• Journal of Intelligence and Information Systems
• /
• v.19 no.2
• /
• pp.55-71
• /
• 2013

The prediction model is the main factor affecting the performance of a knowledge-based system for bankruptcy prediction. Earlier studies on prediction modeling have focused on the building of a single best model using statistical and artificial intelligence techniques. However, since the mid-1980s, integration of multiple techniques (hybrid techniques) and, by extension, combinations of the outputs of several models (ensemble techniques) have, according to the experimental results, generally outperformed individual models. An ensemble is a technique that constructs a set of multiple models, combines their outputs, and produces one final prediction. The way in which the outputs of ensemble members are combined is one of the important issues affecting prediction accuracy. A variety of combination schemes have been proposed in order to improve prediction performance in ensembles. Each combination scheme has advantages and limitations, and can be influenced by domain and circumstance. Accordingly, decisions on the most appropriate combination scheme in a given domain and contingency are very difficult. This paper proposes a confidence-based selection approach as part of an ensemble bankruptcy-prediction scheme that can measure unified confidence, even if ensemble members produce different types of continuous-valued outputs. The present experimental results show that when varying the number of models to combine, according to the creation type of ensemble members, the proposed combination method offers the best performance in the ensemble having the largest number of models, even when compared with the methods most often employed in bankruptcy prediction.

• Journal of Internet of Things and Convergence
• /
• v.8 no.6
• /
• pp.55-63
• /
• 2022

In recent years, with the rapid development of large and medium-sized urban rail transit in China, the total operating mileage of high-speed railway and the total number of EMUs(Electric Multiple Units) are rising. The system complexity of high-speed EMU is constantly increasing, which puts forward higher requirements for the safety of equipment and the efficiency of maintenance.At present, the maintenance mode of high-speed EMU in China still adopts the post maintenance method based on planned maintenance and fault maintenance, which leads to insufficient or excessive maintenance, reduces the efficiency of equipment fault handling, and increases the maintenance cost. Based on the intelligent operation and maintenance technology of PHM(prognostics and health management). This thesis builds an integrated PHM platform of "vehicle system-communication system-ground system" by integrating multi-source heterogeneous data of different scenarios of high-speed EMU, and combines the equipment fault mechanism with artificial intelligence algorithms to build a fault prediction model for traction motors of high-speed EMU.Reliable fault prediction and accurate maintenance shall be carried out in advance to ensure safe and efficient operation of high-speed EMU.

• Journal of Intelligence and Information Systems
• /
• v.19 no.2
• /
• pp.125-140
• /
• 2013

