of unlabeled examples utilized in each iteration is conservatively selected taking in account the size of the initial labeled set, using also a control parameter T, setting the percentage of unlabeled examples related to the size of the initial labeled set. The number of maximum unlabeled instances selected in each iteration is calculated as follows:

$$\max UnlabPerIter=T*R*size(D)$$

(2)

In each iteration (i), one of the two learning approaches is employed successively. The self-training loop terminates in a maximum number of iterations MaxIter or in the case of exhaustion of the pool of unlabeled examples.

Starting with the SSL round, the CLS is applied on the current unlabeled set Uⁱ and a matrix of predictions M_pr is constructed along with the prediction probability for each unlabeled instance, resulting in a $size\left({\mathit{U}}^{\mathit{i}}\right) x \left(l+2\right)$ dimensions matrix, where $l+2$ is the number of features, including the predicted labels and the corresponding prediction probabilities. The SSL round uses machine labeling in order to balance the expensive human effort and examination process required to label the data. The M_pr is sorted descendingly utilizing the prediction probabilities while the rest of the maxUnlabPerIter elements are discarded. The maxUnlabPerIter instances along with their predicted labels are stored in M_final.

Following the method flow, an AL round is deployed in every other iteration. In this round, the algorithm attempts to construct a matrix containing the entropy estimation of each unlabeled instance EntrUⁱ. The base learner is applied on Uⁱ and the distribution of probabilities are exported in matrix DistUⁱ of dimensions $size\left({\mathit{U}}^{\mathit{i}}\right) x num\_classes\left(\mathit{D}\right)$, where the $num\_classes\left(X\right)$ function returns the number of classes of a dataset. Having produced DistUⁱ, the calculation of entropy estimation matrix is performed using the next formula, to compute each one of its elements (j):

$$Entrop{y}_j={\displaystyle \sum _{k=1}^{num\_classes(D)}-p_k*{\log }_2{p}_k}$$

(3)

where $p_k$ denotes the probability of k class for instance j, already contained in DistUⁱ.

Subsequently, EntrUⁱ is sorted in descending order, as the most confusing examples, with entropy values near one, should be placed on the top of the matrix. The top maxUnlabPerIter instances are kept in EntrUⁱ with the rest of them being discarded. Human expertise is utilized to label the maxUnlabPerIter instances and a matrix containing the human-labeled instances M_final is constructed with the size of $\mathit{maxUnlabPerIter} x \left(l+1\right)$, where $l+1$ is the number of features, including the class.

During each iteration, the M_final instances are added to the current labeled set Lⁱ and removed from the current unlabeled set Uⁱ. The CLS is re-trained at the start of each self-training iteration in order to be utilized again. When the termination criteria are met, the algorithm exits the self-training loop having constructed the augmented labeled set L^augmented ($\equiv$L^{last iteration}). As a final step, the CLS is trained on the augmented labeled set in order to be applied on the unknown test cases. The exact implementation of the combination scheme is presented in Algorithm 1.

4. Experimentation and Results

In order to examine the efficacy of the proposed scheme, an exhaustive experimentation procedure was followed. At first, fifty-five (55) benchmark datasets were extracted from the UCI repository [14], related to a wide range of classification problems. To further enhance the variance and complexity of the classification process, all datasets were partitioned and examined according to the resampling procedure of k-fold cross-validation [37]. Following the method’s steps, each subject dataset is shuffled and then divided into k unique data groups. By holding out one of the groups as a test set and utilizing the rest as a train set, k new datasets are generated. The k parameter was set equal to ten, as it is commonly selected by the majority of the literature.

The main aim of the experimentation process was to prove the superiority of the combination scheme against the competing methods of the supervised, semi-supervised and active learning using always the same amounts of labeled and unlabeled data under the same base learner model. In more detail, the supervised method is trained only on the initial labeled set while the semi-supervised rival method utilizes also the initial unlabeled set in the same manner that is also exploited in the proposed combination scheme. Moreover, as baseline AL opponent the random sampling [7] process is implemented in a similar way with the rest of the combination self-training procedure, also utilizing the initial unlabeled set.

For this purpose, all training subsets were further divided into two sets, an initial labeled set and an initial unlabeled set, using four different labeled ratios R. As the initial datasets contained a hundred percent of the instance labels, in order to simulate the human expert labeling process, all the original labels for the constructed unlabeled sets were stored separately in order to be retrieved whenever the algorithm required to query the human expert. Thus, each original dataset was augmented into forty derived datasets. In detail, the R values were set to 10%, 20%, 30%, and 40%. As regards the proposed algorithm’s parameters, the control parameter T was set equal to 10%, while the MaxIter parameter was empirically selected equal to 10 in order to impose a maximum of 40%, in relation to the original dataset size, limit (can be calculated using Equation (2) multiplied by the MaxIter parameter of unlabeled instances for selection and augmentation of the initial labeled set in the case of R = 40%.

As a comparison measure, the average classification accuracy over each R was used. In order to draw general conclusions for the efficacy of the combination scheme, a wide range of classification models and meta-techniques were employed, incorporated in each one of the four learning methods. A brief description for each one of the base learners is presented:

• BagDT: In this model, the bootstrap aggregating (bagging) [38] meta-algorithm was applied along with the use of the C4.5 decision trees [39] classifier. The bagging technique is often adopted to reduce the variance and overfitting of a base learner and enhance its accuracy stability. The basic idea behind this technique is the generation of multiple training sets by uniformly sampling the original dataset.

• 5NN: The k-nearest neighbors [40] classifier belongs to the family of lazy learning algorithms. By examining the k closest instances in a defined feature space, it classifies a given test instance by plurality voting on the labels of the k instances.

• Logistic: The logistic regression, also commonly referenced as the logit model, is a statistical model that utilizes the logistic function in order to model binary dependent variables, thus fitting very well with categorical targets. In problems where the target variable has more than two values, multinomial logistic regression is applied [41].

• LMT: The logistic model tree [42] classification model combines logistic regression with decision trees. The main idea behind the classifier is the use of linear regression models as leaves of a classification tree.

• LogitBoost: This classifier is a boosting model proposed by Friedman et al. [43]. It is based on the idea that the adaptive boosting [44] method can be thought as a generalized additive model, thus the cost function of logistic regression can be applied.

• RF: One of the most robust ML learners is the random forests [45] model, which is capable of tackling regression and classification problems. Its operation is based on the construction of multiple decision trees using random subsamples of the original feature space. The aggregation of the results is achieved via majority voting. Due to its inner architecture, it is known to efficiently handle the overfitting phenomena.

• RotF: The rotation forest model constitutes an ensemble [46] classifier proposed by Rodriguez and Kuncheva [47]. Following the flow of this algorithm, the initial feature space is divided in random subspaces. The default feature extraction algorithm applied to create the subspaces is the principal component analysis (PCA) [48], aiming to increase the diversity amongst the base learners.

• XGBoost: The extreme gradient boosted trees [49] algorithm, is a powerful implementation of gradient boosted decision trees. Under this boosting [50] scheme, a number of trees are built sequentially with each time the goal to reduce errors produced from the previous tree, thus each tree is fitted on the gradient loss of the previous step. The final decision is produced from the weighted voting of the trees. The XGBoost algorithm is a very scalable algorithm that has shown to perform very well on large datasets or sparse datasets utilizing parallel and distributed execution methods.

• Voting (RF, RotF, XGBoost): As a last effort to further explore the potential of more complex classification models in the combination scheme, the construction of an ensemble classifier by majority voting the results of three of the most robust models: RF, RotF, and XGBoost was put forward. As regards the extraction of probabilities, the average of the exported probabilities for the three classifiers was considered as the best option.

The experimental results in terms of classification accuracy for each base learner are organized in Table 1, Table 2, Table 3, Table 4 and Table 5 and supplementary material Tables S1–S4, categorized according to the four label ratios (10%, 20%, 30%, 40%) for each learning method. The bold values in the tables indicate the highest accuracy for the corresponding dataset and the subject labeled ratio.

Table 1. Classification accuracies of bagging-decision trees (BagDT) on four different ratios.