We call a data set in which the number of records belonging to a certain class far outnumbers the number of records belonging to the other class, 'imbalanced data set'. Most of the classification techniques perform poorly on imbalanced data sets. When we evaluate the performance of a certain classification technique, we need to measure not only 'accuracy' but also 'sensitivity' and 'specificity'. In a customer churn prediction problem, 'retention' records account for the majority class, and 'churn' records account for the minority class. Sensitivity measures the proportion of actual retentions which are correctly identified as such. Specificity measures the proportion of churns which are correctly identified as such. The poor performance of the classification techniques on imbalanced data sets is due to the low value of specificity. Many previous researches on imbalanced data sets employed 'oversampling' technique where members of the minority class are sampled more than those of the majority class in order to make a relatively balanced data set. When a classification model is constructed using this oversampled balanced data set, specificity can be improved but sensitivity will be decreased. In this research, we developed a hybrid model of support vector machine (SVM), artificial neural network (ANN) and decision tree, that improves specificity while maintaining sensitivity. We named this hybrid model 'hybrid SVM model.' The process of construction and prediction of our hybrid SVM model is as follows. By oversampling from the original imbalanced data set, a balanced data set is prepared. SVM_I model and ANN_I model are constructed using the imbalanced data set, and SVM_B model is constructed using the balanced data set. SVM_I model is superior in sensitivity and SVM_B model is superior in specificity. For a record on which both SVM_I model and SVM_B model make the same prediction, that prediction becomes the final solution. If they make different prediction, the final solution is determined by the discrimination rules obtained by ANN and decision tree. For a record on which SVM_I model and SVM_B model make different predictions, a decision tree model is constructed using ANN_I output value as input and actual retention or churn as target. We obtained the following two discrimination rules: 'IF ANN_I output value <0.285, THEN Final Solution = Retention' and 'IF ANN_I output value ${\geq}0.285$, THEN Final Solution = Churn.' The threshold 0.285 is the value optimized for the data used in this research. The result we present in this research is the structure or framework of our hybrid SVM model, not a specific threshold value such as 0.285. Therefore, the threshold value in the above discrimination rules can be changed to any value depending on the data. In order to evaluate the performance of our hybrid SVM model, we used the 'churn data set' in UCI Machine Learning Repository, that consists of 85% retention customers and 15% churn customers. Accuracy of the hybrid SVM model is 91.08% that is better than that of SVM_I model or SVM_B model. The points worth noticing here are its sensitivity, 95.02%, and specificity, 69.24%. The sensitivity of SVM_I model is 94.65%, and the specificity of SVM_B model is 67.00%. Therefore the hybrid SVM model developed in this research improves the specificity of SVM_B model while maintaining the sensitivity of SVM_I model.

• Journal of Intelligence and Information Systems
• /
• v.25 no.3
• /
• pp.201-220
• /
• 2019

Fake news has emerged as a significant issue over the last few years, igniting discussions and research on how to solve this problem. In particular, studies on automated fact-checking and fake news detection using artificial intelligence and text analysis techniques have drawn attention. Fake news detection research entails a form of document classification; thus, document classification techniques have been widely used in this type of research. However, document summarization techniques have been inconspicuous in this field. At the same time, automatic news summarization services have become popular, and a recent study found that the use of news summarized through abstractive summarization has strengthened the predictive performance of fake news detection models. Therefore, the need to study the integration of document summarization technology in the domestic news data environment has become evident. In order to examine the effect of extractive summarization on the fake news detection model, we first summarized news articles through extractive summarization. Second, we created a summarized news-based detection model. Finally, we compared our model with the full-text-based detection model. The study found that BPN(Back Propagation Neural Network) and SVM(Support Vector Machine) did not exhibit a large difference in performance; however, for DT(Decision Tree), the full-text-based model demonstrated a somewhat better performance. In the case of LR(Logistic Regression), our model exhibited the superior performance. Nonetheless, the results did not show a statistically significant difference between our model and the full-text-based model. Therefore, when the summary is applied, at least the core information of the fake news is preserved, and the LR-based model can confirm the possibility of performance improvement. This study features an experimental application of extractive summarization in fake news detection research by employing various machine-learning algorithms. The study's limitations are, essentially, the relatively small amount of data and the lack of comparison between various summarization technologies. Therefore, an in-depth analysis that applies various analytical techniques to a larger data volume would be helpful in the future.

• Journal of Intelligence and Information Systems
• /
• v.23 no.2
• /
• pp.71-88
• /
• 2017