		R = 10%				R = 20%				R = 30%				R = 40%
	Method	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination
Dataset
anneal		88.866	88.524	93.211	94.660	95.210	95.988	96.770	99.107	96.659	96.547	97.551	98.663	97.773	97.215	98.326	98.885
arrhythmia		58.425	57.981	62.614	63.285	65.063	63.314	68.807	69.691	70.140	70.589	69.483	70.812	70.348	69.498	73.242	73.464
audiology		52.253	50.474	56.166	56.640	61.087	61.482	65.593	68.636	65.138	64.684	73.004	74.308	71.660	70.791	75.178	80.079
autos		45.381	44.286	50.167	48.286	52.119	56.167	61.000	55.619	60.976	58.143	63.405	66.381	65.405	59.048	69.738	72.690
balance-scale		73.103	75.832	74.560	75.996	76.487	75.540	80.476	81.608	81.129	78.571	81.941	81.764	83.372	78.093	83.518	83.372
breast-cancer		69.926	70.628	70.603	69.544	69.963	70.616	70.603	70.961	70.961	70.616	70.591	71.675	71.293	70.616	70.948	74.803
bridges-version1		45.727	45.727	45.727	45.727	53.273	57.909	57.273	56.909	60.000	60.727	60.000	62.455	61.000	61.091	64.909	62.909
bridges-version2		43.000	43.000	43.000	43.000	48.455	51.091	59.818	53.909	60.000	64.000	63.091	62.818	62.091	59.182	63.636	62.000
clevalend		78.215	76.559	75.849	75.559	73.247	74.570	79.204	78.194	79.516	76.839	80.462	81.484	83.151	78.516	80.495	81.806
cmc		48.066	48.329	49.356	49.691	50.302	51.455	54.307	51.260	50.576	52.339	51.527	53.154	53.220	54.100	52.674	53.428
column_2C		76.452	75.806	75.806	80.645	80.323	80.323	80.323	84.516	81.613	80.968	82.258	82.903	82.258	82.258	82.903	83.226
column_3C		79.032	79.032	77.419	77.742	79.355	78.710	79.677	83.226	80.323	80.645	78.065	83.871	79.355	79.032	82.258	82.903
credit-rating		84.783	83.768	85.217	85.362	84.638	84.783	85.217	85.507	85.652	85.507	85.652	86.667	85.072	85.942	86.087	86.232
cylinder-bands		58.333	57.037	57.222	59.444	59.630	60.000	59.259	60.556	58.889	58.704	58.148	59.074	60.556	57.963	58.519	59.259
dermatology		71.089	70.578	82.770	87.447	84.977	83.348	91.036	95.931	91.006	91.029	93.483	95.113	93.461	89.647	92.658	96.742
ecoli		67.282	68.173	76.471	76.194	79.144	76.185	80.936	81.230	80.936	80.963	81.827	83.324	80.348	79.162	83.324	83.316
flags		50.053	48.553	48.526	49.921	51.105	51.079	52.026	52.579	55.658	52.184	51.132	56.763	56.237	52.605	55.289	52.711
german_credit		70.500	69.200	69.300	70.800	70.900	69.200	69.500	71.400	69.500	70.500	74.000	74.000	73.600	71.800	72.500	73.000
glass		50.498	47.684	51.991	53.810	64.524	59.848	57.468	65.952	60.303	60.779	67.727	67.338	69.113	64.372	67.229	69.113
haberman		72.538	71.591	71.559	70.882	72.860	73.204	70.882	73.860	71.871	72.538	72.204	73.194	71.871	71.538	71.839	71.237
heart-statlog		72.963	71.481	74.444	75.185	75.556	77.037	78.148	80.741	77.778	76.667	80.370	84.074	79.630	77.037	82.222	82.963
hepatitis		80.083	81.958	79.375	81.917	78.792	80.083	80.000	81.875	82.542	82.542	79.958	79.417	82.542	80.000	78.000	79.875
horse-colic		78.551	78.544	82.605	84.219	84.775	83.138	82.868	85.308	83.401	82.590	84.219	85.300	83.949	85.030	85.571	85.841
hungarian-heart		81.000	78.632	78.310	82.391	81.023	80.701	78.276	78.966	78.621	78.644	78.977	83.057	76.897	76.885	81.655	79.310
hypothyroid		98.357	98.199	98.808	99.602	99.072	98.887	99.046	99.602	99.099	99.099	99.417	99.549	99.285	99.311	99.443	99.576
ionosphere		75.802	77.802	86.643	84.667	88.040	88.032	88.333	92.317	86.619	86.357	91.183	92.317	90.611	89.746	90.611	92.032
iris		72.667	77.333	84.667	86.667	88.000	89.333	90.667	90.667	93.333	93.333	91.333	92.000	93.333	93.333	93.333	92.667
kr-vs-kp		95.025	95.244	96.340	98.499	97.027	97.310	98.091	99.249	97.998	98.216	98.748	99.343	98.998	98.811	99.218	99.343
labor		66.333	66.333	66.333	66.333	65.667	70.333	70.000	66.333	70.000	73.667	77.000	87.667	77.000	79.000	78.667	79.000
letter		77.955	77.235	81.320	84.490	83.775	82.470	86.570	89.840	86.570	85.770	89.165	92.175	88.335	86.375	90.280	92.620
lymphography		65.476	66.143	68.952	67.619	77.571	72.381	71.571	71.667	72.905	70.857	75.571	79.048	74.286	76.238	77.619	79.667
mushroom		99.163	99.323	99.729	100.000	99.877	99.889	99.914	100.000	99.951	99.902	99.975	100.000	99.963	99.975	99.975	100.000
optdigits		89.021	86.174	90.783	92.829	92.189	90.036	93.043	95.979	92.936	90.587	93.719	95.712	94.288	90.854	94.555	95.498
page-blocks		95.414	95.468	95.980	97.058	96.236	95.926	96.547	97.113	96.620	96.583	97.369	97.241	97.168	97.004	97.150	97.168
pendigits		93.122	92.849	94.796	96.516	95.515	94.860	96.061	97.862	96.179	96.006	96.871	98.335	96.880	96.243	97.362	98.071
pima_diabetes		71.880	73.055	75.135	74.479	75.911	73.970	74.747	73.445	75.270	75.261	73.841	76.307	76.309	76.044	74.498	75.930
postoperative		65.556	65.556	65.556	65.556	67.778	71.111	64.444	66.667	62.222	67.778	66.667	67.778	64.444	67.778	65.556	67.778
primary-tumor		29.474	30.339	35.954	34.795	35.651	34.750	38.610	36.248	38.610	34.198	37.745	37.469	40.401	35.677	41.578	38.930
segment		90.736	90.779	91.472	94.329	93.420	93.117	94.372	97.229	93.853	94.502	95.455	97.359	94.589	94.892	96.017	97.186
sick		97.640	97.587	97.932	98.648	98.038	98.118	98.117	98.754	98.118	98.144	98.223	98.728	98.277	98.144	98.595	98.621
solar-flare		67.914	64.807	68.439	68.164	69.790	68.941	70.081	70.322	70.115	70.538	71.532	71.260	70.336	71.536	71.448	71.905
sonar		59.524	60.976	64.881	60.548	66.833	64.429	68.714	67.881	69.214	70.595	74.048	71.119	70.714	69.214	76.476	75.024
soybean		64.241	61.624	75.980	70.119	79.211	75.835	84.633	86.091	84.486	82.564	87.551	90.635	86.520	84.318	91.355	93.116
spambase		89.937	90.285	90.545	92.827	91.328	91.784	92.349	94.305	92.132	92.827	92.805	94.631	92.741	92.067	93.219	94.610
spect		61.316	61.494	61.864	63.367	68.506	65.344	66.099	66.902	66.717	71.978	73.390	80.342	79.186	74.895	77.025	76.084
sponge		86.250	86.250	86.250	86.250	91.071	92.500	92.500	92.500	92.500	92.500	92.500	92.500	93.750	92.500	92.500	93.750
tae		32.375	34.375	40.417	38.375	38.333	36.958	38.333	41.000	42.958	41.708	40.333	52.250	44.375	42.333	53.583	54.875
tic-tac-toe		70.254	70.985	74.221	76.620	78.189	74.326	80.895	83.091	82.467	76.404	84.864	86.951	84.444	80.684	89.457	91.864
vehicle		64.189	62.289	67.150	67.615	66.916	67.134	70.445	69.618	70.221	70.091	73.175	71.161	73.060	70.454	72.108	72.696
vote		94.049	94.952	94.276	95.174	95.412	95.418	94.049	96.321	95.412	95.640	96.327	95.872	95.640	95.645	96.781	96.327
vowel		48.283	49.192	52.424	54.040	60.808	60.707	67.980	70.202	70.707	65.758	74.545	77.172	73.939	68.485	80.303	83.737
waveform		78.020	75.600	79.180	79.700	80.300	77.540	79.160	81.340	79.160	78.380	81.200	81.260	80.700	78.160	80.880	81.740
wine		74.314	78.203	80.425	84.935	88.758	85.425	91.601	93.268	90.458	89.379	93.301	96.078	92.157	91.601	92.712	94.379
wisconsin-breast		91.986	92.418	92.416	95.422	94.559	93.704	94.994	96.422	95.565	94.277	95.565	95.994	94.996	95.277	95.137	96.137
zoo		57.455	57.455	57.455	57.455	75.273	75.273	78.273	85.182	81.273	82.273	86.273	88.273	84.273	86.273	88.273	93.182
Average		71.270	71.158	73.611	74.383	76.216	75.847	77.631	78.817	78.125	77.863	79.614	81.348	79.913	78.623	81.062	81.867

Table 2. Classification accuracies of random forests (RF) on four different ratios.

		R = 10%				R = 20%				R = 30%				R = 40%
	Method	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination
Dataset
anneal		83.740	83.965	85.414	87.976	87.418	87.194	90.979	92.203	90.869	90.087	92.985	93.762	92.875	92.427	93.871	94.871
arrhythmia		56.638	56.633	60.406	60.623	59.720	58.841	60.396	64.599	59.285	58.184	63.275	64.821	62.377	59.517	63.937	65.488
audiology		42.490	36.739	53.518	45.652	56.700	54.466	63.794	59.308	61.976	63.241	72.134	68.617	68.123	67.767	73.874	73.379
autos		48.833	49.714	52.119	51.238	56.929	57.452	65.381	66.286	68.286	69.262	70.667	74.595	68.738	69.738	77.976	79.476
balance-scale		78.067	77.588	79.027	80.461	79.519	78.241	79.841	80.161	82.084	79.524	82.087	81.935	82.081	80.806	81.910	81.129
breast-cancer		69.249	70.283	67.131	72.044	67.180	68.559	64.310	68.276	66.773	68.153	67.192	66.441	67.106	69.212	64.754	68.596
bridges-version1		44.636	44.636	44.636	44.636	44.636	38.000	46.545	44.727	46.636	44.545	52.545	45.545	51.545	45.545	52.273	55.000
bridges-version2		41.818	41.818	41.818	41.818	45.545	37.000	49.273	46.455	48.364	40.000	53.182	45.545	53.273	47.545	53.182	52.273
clevalend		76.849	75.172	79.505	78.817	80.161	79.495	83.430	81.108	81.086	79.774	82.806	81.774	82.452	82.108	83.774	81.806
cmc		48.881	48.406	50.102	48.401	49.964	51.322	51.798	50.239	51.525	51.730	50.914	52.544	51.116	52.000	50.908	52.200
column_2C		77.097	76.774	77.419	79.355	80.000	79.677	80.968	81.290	81.935	81.935	83.548	83.871	83.871	81.613	82.581	85.484
column_3C		77.742	79.032	79.677	82.258	82.903	82.258	84.516	82.903	82.581	83.226	81.290	84.839	81.935	82.258	81.290	84.194
credit-rating		82.319	82.464	83.478	85.217	83.768	84.203	84.493	85.797	83.768	84.203	84.928	85.362	84.638	84.348	84.783	86.087
cylinder-bands		59.259	58.519	63.333	64.259	65.185	59.444	67.778	65.741	67.593	60.185	70.185	70.000	67.222	60.000	71.296	70.185
dermatology		78.709	79.797	86.877	92.598	90.961	89.595	94.805	95.616	93.716	93.724	95.090	96.456	95.638	94.264	93.986	96.179
ecoli		74.091	72.906	77.674	78.574	80.339	77.968	84.189	84.777	84.189	80.651	84.225	85.695	83.307	84.216	85.410	86.292
flags		46.447	41.789	44.395	48.421	52.763	38.711	52.711	51.053	54.737	45.921	50.132	52.184	54.289	47.000	57.789	52.132
german_credit		72.900	72.200	73.300	72.700	74.000	73.600	73.900	75.600	73.900	73.700	75.300	75.300	74.900	71.600	74.500	75.400
glass		55.216	53.333	57.987	55.216	66.385	66.840	68.701	68.247	68.723	69.156	71.948	75.714	72.381	67.706	74.762	74.286
haberman		62.043	65.656	66.344	64.258	68.591	69.903	67.269	68.581	66.978	67.634	66.355	69.237	67.022	68.989	66.301	66.677
heart-statlog		74.074	75.185	74.815	77.037	78.519	78.519	78.148	81.111	79.259	78.519	81.481	83.333	80.741	78.889	82.222	81.852
hepatitis		80.667	80.708	80.125	83.792	78.167	79.458	81.250	83.125	82.000	82.542	78.667	84.458	81.917	81.875	78.667	85.125
horse-colic		79.069	79.595	79.857	82.590	84.227	83.664	83.949	84.767	83.686	82.057	84.219	85.578	83.941	85.856	84.767	86.119
hungarian-heart		79.276	79.264	79.989	82.000	83.046	83.046	81.000	82.011	81.345	81.345	80.287	81.000	81.644	81.655	81.299	80.966
hypothyroid		95.733	95.520	96.660	99.258	97.880	97.429	98.569	99.444	98.330	98.039	98.966	99.391	98.993	98.596	98.914	99.364
ionosphere		81.802	80.381	86.929	88.603	90.603	89.746	90.889	92.889	91.460	90.603	92.889	93.460	91.746	91.746	93.175	94.032
iris		83.333	83.333	88.000	90.000	93.333	92.000	96.000	94.667	96.000	95.333	96.000	94.667	95.333	95.333	94.667	95.333
kr-vs-kp		95.776	94.900	96.120	98.623	96.778	96.621	97.497	99.280	97.559	97.465	98.248	99.406	98.154	98.216	98.937	99.312
labor		65.000	65.000	65.000	65.000	73.667	69.667	82.000	80.333	82.000	77.000	84.000	82.667	84.000	73.333	82.667	82.333
letter		84.795	84.210	87.940	90.915	89.845	89.625	92.215	95.300	92.215	91.870	94.050	96.290	93.415	92.955	95.010	96.435
lymphography		68.238	67.571	73.619	75.000	74.190	76.286	77.619	76.238	74.333	76.333	81.667	80.333	79.048	77.619	85.143	85.095
mushroom		99.741	99.766	99.914	100.000	99.963	99.963	100.000	100.000	99.988	99.988	99.975	100.000	100.000	99.988	100.000	100.000
optdigits		94.573	94.431	95.765	97.473	96.406	96.050	97.011	98.327	97.135	96.762	97.438	98.327	97.438	96.886	97.865	98.363
page-blocks		95.870	95.797	96.218	97.351	96.492	96.455	96.675	97.497	96.748	96.766	97.004	97.552	96.985	96.912	97.479	97.533
pendigits		97.143	96.980	97.607	98.463	98.008	97.844	98.317	99.236	98.372	98.353	98.772	99.245	98.672	98.435	98.917	99.172
pima_diabetes		74.354	72.404	74.231	74.612	75.656	74.614	75.921	76.304	77.092	76.832	75.930	74.879	75.531	75.138	74.624	76.048
postoperative		68.889	68.889	68.889	68.889	62.222	66.667	62.222	67.778	63.333	64.444	64.444	65.556	60.000	65.556	63.333	63.333
primary-tumor		33.601	34.768	37.424	33.592	37.406	38.592	43.039	40.383	43.039	40.348	42.736	43.913	41.551	40.642	43.627	43.039
segment		93.333	93.463	94.372	96.537	95.281	95.238	95.844	97.965	95.628	95.671	96.970	98.009	96.537	96.364	97.403	98.225
sick		96.527	96.394	96.659	98.595	97.455	97.455	97.773	98.542	97.667	97.534	97.958	98.462	97.905	97.826	98.144	98.436
solar-flare		66.769	65.990	69.098	69.375	70.512	70.059	70.131	70.464	70.073	70.037	69.758	71.550	70.041	69.681	70.513	72.590
sonar		62.024	64.024	67.333	69.643	69.238	69.738	75.929	79.333	74.976	75.976	76.929	81.238	79.786	75.952	82.238	82.690
soybean		67.199	64.429	76.281	75.251	79.066	77.006	85.211	87.990	86.816	84.474	89.744	92.822	88.286	88.572	91.360	93.110
spambase		92.871	92.566	93.088	94.588	93.523	93.305	94.197	95.544	94.305	94.110	94.675	95.675	94.523	94.327	94.936	95.588
spect		69.270	69.825	68.522	69.270	72.395	72.133	71.284	78.837	75.582	71.948	78.159	79.193	80.196	77.765	79.895	77.381
sponge		92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	93.750	93.750	92.500	93.750
tae		36.375	37.750	41.750	41.667	39.708	37.042	41.667	43.542	40.375	39.083	49.042	49.625	46.333	45.042	56.333	57.667
tic-tac-toe		76.512	74.942	78.396	83.296	81.735	80.171	84.343	91.544	84.240	81.110	88.515	95.510	89.764	82.156	91.549	96.658
vehicle		65.132	64.311	70.339	69.151	71.761	70.104	70.227	73.167	70.340	70.697	74.115	73.527	71.992	71.169	74.591	73.161
vote		94.503	94.276	94.958	96.327	95.196	94.503	94.947	96.332	95.412	95.180	96.786	96.332	96.559	96.327	97.014	96.781
vowel		34.444	33.737	44.848	42.727	56.566	55.657	71.313	71.111	71.010	69.899	84.545	88.788	80.808	80.000	91.616	96.162
waveform		83.660	82.940	84.340	83.920	84.300	84.140	84.140	84.620	84.140	83.860	84.440	84.840	85.100	83.960	85.220	84.840
wine		86.046	86.601	85.980	94.412	94.379	96.601	95.000	98.301	96.111	96.667	97.222	98.333	98.333	97.222	97.222	98.301
wisconsin-breast		94.414	94.986	95.130	97.280	96.135	95.418	96.420	96.851	96.706	96.565	96.280	96.565	95.994	96.280	96.280	96.422
zoo		75.273	75.273	75.273	75.273	79.273	77.273	79.273	88.182	85.273	79.273	87.273	93.182	88.273	84.273	87.273	92.182
Average		73.015	72.730	75.130	76.137	77.238	76.316	79.047	80.118	79.274	78.255	80.954	81.826	80.694	79.436	81.901	82.701

Table 3. Classification accuracies of rotation forest (RotF) on four different ratios.