Language models were originally developed for speech recognition and language processing. Using a set of example sentences, a language model predicts the next word or character based on sequential input data. N-gram models have been widely used but this model cannot model the correlation between the input units efficiently since it is a probabilistic model which are based on the frequency of each unit in the training set. Recently, as the deep learning algorithm has been developed, a recurrent neural network (RNN) model and a long short-term memory (LSTM) model have been widely used for the neural language model (Ahn, 2016; Kim et al., 2016; Lee et al., 2016). These models can reflect dependency between the objects that are entered sequentially into the model (Gers and Schmidhuber, 2001; Mikolov et al., 2010; Sundermeyer et al., 2012). In order to learning the neural language model, texts need to be decomposed into words or morphemes. Since, however, a training set of sentences includes a huge number of words or morphemes in general, the size of dictionary is very large and so it increases model complexity. In addition, word-level or morpheme-level models are able to generate vocabularies only which are contained in the training set. Furthermore, with highly morphological languages such as Turkish, Hungarian, Russian, Finnish or Korean, morpheme analyzers have more chance to cause errors in decomposition process (Lankinen et al., 2016). Therefore, this paper proposes a phoneme-level language model for Korean language based on LSTM models. A phoneme such as a vowel or a consonant is the smallest unit that comprises Korean texts. We construct the language model using three or four LSTM layers. Each model was trained using Stochastic Gradient Algorithm and more advanced optimization algorithms such as Adagrad, RMSprop, Adadelta, Adam, Adamax, and Nadam. Simulation study was done with Old Testament texts using a deep learning package Keras based the Theano. After pre-processing the texts, the dataset included 74 of unique characters including vowels, consonants, and punctuation marks. Then we constructed an input vector with 20 consecutive characters and an output with a following 21st character. Finally, total 1,023,411 sets of input-output vectors were included in the dataset and we divided them into training, validation, testsets with proportion 70:15:15. All the simulation were conducted on a system equipped with an Intel Xeon CPU (16 cores) and a NVIDIA GeForce GTX 1080 GPU. We compared the loss function evaluated for the validation set, the perplexity evaluated for the test set, and the time to be taken for training each model. As a result, all the optimization algorithms but the stochastic gradient algorithm showed similar validation loss and perplexity, which are clearly superior to those of the stochastic gradient algorithm. The stochastic gradient algorithm took the longest time to be trained for both 3- and 4-LSTM models. On average, the 4-LSTM layer model took 69% longer training time than the 3-LSTM layer model. However, the validation loss and perplexity were not improved significantly or became even worse for specific conditions. On the other hand, when comparing the automatically generated sentences, the 4-LSTM layer model tended to generate the sentences which are closer to the natural language than the 3-LSTM model. Although there were slight differences in the completeness of the generated sentences between the models, the sentence generation performance was quite satisfactory in any simulation conditions: they generated only legitimate Korean letters and the use of postposition and the conjugation of verbs were almost perfect in the sense of grammar. The results of this study are expected to be widely used for the processing of Korean language in the field of language processing and speech recognition, which are the basis of artificial intelligence systems.

• Journal of Intelligence and Information Systems
• /
• v.21 no.3
• /
• pp.101-116
• /
• 2015

Recently, due to the development of smart devices and social media, vast amounts of information with the various forms were accumulated. Particularly, considerable research efforts are being directed towards analyzing unstructured big data to resolve various social problems. Accordingly, focus of data-driven decision-making is being moved from structured data analysis to unstructured one. Also, in the field of recommendation system, which is the typical area of data-driven decision-making, the need of using unstructured data has been steadily increased to improve system performance. Approaches to improve the performance of recommendation systems can be found in two aspects- improving algorithms and acquiring useful data with high quality. Traditionally, most efforts to improve the performance of recommendation system were made by the former approach, while the latter approach has not attracted much attention relatively. In this sense, efforts to utilize unstructured data from variable sources are very timely and necessary. Particularly, as the interests of users are directly connected with their needs, identifying the interests of the user through unstructured big data analysis can be a crew for improving performance of recommendation systems. In this sense, this study proposes the methodology of improving recommendation system by measuring interests of the user. Specially, this study proposes the method to quantify interests of the user by analyzing user's internet usage patterns, and to predict user's repurchase based upon the discovered preferences. There are two important modules in this study. The first module predicts repurchase probability of each category through analyzing users' purchase history. We include the first module to our research scope for comparing the accuracy of traditional purchase-based prediction model to our new model presented in the second module. This procedure extracts purchase history of users. The core part of our methodology is in the second module. This module extracts users' interests by analyzing news articles the users have read. The second module constructs a correspondence matrix between topics and news articles by performing topic modeling on real world news articles. And then, the module analyzes users' news access patterns and then constructs a correspondence matrix between articles and users. After that, by merging the results of the previous processes in the second module, we can obtain a correspondence matrix between users and topics. This matrix describes users' interests in a structured manner. Finally, by using the matrix, the second module builds a model for predicting repurchase probability of each category. In this paper, we also provide experimental results of our performance evaluation. The outline of data used our experiments is as follows. We acquired web transaction data of 5,000 panels from a company that is specialized to analyzing ranks of internet sites. At first we extracted 15,000 URLs of news articles published from July 2012 to June 2013 from the original data and we crawled main contents of the news articles. After that we selected 2,615 users who have read at least one of the extracted news articles. Among the 2,615 users, we discovered that the number of target users who purchase at least one items from our target shopping mall 'G' is 359. In the experiments, we analyzed purchase history and news access records of the 359 internet users. From the performance evaluation, we found that our prediction model using both users' interests and purchase history outperforms a prediction model using only users' purchase history from a view point of misclassification ratio. In detail, our model outperformed the traditional one in appliance, beauty, computer, culture, digital, fashion, and sports categories when artificial neural network based models were used. Similarly, our model outperformed the traditional one in beauty, computer, digital, fashion, food, and furniture categories when decision tree based models were used although the improvement is very small.