		R = 10%				R = 20%				R = 30%				R = 40%
	Method	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination
Dataset
anneal		85.185	86.075	86.859	91.080	88.206	91.541	93.211	94.206	93.432	93.543	95.104	96.215	94.878	94.988	96.105	96.880
arrhythmia		59.271	64.377	61.512	66.357	67.937	67.705	67.266	70.130	69.266	68.150	69.053	71.242	68.159	72.353	72.135	71.469
audiology		51.759	51.759	57.589	58.379	65.929	63.775	67.312	70.810	65.949	67.292	72.095	74.783	70.296	70.771	76.482	79.644
autos		47.333	45.405	54.143	52.262	53.595	59.405	66.333	60.857	66.405	68.214	69.690	69.762	68.738	72.643	75.595	72.643
balance-scale		83.669	83.513	85.261	84.470	84.956	85.750	86.892	86.879	87.837	87.046	87.673	88.966	88.490	89.601	89.913	91.027
breast-cancer		64.335	66.096	69.224	67.180	69.286	70.320	69.963	70.345	73.042	72.709	73.067	73.448	69.224	69.594	69.544	72.007
bridges-version1		50.455	50.455	50.455	50.455	56.091	53.182	58.091	51.091	59.273	65.636	60.182	63.818	58.182	64.727	67.727	60.091
bridges-version2		48.455	48.455	48.455	48.455	54.273	51.273	62.909	57.818	54.727	58.273	58.182	61.818	57.091	58.364	64.909	66.818
clevalend		74.172	77.527	76.892	79.462	81.817	79.860	82.161	81.151	81.151	81.806	82.129	82.194	83.462	82.473	82.817	84.441
cmc		49.764	49.968	49.492	49.899	49.492	52.472	51.048	51.661	51.388	53.764	54.780	52.064	52.677	54.035	52.882	53.899
column_2C		80.645	80.645	78.710	82.258	79.032	80.323	83.548	83.226	81.613	80.645	81.613	82.903	82.258	82.581	83.226	84.516
column_3C		80.000	74.839	80.000	82.258	79.032	80.645	81.613	86.452	80.323	80.000	81.613	86.774	83.871	84.839	84.194	83.871
credit-rating		83.478	83.188	84.493	85.072	85.072	85.072	84.638	85.507	85.507	85.797	86.522	86.232	85.072	85.652	87.101	86.957
cylinder-bands		60.000	61.667	63.519	63.889	64.444	63.519	69.815	68.333	70.000	69.259	69.815	75.000	70.556	68.148	76.852	75.185
dermatology		85.008	84.452	92.342	95.901	95.631	94.272	95.368	97.290	94.827	94.264	96.186	98.101	96.742	96.734	97.830	97.553
ecoli		75.303	73.529	75.276	83.307	80.633	79.750	83.913	86.007	83.913	83.351	85.071	86.292	83.601	84.198	85.107	87.460
flags		50.132	52.158	50.184	51.579	50.658	51.605	55.289	53.263	58.316	54.789	55.763	57.737	56.316	53.763	57.263	60.026
german_credit		68.500	71.500	70.900	72.300	73.200	73.500	73.200	73.300	73.200	74.100	74.000	73.700	72.900	73.900	74.100	76.000
glass		51.407	54.610	57.511	56.039	63.528	63.528	63.506	70.563	61.710	63.203	67.792	69.675	64.004	64.978	70.519	70.498
haberman		69.258	69.935	71.860	71.527	72.505	73.161	71.559	74.849	72.860	72.516	74.516	73.538	73.849	73.839	70.258	72.860
heart-statlog		76.296	72.593	79.630	77.407	80.000	78.148	79.630	80.370	79.259	80.000	82.222	80.741	79.259	78.889	82.222	80.741
hepatitis		82.542	79.833	80.625	77.917	82.625	80.708	82.500	82.583	82.500	84.500	80.583	82.542	81.167	82.500	80.667	87.042
horse-colic		77.155	77.147	78.544	79.324	82.628	83.431	82.583	83.393	82.853	83.408	80.435	83.964	83.408	84.227	83.979	85.045
hungarian-heart		81.310	82.333	82.667	82.632	83.345	83.368	79.632	80.989	82.034	80.023	81.644	82.345	81.310	82.356	82.011	81.333
hypothyroid		96.873	97.270	97.615	99.496	98.728	98.648	98.728	99.603	98.807	98.304	98.781	99.364	99.073	98.781	99.126	99.417
ionosphere		80.095	82.087	89.468	90.357	89.190	90.881	90.889	94.325	91.468	92.032	91.484	94.603	91.460	92.024	92.897	94.032
iris		85.333	85.333	94.000	88.000	92.667	93.333	94.000	94.667	96.667	95.333	94.667	96.000	97.333	95.333	96.000	96.000
kr-vs-kp		94.932	94.647	96.840	98.499	97.090	96.933	98.154	99.156	97.467	97.245	98.717	99.031	97.248	97.935	98.593	99.343
labor		72.667	72.667	72.667	72.667	73.667	66.333	82.333	80.667	82.333	77.000	78.333	85.667	78.333	78.667	84.333	90.000
letter		81.565	81.565	84.680	88.220	87.230	87.190	89.825	93.270	89.825	89.410	92.015	94.575	91.635	90.890	93.400	95.260
lymphography		69.619	71.714	71.714	73.048	74.905	78.381	74.238	79.000	71.619	78.333	80.333	77.143	82.381	77.619	81.000	83.143
mushroom		99.643	99.643	99.889	100.000	99.914	99.914	99.926	100.000	99.951	99.926	99.951	100.000	99.951	99.938	99.963	100.000
optdigits		92.740	92.544	94.644	95.819	94.911	95.071	95.463	97.189	95.925	95.463	96.495	97.740	96.335	96.139	97.153	97.473
page-blocks		95.980	95.432	96.072	97.296	96.528	96.254	96.766	97.460	96.583	96.711	97.040	97.606	96.930	97.058	97.205	97.552
pendigits		96.889	96.934	97.626	98.699	97.771	97.926	98.535	99.045	98.590	98.399	98.826	99.118	98.754	98.672	98.917	99.118
pima_diabetes		72.011	72.927	74.219	74.352	74.222	76.300	77.218	76.174	76.174	74.614	75.005	76.304	75.138	74.610	75.140	75.781
postoperative		64.444	64.444	64.444	64.444	70.000	66.667	64.444	66.667	70.000	70.000	66.667	68.889	64.444	67.778	66.667	70.000
primary-tumor		32.701	30.945	36.881	38.012	37.415	38.333	41.854	42.709	41.854	40.963	39.528	42.442	43.681	42.478	43.351	42.166
segment		93.463	93.853	94.545	95.411	95.411	94.459	95.887	97.965	96.190	96.104	97.186	97.965	96.494	96.450	97.143	98.139
sick		97.720	97.190	97.905	98.621	98.091	98.038	98.144	98.860	98.224	97.985	98.330	98.913	98.383	98.250	98.701	99.046
solar-flare		66.657	68.468	70.017	71.152	70.766	70.821	70.541	71.115	70.724	70.360	70.817	72.454	70.328	71.674	71.551	73.115
sonar		62.429	64.952	67.286	64.429	67.405	67.286	70.143	75.429	73.548	73.571	79.786	73.071	76.929	79.286	81.333	79.310
soybean		72.594	74.205	81.831	83.282	86.087	85.789	89.156	92.093	88.574	88.568	92.822	93.849	90.774	91.213	93.129	94.286
spambase		91.632	91.828	91.958	93.740	92.806	93.306	93.414	95.088	93.306	93.892	94.088	95.305	93.784	93.697	94.762	95.196
spect		67.219	67.603	68.166	68.137	72.349	72.025	73.351	82.388	74.594	73.515	77.094	74.402	74.827	79.224	77.171	77.828
sponge		92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	91.071	92.500	93.750	92.500
tae		33.708	39.042	44.375	43.042	41.750	37.708	46.333	41.000	44.333	48.333	50.333	56.333	47.583	50.250	56.875	56.875
tic-tac-toe		74.637	74.536	76.724	80.484	80.586	81.109	83.813	89.148	84.766	83.409	89.038	94.677	89.878	88.315	93.940	96.453
vehicle		69.979	66.085	71.535	71.625	73.889	71.406	75.892	76.714	73.641	72.580	74.711	77.076	75.416	72.464	75.305	76.134
vote		91.723	91.047	93.108	94.937	95.645	94.271	94.963	95.180	94.244	95.190	95.634	96.327	96.327	94.952	96.559	96.559
vowel		45.152	45.657	56.667	57.980	67.374	62.222	74.646	78.384	77.374	73.636	84.747	88.889	84.141	82.121	91.313	95.253
waveform		81.320	81.820	82.480	82.200	82.180	83.140	82.700	83.340	82.700	83.820	83.160	82.940	83.320	84.080	83.180	83.880
wine		87.680	86.536	87.157	96.078	90.425	92.157	95.556	94.967	93.333	93.856	97.190	96.634	94.412	96.078	96.078	96.634
wisconsin-breast		95.418	95.277	95.994	97.137	96.277	96.422	96.565	96.994	96.849	96.708	96.708	96.851	96.563	96.565	96.708	97.280
zoo		70.455	70.455	70.455	70.455	81.273	78.273	81.364	87.273	83.273	83.273	88.273	93.091	88.273	89.273	89.273	92.182
Average		73.913	74.205	76.356	77.264	78.418	78.171	80.169	81.263	80.306	80.424	81.636	82.975	81.213	81.645	83.163	83.963

Table 4. Classification accuracies of XGBoost on four different ratios.

		R = 10%				R = 20%				R = 30%				R = 40%
	Method	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination
Dataset
anneal		92.759	92.648	95.100	96.994	96.323	96.101	96.881	98.885	96.993	96.881	97.437	98.884	97.438	97.105	97.772	98.884
arrhythmia		55.531	56.420	63.092	61.947	63.285	64.831	65.498	68.155	65.498	68.372	68.174	70.372	67.493	67.053	70.357	72.150
audiology		48.261	45.217	52.668	50.119	57.134	58.024	62.905	61.107	63.360	65.534	68.103	69.032	66.364	67.668	72.964	76.976
autos		46.333	44.857	51.143	50.738	58.952	55.548	65.357	65.310	64.476	64.429	70.643	77.571	72.690	72.167	77.952	78.976
balance-scale		76.792	75.202	77.747	79.813	80.632	80.952	82.396	81.126	82.568	82.401	83.689	83.039	83.039	83.835	84.944	83.041
breast-cancer		65.788	64.704	66.084	67.488	67.167	66.810	66.392	67.857	65.000	67.475	66.392	68.140	66.047	68.842	65.382	71.613
bridges-version1		42.818	42.818	42.818	42.818	44.818	47.636	54.455	49.545	60.909	57.182	54.364	59.000	56.364	54.273	57.000	60.909
bridges-version2		43.545	43.545	43.545	43.545	42.727	43.727	55.273	54.091	60.182	57.273	56.182	58.182	57.182	56.273	58.000	67.909
clevalend		71.172	72.516	75.849	73.796	79.548	80.860	80.785	80.505	79.484	80.473	84.430	80.473	82.796	81.140	82.441	81.785
cmc		49.220	49.488	50.776	50.576	51.253	52.951	53.358	54.312	53.563	54.987	54.442	55.052	53.966	54.848	54.304	54.852
column_2C		76.129	76.452	76.129	80.645	81.613	80.968	80.323	83.548	79.677	80.968	81.613	81.935	85.806	80.000	83.226	83.226
column_3C		79.032	77.419	76.129	78.065	81.290	81.290	80.645	81.613	80.645	80.323	79.355	82.581	80.323	82.903	80.968	84.839
credit-rating		82.174	82.029	83.478	83.478	83.188	84.058	83.913	85.652	84.203	84.928	84.638	86.812	84.783	85.072	86.522	86.522
cylinder-bands		65.926	63.333	70.370	69.444	71.667	71.667	71.852	74.630	75.741	74.815	77.963	78.148	75.000	74.444	79.259	80.000
dermatology		83.889	85.000	89.872	92.342	93.183	92.102	94.820	96.749	94.820	94.002	95.375	96.194	94.827	94.550	95.375	97.020
ecoli		68.146	67.273	70.865	76.203	75.000	75.000	81.836	82.121	81.836	80.339	81.827	84.804	80.321	81.221	82.727	84.528
flags		49.132	48.632	48.579	51.053	48.632	50.711	52.237	54.711	53.237	51.632	53.632	49.000	53.658	50.579	59.737	56.289
german_credit		68.900	68.800	70.100	70.700	72.500	71.500	73.500	73.800	73.500	71.900	72.800	73.900	73.200	73.800	74.100	75.100
glass		46.818	46.364	54.264	55.130	62.121	60.714	64.978	63.117	64.545	65.952	66.342	67.424	67.359	67.338	69.524	72.381
haberman		66.591	67.269	69.290	71.903	69.247	69.570	67.645	68.914	68.290	67.602	67.333	70.280	70.925	69.935	66.656	67.312
heart-statlog		71.852	72.963	74.444	74.074	75.926	74.074	75.556	75.185	75.926	76.296	75.926	80.370	74.815	74.444	80.370	78.519
hepatitis		74.042	73.417	76.000	77.333	74.792	75.333	79.292	77.333	80.542	80.583	77.250	80.000	77.958	78.000	77.208	81.167
horse-colic		75.788	75.556	80.165	79.092	79.617	80.721	81.809	82.875	79.610	80.691	81.239	84.767	80.976	80.698	81.246	85.586
hungarian-heart		80.989	79.644	79.310	81.977	80.667	81.690	79.966	78.563	78.931	78.931	76.862	78.609	79.943	79.230	79.943	80.264
hypothyroid		98.304	98.304	98.436	99.470	98.781	98.754	99.046	99.682	99.046	99.046	99.311	99.682	99.284	99.205	99.523	99.655
ionosphere		80.659	80.659	84.635	84.937	86.643	86.365	88.603	90.333	87.183	85.190	88.611	92.032	91.175	88.317	90.897	92.032
iris		84.667	82.000	86.667	91.333	92.000	91.333	94.667	93.333	94.667	93.333	94.667	95.333	94.667	94.667	96.000	96.000
kr-vs-kp		96.339	96.151	97.059	98.561	97.685	97.434	98.279	99.437	98.310	98.216	98.873	99.500	98.811	98.780	99.062	99.343
labor		65.000	65.000	65.000	65.000	64.667	72.000	73.667	59.333	73.667	78.667	79.000	78.667	79.000	82.667	80.333	82.000
letter		82.325	82.095	84.905	87.820	86.930	86.430	89.330	92.050	89.330	88.795	90.910	93.235	90.585	89.990	91.760	93.425
lymphography		70.238	69.571	74.286	66.952	76.333	75.000	76.286	77.000	77.762	76.333	78.333	82.524	79.667	77.667	83.095	85.190
mushroom		99.729	99.717	99.877	100.000	99.914	99.914	99.914	100.000	99.914	99.963	99.963	100.000	99.951	100.000	99.963	100.000
optdigits		91.299	90.694	93.345	95.071	94.537	94.431	95.196	97.349	95.214	95.107	96.068	97.331	96.050	95.516	96.370	97.331
page-blocks		95.834	95.889	96.199	97.296	96.382	96.437	96.602	97.387	96.583	96.583	97.186	97.223	96.857	96.985	97.241	97.332
pendigits		94.642	94.214	90.897	97.844	96.761	96.543	97.398	98.854	97.380	97.416	97.971	98.890	97.726	97.689	98.308	98.836
pima_diabetes		72.138	70.976	72.143	74.096	74.482	75.784	74.229	75.511	74.626	73.959	73.717	74.098	74.621	74.489	72.931	75.665
postoperative		60.000	60.000	60.000	60.000	58.889	60.000	60.000	67.778	58.889	58.889	61.111	61.111	57.778	57.778	57.778	55.556
primary-tumor		34.777	37.736	40.677	39.519	38.592	40.963	43.619	40.695	43.619	44.795	43.324	46.310	42.754	43.048	44.822	43.057
segment		92.857	91.948	93.680	95.844	94.719	94.156	95.714	98.052	95.671	95.584	96.623	98.398	96.364	96.190	97.619	98.312
sick		97.614	97.640	97.799	98.860	98.171	98.250	98.356	99.046	98.356	98.330	98.356	99.046	98.303	98.250	98.781	99.072
solar-flare		68.911	68.786	70.739	70.238	69.480	69.089	70.392	70.956	70.408	70.345	70.122	72.730	70.513	72.118	72.127	72.945
sonar		58.143	59.071	63.833	62.976	68.714	65.786	71.619	69.619	75.405	71.119	77.333	74.952	74.929	74.857	77.381	80.667
soybean		72.598	72.157	81.249	79.776	85.200	84.327	89.009	90.473	89.450	87.835	90.190	93.847	90.040	90.332	92.971	93.252
spambase		91.480	90.936	92.480	93.827	93.284	92.610	93.284	94.870	93.588	93.306	94.392	95.153	94.371	93.870	94.653	94.957
spect		62.936	63.129	63.478	65.166	66.863	69.601	67.959	70.157	70.467	69.464	73.853	73.722	70.027	72.094	76.070	76.400
sponge		92.500	92.500	92.500	92.500	92.500	92.500	91.071	92.500	93.750	93.750	93.750	92.321	93.750	92.321	93.750	93.571
tae		34.417	31.750	37.667	37.083	34.958	39.000	39.083	48.292	41.042	40.333	49.667	42.292	46.333	46.333	57.625	53.000
tic-tac-toe		78.195	78.922	81.427	87.067	88.207	86.851	91.132	95.827	91.757	91.237	95.095	97.705	95.616	94.991	96.346	98.224
vehicle		61.941	63.489	68.331	68.922	67.265	67.034	69.634	72.224	70.339	70.571	73.646	73.637	71.408	70.455	74.933	75.312
vote		95.185	95.185	95.412	96.105	95.640	94.958	95.877	95.640	95.185	95.645	96.094	95.640	95.407	95.412	95.645	95.640
vowel		49.293	50.707	57.273	56.667	63.333	62.323	72.222	74.848	71.818	70.606	78.283	82.525	77.273	75.455	83.939	88.384
waveform		81.780	80.480	82.160	83.160	82.340	82.220	83.160	83.780	83.160	83.440	83.620	84.380	84.280	83.960	84.300	84.580
wine		83.203	83.758	85.980	87.647	92.124	92.092	93.856	94.967	93.268	93.824	94.412	94.412	94.412	93.235	94.935	96.634
wisconsin-breast		93.561	94.275	94.135	96.420	94.418	94.133	95.137	95.994	95.137	94.849	95.280	95.851	95.422	94.708	94.994	95.851
zoo		60.545	60.545	60.545	60.545	79.273	80.273	83.273	84.091	87.273	87.273	90.273	93.091	90.273	89.273	89.273	96.000
Average		72.413	72.179	74.557	75.454	76.734	76.971	78.896	79.632	79.378	79.232	80.474	81.640	80.380	80.110	81.844	83.056

Table 5. Classification accuracies of voting (RF, RotF, XGBoost) on four different ratios.