• Journal of Intelligence and Information Systems
• /
• v.27 no.1
• /
• pp.23-46
• /
• 2021

Collaborative filtering, which is often used in personalization recommendations, is recognized as a very useful technique to find similar customers and recommend products to them based on their purchase history. However, the traditional collaborative filtering technique has raised the question of having difficulty calculating the similarity for new customers or products due to the method of calculating similaritiesbased on direct connections and common features among customers. For this reason, a hybrid technique was designed to use content-based filtering techniques together. On the one hand, efforts have been made to solve these problems by applying the structural characteristics of social networks. This applies a method of indirectly calculating similarities through their similar customers placed between them. This means creating a customer's network based on purchasing data and calculating the similarity between the two based on the features of the network that indirectly connects the two customers within this network. Such similarity can be used as a measure to predict whether the target customer accepts recommendations. The centrality metrics of networks can be utilized for the calculation of these similarities. Different centrality metrics have important implications in that they may have different effects on recommended performance. In this study, furthermore, the effect of these centrality metrics on the performance of recommendation may vary depending on recommender algorithms. In addition, recommendation techniques using network analysis can be expected to contribute to increasing recommendation performance even if they apply not only to new customers or products but also to entire customers or products. By considering a customer's purchase of an item as a link generated between the customer and the item on the network, the prediction of user acceptance of recommendation is solved as a prediction of whether a new link will be created between them. As the classification models fit the purpose of solving the binary problem of whether the link is engaged or not, decision tree, k-nearest neighbors (KNN), logistic regression, artificial neural network, and support vector machine (SVM) are selected in the research. The data for performance evaluation used order data collected from an online shopping mall over four years and two months. Among them, the previous three years and eight months constitute social networks composed of and the experiment was conducted by organizing the data collected into the social network. The next four months' records were used to train and evaluate recommender models. Experiments with the centrality metrics applied to each model show that the recommendation acceptance rates of the centrality metrics are different for each algorithm at a meaningful level. In this work, we analyzed only four commonly used centrality metrics: degree centrality, betweenness centrality, closeness centrality, and eigenvector centrality. Eigenvector centrality records the lowest performance in all models except support vector machines. Closeness centrality and betweenness centrality show similar performance across all models. Degree centrality ranking moderate across overall models while betweenness centrality always ranking higher than degree centrality. Finally, closeness centrality is characterized by distinct differences in performance according to the model. It ranks first in logistic regression, artificial neural network, and decision tree withnumerically high performance. However, it only records very low rankings in support vector machine and K-neighborhood with low-performance levels. As the experiment results reveal, in a classification model, network centrality metrics over a subnetwork that connects the two nodes can effectively predict the connectivity between two nodes in a social network. Furthermore, each metric has a different performance depending on the classification model type. This result implies that choosing appropriate metrics for each algorithm can lead to achieving higher recommendation performance. In general, betweenness centrality can guarantee a high level of performance in any model. It would be possible to consider the introduction of proximity centrality to obtain higher performance for certain models.