		R = 10%				R = 20%				R = 30%				R = 40%
	Method	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination	Supervised	Semi-Supervised	Active Random	Combination
Dataset
anneal		90.864	89.634	92.869	94.648	94.207	94.096	96.768	98.658	96.323	96.660	97.552	99.107	97.552	97.663	98.218	98.886
arrhythmia		60.836	58.865	63.734	65.729	67.266	66.169	66.604	68.362	67.715	67.271	70.821	71.705	69.478	67.261	72.575	73.681
audiology		51.759	50.000	58.419	57.036	59.842	62.451	67.273	66.383	66.818	68.123	72.964	73.458	68.577	71.166	76.462	78.241
autos		49.262	45.905	51.667	55.619	59.905	59.357	67.810	65.786	68.357	67.810	74.548	78.095	73.143	73.143	78.452	79.857
balance-scale		80.947	79.355	81.423	83.187	83.034	83.510	85.289	84.002	84.808	84.647	85.123	84.813	86.078	86.400	87.355	87.353
breast-cancer		66.810	68.214	66.798	69.273	67.512	67.894	68.165	72.759	69.187	69.557	67.475	71.342	68.498	68.867	67.118	73.054
bridges-version1		46.545	46.545	46.545	46.545	53.273	49.455	56.273	57.182	62.909	59.909	60.182	65.727	60.182	65.727	65.727	67.545
bridges-version2		47.545	47.545	47.545	47.545	47.545	51.091	59.818	54.455	59.182	61.909	62.091	65.909	61.091	64.818	67.545	63.909
clevalend		72.495	74.828	78.172	79.817	82.183	79.871	81.430	84.140	81.441	81.785	83.774	82.796	83.774	83.441	84.452	82.462
cmc		50.376	50.645	51.187	51.732	51.731	53.292	52.883	52.753	52.475	55.661	53.967	54.238	52.679	55.933	53.422	53.972
column_2C		78.387	78.065	78.710	80.000	80.645	81.290	83.226	85.484	82.258	82.258	83.226	83.548	85.161	83.226	86.129	85.484
column_3C		79.677	78.710	79.032	82.258	81.613	81.290	83.871	86.129	83.548	82.581	82.258	83.871	82.581	83.226	82.258	83.871
credit-rating		83.768	84.493	84.493	85.797	85.072	83.768	85.362	86.232	85.652	85.507	85.942	86.232	86.522	86.232	86.812	87.536
cylinder-bands		67.407	66.852	70.556	70.000	72.222	70.741	75.556	73.333	75.926	70.926	78.148	78.519	76.481	72.963	81.111	82.037
dermatology		86.089	86.622	93.431	95.616	95.901	95.623	96.456	97.020	96.456	97.005	97.568	97.553	96.742	97.275	96.734	97.568
ecoli		74.706	73.779	76.488	80.945	79.135	78.547	83.021	84.822	83.021	81.818	85.704	85.971	84.207	83.030	85.722	87.469
flags		53.211	47.579	53.263	51.053	55.921	52.184	55.316	56.237	59.395	56.289	56.289	57.789	55.237	56.789	60.816	58.921
german_credit		70.900	70.700	71.600	73.900	74.300	74.200	74.800	75.500	74.800	75.400	74.900	74.800	75.700	73.900	75.500	76.300
glass		54.719	55.173	60.390	61.710	66.364	65.909	67.792	69.221	66.840	68.247	71.970	74.762	70.519	70.022	75.173	76.212
haberman		70.591	70.258	67.624	69.925	70.237	70.559	69.280	73.151	70.581	68.946	71.570	71.237	71.269	70.914	70.892	70.570
heart-statlog		72.963	74.074	76.296	77.778	78.889	76.667	77.778	80.741	77.037	77.778	79.630	82.222	79.630	79.259	82.963	81.111
hepatitis		79.250	81.833	76.708	81.875	78.792	80.667	80.000	83.792	81.875	81.875	79.875	82.500	77.958	81.833	80.500	82.542
horse-colic		77.995	78.799	81.802	82.883	81.794	83.956	83.694	84.505	83.138	83.949	82.868	86.396	84.767	84.227	84.234	86.404
hungarian-heart		80.655	80.621	80.977	83.391	82.690	83.345	80.632	81.678	82.000	82.356	78.897	79.966	80.632	80.966	82.333	80.644
hypothyroid		98.013	97.827	98.463	99.709	98.940	98.781	99.099	99.709	99.046	99.125	99.284	99.735	99.231	99.258	99.497	99.655
ionosphere		81.516	80.944	88.063	91.206	90.341	89.778	90.317	93.175	90.325	91.175	92.603	92.603	91.460	92.889	92.889	93.746
iris		82.000	82.667	91.333	92.667	92.667	92.667	94.667	95.333	94.667	94.667	95.333	94.667	94.667	94.667	95.333	95.333
kr-vs-kp		96.402	96.308	97.121	98.686	97.747	97.403	98.279	99.468	98.435	98.310	98.967	99.374	98.842	98.779	99.124	99.437
labor		68.000	68.000	68.000	68.000	75.333	73.667	78.667	74.667	78.667	78.667	80.333	80.667	80.333	79.000	82.667	84.667
letter		85.625	85.410	88.460	91.850	90.355	89.950	92.360	95.315	92.360	91.950	93.955	96.255	93.450	93.225	94.875	96.335
lymphography		71.000	71.714	75.667	69.714	76.905	80.333	77.667	77.524	76.381	76.286	81.000	81.714	80.381	80.333	83.762	83.667
mushroom		99.729	99.729	99.914	100.000	99.914	99.926	99.926	100.000	99.951	99.926	99.975	100.000	99.963	99.963	99.975	100.000
optdigits		94.217	93.968	95.534	97.384	96.246	96.157	96.922	98.238	97.064	96.779	97.456	98.274	97.171	97.046	97.562	98.025
page-blocks		95.980	96.071	96.346	97.643	96.602	96.620	96.784	97.570	96.857	96.839	97.333	97.424	97.077	97.186	97.479	97.607
pendigits		97.034	97.052	97.498	98.717	98.008	97.999	98.544	99.327	98.408	98.581	98.790	99.300	98.699	98.581	98.999	99.263
pima_diabetes		72.925	74.238	74.231	76.176	75.137	76.049	76.316	75.658	77.093	75.531	75.540	75.911	75.398	74.879	74.626	75.660
postoperative		67.778	67.778	67.778	67.778	64.444	65.556	64.444	68.889	63.333	65.556	63.333	65.556	62.222	64.444	63.333	64.444
primary-tumor		35.365	36.257	37.727	36.533	39.492	40.695	46.586	43.922	46.586	45.089	43.039	45.989	44.840	42.763	45.989	44.831
segment		93.853	93.853	94.935	96.667	95.498	95.844	96.104	98.485	96.277	96.320	97.186	98.485	96.840	96.883	97.835	98.528
sick		97.826	97.720	97.985	98.595	98.197	98.303	98.356	99.072	98.356	98.276	98.409	99.019	98.383	98.409	98.807	99.019
solar-flare		69.171	68.081	70.975	71.101	70.275	70.458	70.643	72.026	70.477	70.524	70.394	72.741	70.805	70.931	70.971	72.838
sonar		62.452	61.452	66.786	66.357	70.190	68.286	74.024	75.500	73.524	74.452	79.262	76.905	78.786	76.881	78.857	81.214
soybean		75.816	74.938	82.711	85.049	85.931	86.228	90.030	92.530	90.471	89.595	92.234	94.290	91.211	92.238	94.150	94.578
spambase		92.349	92.088	93.241	94.370	93.654	93.545	94.045	95.349	94.175	94.088	94.631	95.544	94.479	94.284	94.979	95.327
spect		63.985	64.911	65.652	66.592	73.421	71.902	69.779	72.203	73.337	72.457	77.842	74.455	75.135	76.763	81.068	77.279
sponge		92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	92.500	93.750	92.500
tae		35.708	37.750	43.708	41.708	38.958	37.000	41.667	43.583	42.333	41.083	47.667	51.708	51.000	49.000	56.917	56.333
tic-tac-toe		78.502	78.712	80.907	86.634	87.477	86.435	90.919	96.036	90.611	88.732	94.781	98.330	95.198	92.177	97.598	98.641
vehicle		66.203	66.326	71.175	69.273	71.052	70.920	72.944	75.661	72.825	69.864	75.190	73.529	72.469	71.164	76.361	76.480
vote		94.958	94.271	95.180	96.554	95.185	95.418	96.099	96.327	95.872	96.094	96.554	96.099	95.867	96.321	96.099	95.872
vowel		49.697	49.798	60.707	59.091	68.485	65.051	77.980	80.101	77.374	75.960	86.162	91.212	84.949	83.838	92.222	95.758
waveform		83.580	83.320	83.980	84.620	84.180	84.100	84.640	84.880	84.640	84.420	85.400	84.960	85.140	85.100	85.380	85.220
wine		86.569	84.869	85.458	96.667	93.791	95.458	96.111	98.333	96.111	94.967	96.634	97.745	96.667	96.634	96.634	97.745
wisconsin-breast		94.849	94.563	95.422	96.708	95.708	95.275	95.994	96.280	95.849	95.708	95.994	96.280	96.137	95.565	96.280	96.280
zoo		75.273	75.273	75.273	75.273	83.273	82.273	82.273	91.182	87.273	86.273	88.273	95.091	88.273	88.273	89.273	94.273
Average		74.666	74.500	76.772	78.038	78.909	78.737	80.614	81.839	80.962	80.692	82.244	83.435	81.928	81.968	83.742	84.294

The superiority of the proposed combination scheme regarding the classification accuracy is prominent. The following important observations are derived from the accuracy tables:

• The proposed combination method outperforms all other four learning methods in all four labeled ratios and for all the nine base learners used as control methods, in terms of average accuracy. This argument is also validated in Figure 3, where the comparisons are visually assembled and a progressive picture of the performance of the two dominant methods is presented as the R increases. The SL method was also included as a baseline performance metric.
  

  Figure 3. Performance comparison of the proposed combination scheme, in terms of average accuracies over fifty-five datasets and four labeled ratios, against the corresponding methods of active learning (AL) and supervised learning (SL) for the nine base classifiers.
• It is observed by the accuracy tables that the proposed method steadily produces significantly more wins on each individual dataset through all the experiments carried out.

Following the accuracy examination, the Friedman aligned ranks test [51] was conducted. In Table 6, Table 7, Table 8, Table 9, Table 10, Table 11, Table 12, Table 13 and Table 14, the results of the statistical tests for each one of the nine base learners divided into the four labeled ratios used, are presented. These lead to the following assumptions:

Table 6. Friedman aligned ranking test and Holm’s post hoc test regarding BagDT (a = 0.10).

Labeled Ratio (R)	Classifier (BagDT)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (76.02618)	55.83636	-	-
	Active Random		81.33636	0.03566	rejected
	Supervised		148.77273	0.00000	rejected
	Semi-supervised		156.05455	0.00000	rejected
20%	Combination	0.00000 (59.02259)	57.93636	-	-
	Active Random		91.92727	0.00510	rejected
	Supervised		141.07273	0.00000	rejected
	Semi-supervised		151.06364	0.00000	rejected
30%	Combination	0.00000 (78.32732)	48.57273	-	-
	Active Random		91.35455	0.00042	rejected
	Supervised		148.44545	0.00000	rejected
	Semi-supervised		153.62727	0.00000	rejected
40%	Combination	0.00000 (65.36953)	61.73636	-	-
	Active Random		85.82727	0.04717	rejected
	Supervised		129.42727	0.00000	rejected
	Semi-supervised		165.00909	0.00000	rejected

Table 7. Friedman aligned ranking test and Holm’s post hoc test regarding RF (a = 0.10).

Labeled Ratio (R)	Classifier (Random Forests)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (90.74521)	50.23636	-	-
	Active Random		80.05455	0.01403	rejected
	Supervised		153.30000	0.00000	rejected
	Semi-supervised		158.40909	0.00000	rejected
20%	Combination	0.00000 (76.85983)	52.23636	-	-
	Active Random		88.34545	0.00293	rejected
	Supervised		142.39091	0.00000	rejected
	Semi-supervised		159.02727	0.00000	rejected
30%	Combination	0.00000 (76.55845)	55.79091	-	-
	Active Random		83.12727	0.02432	rejected
	Supervised		140.62727	0.00000	rejected
	Semi-supervised		162.45455	0.00000	rejected
40%	Combination	0.00000 (59.66724)	61.71818	-	-
	Active Random		90.20000	0.01895	rejected
	Supervised		129.20000	0.00000	rejected
	Semi-supervised		160.88182	0.00000	rejected

Table 8. Friedman aligned ranking test and Holm’s post hoc test regarding RotF (a = 0.10).

Labeled Ratio (R)	Classifier (Rotation Forest)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (86.92304)	53.12727	-	-
	Active Random		78.83636	0.03417	rejected
	Semi-supervised		149.38182	0.00000	rejected
	Supervised		160.65455	0.00000	rejected
20%	Combination	0.00000 (68.42331)	53.53636	-	-
	Active Random		91.30909	0.00186	rejected
	Supervised		148.11818	0.00000	rejected
	Semi-supervised		149.03636	0.00000	rejected
30%	Combination	0.00000 (61.06200)	54.55455	-	-
	Active Random		95.74545	0.00069	rejected
	Semi-supervised		145.80909	0.00000	rejected
	Supervised		145.89091	0.00000	rejected
40%	Combination	0.00000 (71.23507)	56.67273	-	-
	Active Random		84.93636	0.01989	rejected
	Semi-supervised		141.01818	0.00000	rejected
	Supervised		159.37273	0.00000	rejected

Table 9. Friedman aligned ranking test and Holm’s post hoc test regarding extreme gradient boosted trees (XGBoost) (a = 0.10).

Labeled Ratio (R)	Classifier (XGBoost)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (100.32586)	48.66364	-	-
	Active Random		75.80000	0.02538	rejected
	Supervised		153.99091	0.00000	rejected
	Semi-supervised		163.54545	0.00000	rejected
20%	Combination	0.00000 (79.07341)	53.10000	-	-
	Active Random		83.52727	0.01218	rejected
	Semi-supervised		149.70000	0.00000	rejected
	Supervised		155.67273	0.00000	rejected
30%	Combination	0.00000 (64.21611)	53.01818	-	-
	Active Random		95.96364	0.00040	rejected
	Supervised		144.96364	0.00000	rejected
	Semi-supervised		148.05455	0.00000	rejected
40%	Combination	0.00000 (73.61879)	52.05455	-	-
	Active Random		89.32727	0.00214	rejected
	Supervised		147.56364	0.00000	rejected
	Semi-supervised		153.05455	0.00000	rejected

Table 10. Friedman aligned ranking test and Holm’s post hoc test regarding voting (RF, RotF, XGBoost) (a = 0.10).

Labeled Ratio (R)	Classifier (Voting (RF,RotF,XGBoost))	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (94.26061)	47.87273	-	-
	Active Random		80.97273	0.00639	rejected
	Supervised		155.69091	0.00000	rejected
	Semi-supervised		157.46364	0.00000	rejected
20%	Combination	0.00000 (84.82332)	47.57273	-	-
	Active Random		87.92727	0.00089	rejected
	Supervised		151.62727	0.00000	rejected
	Semi-supervised		154.87273	0.00000	rejected
30%	Combination	0.00000 (71.00226)	53.01818	-	-
	Active Random		89.65455	0.00254	rejected
	Supervised		145.72727	0.00000	rejected
	Semi-supervised		153.60000	0.00000	rejected
40%	Combination	0.00000 (76.77322)	58.08182	-	-
	Active Random		78.40909	0.09400	rejected
	Semi-supervised		150.10909	0.00000	rejected
	Supervised		155.40000	0.00000	rejected

Table 11. Friedman aligned ranking test and Holm’s post hoc test regarding k-nearest neighbors (5NN) (a = 0.10).

Labeled Ratio (R)	Classifier (5NN)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-value	Null Hypothesis
10%	Combination	0.00000 (71.86930)	58.98182	-	-
	Active Random		81.24545	0.06662	rejected
	Supervised		141.71818	0.00000	rejected
	Semi-supervised		160.05455	0.00000	rejected
20%	Combination	0.00000 (73.09232)	56.09091	-	-
	Active Random		84.64545	0.01865	rejected
	Supervised		140.45455	0.00000	rejected
	Semi-supervised		160.80909	0.00000	rejected
30%	Combination	0.00000 (62.25286)	57.79091	-	-
	Active Random		89.60000	0.00878	rejected
	Supervised		143.74545	0.00000	rejected
	Semi-supervised		150.86364	0.00000	rejected
40%	Combination	0.00000 (74.33081)	57.08182	-	-
	Active Random		81.72727	0.04231	rejected
	Supervised		144.50909	0.00000	rejected
	Semi-supervised		158.68182	0.00000	rejected

Table 12. Friedman aligned ranking test and Holm’s post hoc test regarding logistic (a = 0.10).

Labeled Ratio (R)	Classifier (Logistic)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (49.05320)	64.44545	-	-
	Active Random		90.40000	0.03249	rejected
	Semi-supervised		142.15455	0.00000	rejected
	Supervised		145.00000	0.00000	rejected
20%	Combination	0.00000 (59.73571)	58.75455	-	-
	Active Random		90.32727	0.00929	rejected
	Semi-supervised		143.43636	0.00000	rejected
	Supervised		149.48182	0.00000	rejected
30%	Combination	0.00000 (40.03025)	71.25455	-	-
	Active Random		89.02727	0.14314	accepted
	Supervised		139.65455	0.00000	rejected
	Semi-supervised		142.06364	0.00000	rejected
40%	Combination	0.00000 (65.16112)	61.48182	-	-
	Active Random		81.70909	0.09563	rejected
	Supervised		146.31818	0.00000	rejected
	Semi-supervised		152.49091	0.00000	rejected

Table 13. Friedman aligned ranking Test and Holm’s post hoc test regarding logistic model tree (LMT) (a = 0.10).

Labeled Ratio (R)	Classifier (LMT)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (74.72391)	55.78182	-	-
	Active Random		82.53636	0.02751	rejected
	Supervised		150.05455	0.00000	rejected
	Semi-supervised		153.62727	0.00000	rejected
20%	Combination	0.00000 (76.73213)	59.54545	-	-
	Active Random		76.80909	0.15495	accepted
	Semi-supervised		148.55455	0.00000	rejected
	Supervised		157.09091	0.00000	rejected
30%	Combination	0.00000 (50.01495)	56.80000	-	-
	Active Random		103.50909	0.00012	rejected
	Semi-Supervised		139.15455	0.00000	rejected
	Supervised		142.53636	0.00000	rejected
40%	Combination	0.00000 (79.76665)	56.98182	-	-
	Active Random		77.71818	0.08757	rejected
	Semi-Supervised		147.13636	0.00000	rejected
	Supervised		160.16364	0.00000	rejected

Table 14. Friedman aligned ranking test and Holm’s post hoc test regarding LogitBoost (a = 0.10).

Labeled Ratio (R)	Classifier (LogitBoost)	Friedman p-Value (Statistic)	Friedman Ranking	Holm’s Post Hoc Test
				p-Value	Null Hypothesis
10%	Combination	0.00000 (75.28847)	52.08182	-	-
	Active Random		87.89091	0.00318	rejected
	Semi-supervised		149.69091	0.00000	rejected
	Supervised		152.33636	0.00000	rejected
20%	Combination	0.00000 (68.38871)	60.38182	-	-
	Active Random		80.80909	0.09239	rejected
	Supervised		147.73636	0.00000	rejected
	Semi-supervised		153.07273	0.00000	rejected
30%	Combination	0.00000 (60.68458)	53.25455	-	-
	Active Random		99.91818	0.00012	rejected
	Supervised		136.14545	0.00000	rejected
	Semi-supervised		152.68182	0.00000	rejected
40%	Combination	0.00000 (46.30018)	70.32727	-	-
	Active Random		86.01818	0.19611	accepted
	Supervised		138.06364	0.00000	rejected
	Semi-supervised		147.59091	0.00000	rejected

• The non-parametric tests assess the null hypothesis that the means of the results of two or more of the compared methods are the same by calculating the related p-value. This hypothesis can be rejected for all the nine algorithms and for all labeled ratios as all calculated p-values are significantly lower than the significance level of a = 0.10.
• Moreover, the Friedman rankings confirm that for all nine base learners and regardless of the labeled ratio, the proposed combination scheme ranks first ahead of all other learning methods in coincidence with the accuracy experimental results.