• Journal of Intelligence and Information Systems
• /
• v.24 no.3
• /
• pp.21-44
• /
• 2018

In recent years, the rapid development of internet technology and the popularization of smart devices have resulted in massive amounts of text data. Those text data were produced and distributed through various media platforms such as World Wide Web, Internet news feeds, microblog, and social media. However, this enormous amount of easily obtained information is lack of organization. Therefore, this problem has raised the interest of many researchers in order to manage this huge amount of information. Further, this problem also required professionals that are capable of classifying relevant information and hence text classification is introduced. Text classification is a challenging task in modern data analysis, which it needs to assign a text document into one or more predefined categories or classes. In text classification field, there are different kinds of techniques available such as K-Nearest Neighbor, Naïve Bayes Algorithm, Support Vector Machine, Decision Tree, and Artificial Neural Network. However, while dealing with huge amount of text data, model performance and accuracy becomes a challenge. According to the type of words used in the corpus and type of features created for classification, the performance of a text classification model can be varied. Most of the attempts are been made based on proposing a new algorithm or modifying an existing algorithm. This kind of research can be said already reached their certain limitations for further improvements. In this study, aside from proposing a new algorithm or modifying the algorithm, we focus on searching a way to modify the use of data. It is widely known that classifier performance is influenced by the quality of training data upon which this classifier is built. The real world datasets in most of the time contain noise, or in other words noisy data, these can actually affect the decision made by the classifiers built from these data. In this study, we consider that the data from different domains, which is heterogeneous data might have the characteristics of noise which can be utilized in the classification process. In order to build the classifier, machine learning algorithm is performed based on the assumption that the characteristics of training data and target data are the same or very similar to each other. However, in the case of unstructured data such as text, the features are determined according to the vocabularies included in the document. If the viewpoints of the learning data and target data are different, the features may be appearing different between these two data. In this study, we attempt to improve the classification accuracy by strengthening the robustness of the document classifier through artificially injecting the noise into the process of constructing the document classifier. With data coming from various kind of sources, these data are likely formatted differently. These cause difficulties for traditional machine learning algorithms because they are not developed to recognize different type of data representation at one time and to put them together in same generalization. Therefore, in order to utilize heterogeneous data in the learning process of document classifier, we apply semi-supervised learning in our study. However, unlabeled data might have the possibility to degrade the performance of the document classifier. Therefore, we further proposed a method called Rule Selection-Based Ensemble Semi-Supervised Learning Algorithm (RSESLA) to select only the documents that contributing to the accuracy improvement of the classifier. RSESLA creates multiple views by manipulating the features using different types of classification models and different types of heterogeneous data. The most confident classification rules will be selected and applied for the final decision making. In this paper, three different types of real-world data sources were used, which are news, twitter and blogs.

• Journal of Intelligence and Information Systems
• /
• v.22 no.2
• /
• pp.127-142
• /
• 2016