Since the Friedman test null hypothesis was rejected, the Holm’s [52] post-hoc statistical test was also applied with an alpha value of 0.10. The aim of the Holm’s test is to detect the specific differences between the combination scheme and the other learning methods, thus the null hypothesis under evaluation is that the mean of the results of the proposed method and against each other group is equal (compared in pairs). The post-hoc results are also presented in the corresponding ranking test tables for each one of the base learners. By observing the adjusted p-values of the Holm’s tests, it is concluded that:

• The proposed combination method performs significantly better on 105 of the total 108 compared method variations for the nine base learners over the four labeled ratios.
• The AL methods for the Logistic, the LMT and the LogitBoost classifiers accept the mean significant difference test for one label ratio each, 30%, 20%, 40% accordingly. However, the adjusted p-values show small differences over the alpha of 0.10.

Summarizing the test results, both Friedman Aligned Ranks tests and Holm’s one vs all comparison tests verify the superior performance of the proposed method over a wide range of scenarios and algorithm comparisons.

To better observe the individual results regarding the combination schemes and the role of the base learners incorporated, the average accuracies were plotted in Figure 4. The outcome was as expected the following: The ensemble voting (RF, RotF, XGBoost) classifier outperforms the rest models in all labeled ratios. As the first indication of such an outcome, the improved prediction probabilities derived from the averaging of the three classifier probabilities, on which the combination scheme relies, it would be a promising starting point for seeking a robust proof to strictly explain the performance boost. Thus, on the one hand, the most confusing unlabeled instances, through the entropy calculation, and on the other hand, the most confident unlabeled instances, through the distribution of prediction probabilities, are detected using the distribution of prediction probabilities. Such behaviors seem to also emerge in other relevant ensemble wrapper algorithms [53].

Figure 4. Average accuracies for the proposed combination scheme regarding different base learners and labeled ratios.

5. Modification

Pointing towards the improvement of the proposed method, it is obvious by the statistical analysis and ranking results that a slight increase in the performance of the SSL part could have a significant impact on the overall efficiency of the combination scheme.

In this direction, careful observation of the execution of the proposed algorithm revealed the weakness of the SSL prediction probabilities, which, in many cases, leads to the selection of the wrong instances to be labeled. In order to augment the probabilistic information available for the proposed method, as regards the SSL part, a lazy classifier (kNN) was integrated into the instance selection process. Such a development, on the one hand, augments the proposed method with a second view of the labels for the unlabeled set, and on the other hand, does not significantly increase the computational overhead as this family of classifiers does not need training. As a second measure to strengthen the SSL instance selection criteria, the empirical approach of setting a lower limit on the minimum accepted probability for an unlabeled instance was adopted using the formula:

$$probaThreshold=\frac{num\_classes(D)+1}{2*num\_classes(D)}$$

(4)

whereby utilizing the $num\_classes\left(X\right)$ function the dependence on the dataset characteristics is lifted. A more compact representation of the SSL part modifications is given in Algorithm 2, while the abstract flow chart of the improved combination framework is presented in Figure 5.

Figure 5. Graphical abstract of the proposed combination framework after the introduction of the semi-supervised learning (SSL) improvements.

Algorithm 2: SSL modification

10: [Execute Algorithm 1 steps (until Alg. 1 line 10)] 11:  IF i modulo 2 == 0: 12:   SET probaThreshold = [num_classes(D) + 1]/[2 * num_classes(D)] 13:   SET the number of nearest neighbors numNeib 14:   INITIALIZE the NN classifier on Li using numNeib 15:    16:   Classify(Ui) using CLS and construct matrix Mpr containing corresponding prediction probabilities along with the predicted labels 17:   FOR_EACH instance of Mpr: 18:    IF cls_predicted_class(instance) != nn_predicted_class(instance)    OR cls_probability(instance) < probaThreshold: 19:     DISCARD instance from Mpr 20:    END_IF 21:   END_FOR_EACH 22:   SORT Mpr descending according to the prediction probabilities 23:   STORE the top maxUnlabPerIter instances of Mpr in a matrix Mfinal 24:   /* now containing the most confident instances along with their predictions */ 25: [Continue Algorithm 1 steps (from Alg. 1 line 16)]

The improved combination scheme was further tested against the most robust AL frameworks found in the literature. In detail, the query strategies of least confidence (LC), margin sampling (MS) and entropy sampling (ES) were considered to be compared with the modified proposed scheme. The major aspects concerning these strategies [7] follow below.

• LC: The objective of this strategy is to identify the least confident unlabeled instances by examining the probability of the most probable label for each unlabeled instance. The strategy continues by selecting the instances having the lowest probable labels and presents them to the human expert to be labeled in order to augment the initial labeled set.

• MS: As an improvement of the LC strategy, MS attempts to overcome the disadvantageous selection process of only considering the most probable labels by calculating the differences of the most probable and the second most probable label for an unlabeled instance. Afterwards, those calculated differences are sorted and the instances with the lowest differences are selected to be labialized.

• ES: This strategy, part of which is also integrated into the AL counterpart of the proposed scheme, computes the entropy measure (similar to Equation (3)) for each unlabeled instance using the distribution of prediction probabilities. The most entropic instances are then selected to be displayed to the human expert in order to enlarge the original labeled set.

In Figure 6, ten experiments display the performance comparison in terms of classification accuracy regarding the three AL methods against the modified combination scheme. The experiments are categorized by the five base learner models that were integrated into the methods. In each experiment, a different benchmark dataset was deployed using four different Rs equal to 10%, 20%, 30%, and 40% accordingly.

Figure 6. Progression of accuracies for the modified combination scheme against the four AL strategies (least confidence (LC), margin sampling (MS), entropy sampling (ES)) for five different base learners on ten different benchmark datasets.

The experimental results confirm the efficiency of the modified combination scheme against the AL methods. It can be extracted from the figure that the proposed technique in all ten cases performs equally or better from its rivals’ accuracies. Moreover, the figure suggests that the three AL methods produce closely related accuracy results, as in four of the ten test cases, their performance was almost identical. The previous outcome can be explained by exploring the metrics utilized in these strategies, which are all derived from the prediction probabilities of the base learners.

Closing this section, in the conducted experiments on real-world benchmark datasets, the proposed combination scheme was compared with the SL, the SSL, and the AL methods. The experiments show that the proposed method outperforms the compared methods. Therefore, in the future, it is very important to conduct more insightful theoretical analyses on the effectiveness of the proposed approach and explore other appropriate selection criteria for filtering the informative unlabeled instances, in order to generalize the results with more confidence.

6. Conclusions

In this research work, a new wrapper algorithm was proposed combining the AL and SSL methods with the aim of efficiently utilizing the available unlabeled data. A plethora of experiments was conducted for evaluating the efficacy of the proposed algorithm in a wide range of benchmark datasets against other learning methods using a variety of classifiers as base models. In addition, four different labeled ratios were investigated. The proposed algorithm prevails over the other learning methods as statistically confirmed by the Friedman aligned ranks non-parametric tests and the Holm’s post-hoc tests. To further promote the use of the proposed algorithm, a software package was developed while more details about this package can be obtained from the link found in the Appendix A.

Regarding the performance boost that was experimentally observed while applying the proposed combination scheme on the numerous datasets, there is strong evidence that the vigorous AL method can efficiently improve its performance utilizing SSL schemes such as the self-training technique. Even in cases were the individual SSL method was not performing dexterously; when integrated in the AL and SSL proposed wrapper the performance of the overall scheme was significantly improved compared to the plain AL method. Moreover, in the case that the majority of the instances used in a learning scheme are automatically labeled, the performance may be unsatisfactory, and in some cases, it may even be worse than the SL baseline accuracy. For this reason, a fundamental requirement arises; that of defining a sufficient threshold of human expert intervention on the labeling process to successfully combine AL and SSL methods. Such a fine-tuning process is criticized as highly application-specific and challenging to automate. Furthermore, it can be noticed by the results, that on datasets with very small initial labeled sets, the proposed scheme can be beneficiary as the initially learned decision boundaries of such datasets can be possibly inaccurate, thus unlabeled instances near these boundaries could be falsely classified. This is an implication that the AL part of the proposed scheme could efficiently tackle.

For future work, a number of areas have been identified and are worth exploring as they seem promising in the direction of improving the classification abilities of the proposed algorithm. As a major first research area, that is expected to have a high impact on the combination scheme’s performance in terms of accuracy and execution time would be the investigation of different instance selection strategies than those that are currently employed. In the AL part of the proposed algorithm, two common alternatives are the least confidence [54] and the margin sampling [55] algorithms, which utilize the unlabeled data under a different scope. Moreover, more complex query scenarios than the plain pool-based sampling used, like query synthesis [56] could also be beneficial. As regards the semi-supervised part, simple techniques like the integration of weights annotating the instances assessed as informative by the SSL part of the algorithm could further improve the overall accuracy of the combination scheme as suggested in [35,57].

Another interesting research area would be that of the extreme outlier detection algorithms. The incorporation of such algorithms in the proposed algorithm would have an immediate impact on the quality of the selected candidate unlabeled instances that are used to augment the labeled set in each self-training iteration, thus resulting in more robust inner models. A few of the very well-known techniques that could be directly implemented in the combination scheme are the local outlier factor [58] for detecting anomalous values based on neighboring data or the isolation forest [59], which is a tree-based outlier detector.

Other research areas that would bear further improvement to the proposed algorithm include preprocessing algorithms, for instance, PCA for dimensionality reduction and production of more informative features or other feature selection techniques such as univariate feature selection [60]. Speaking of the integrated base learners, the introduction of online learners like the Hoeffding adaptive tree [61] and Pegasos [62] or deep learning architectures based on deep neural networks [63] and deep ensembles [64] could make the proposed algorithm sufficient for tackling streaming and big data problems.

Finally, by combining schemes from the fields of active regression learning [65,66] and semi-supervised regression [53] along with the proposed classification algorithm, a general combination scheme could be put forward that would be able to handle numeric and categorical targets.