Deep learning model is a kind of neural networks that allows multiple hidden layers. There are various deep learning architectures such as convolutional neural networks, deep belief networks and recurrent neural networks. Those have been applied to fields like computer vision, automatic speech recognition, natural language processing, audio recognition and bioinformatics where they have been shown to produce state-of-the-art results on various tasks. Among those architectures, convolutional neural networks and recurrent neural networks are classified as the supervised learning model. And in recent years, those supervised learning models have gained more popularity than unsupervised learning models such as deep belief networks, because supervised learning models have shown fashionable applications in such fields mentioned above. Deep learning models can be trained with backpropagation algorithm. Backpropagation is an abbreviation for "backward propagation of errors" and a common method of training artificial neural networks used in conjunction with an optimization method such as gradient descent. The method calculates the gradient of an error function with respect to all the weights in the network. The gradient is fed to the optimization method which in turn uses it to update the weights, in an attempt to minimize the error function. Convolutional neural networks use a special architecture which is particularly well-adapted to classify images. Using this architecture makes convolutional networks fast to train. This, in turn, helps us train deep, muti-layer networks, which are very good at classifying images. These days, deep convolutional networks are used in most neural networks for image recognition. Convolutional neural networks use three basic ideas: local receptive fields, shared weights, and pooling. By local receptive fields, we mean that each neuron in the first(or any) hidden layer will be connected to a small region of the input(or previous layer's) neurons. Shared weights mean that we're going to use the same weights and bias for each of the local receptive field. This means that all the neurons in the hidden layer detect exactly the same feature, just at different locations in the input image. In addition to the convolutional layers just described, convolutional neural networks also contain pooling layers. Pooling layers are usually used immediately after convolutional layers. What the pooling layers do is to simplify the information in the output from the convolutional layer. Recent convolutional network architectures have 10 to 20 hidden layers and billions of connections between units. Training deep learning networks has taken weeks several years ago, but thanks to progress in GPU and algorithm enhancement, training time has reduced to several hours. Neural networks with time-varying behavior are known as recurrent neural networks or RNNs. A recurrent neural network is a class of artificial neural network where connections between units form a directed cycle. This creates an internal state of the network which allows it to exhibit dynamic temporal behavior. Unlike feedforward neural networks, RNNs can use their internal memory to process arbitrary sequences of inputs. Early RNN models turned out to be very difficult to train, harder even than deep feedforward networks. The reason is the unstable gradient problem such as vanishing gradient and exploding gradient. The gradient can get smaller and smaller as it is propagated back through layers. This makes learning in early layers extremely slow. The problem actually gets worse in RNNs, since gradients aren't just propagated backward through layers, they're propagated backward through time. If the network runs for a long time, that can make the gradient extremely unstable and hard to learn from. It has been possible to incorporate an idea known as long short-term memory units (LSTMs) into RNNs. LSTMs make it much easier to get good results when training RNNs, and many recent papers make use of LSTMs or related ideas.

• Journal of Intelligence and Information Systems
• /
• v.28 no.1
• /
• pp.131-154
• /
• 2022

Recently, researches on unstructured data analysis have been actively conducted with the development of information and communication technology. In particular, topic modeling is a representative technique for discovering core topics from massive text data. In the early stages of topic modeling, most studies focused only on topic discovery. As the topic modeling field matured, studies on the change of the topic according to the change of time began to be carried out. Accordingly, interest in dynamic topic modeling that handle changes in keywords constituting the topic is also increasing. Dynamic topic modeling identifies major topics from the data of the initial period and manages the change and flow of topics in a way that utilizes topic information of the previous period to derive further topics in subsequent periods. However, it is very difficult to understand and interpret the results of dynamic topic modeling. The results of traditional dynamic topic modeling simply reveal changes in keywords and their rankings. However, this information is insufficient to represent how the meaning of the topic has changed. Therefore, in this study, we propose a method to visualize topics by period by reflecting the meaning of keywords in each topic. In addition, we propose a method that can intuitively interpret changes in topics and relationships between or among topics. The detailed method of visualizing topics by period is as follows. In the first step, dynamic topic modeling is implemented to derive the top keywords of each period and their weight from text data. In the second step, we derive vectors of top keywords of each topic from the pre-trained word embedding model. Then, we perform dimension reduction for the extracted vectors. Then, we formulate a semantic vector of each topic by calculating weight sum of keywords in each vector using topic weight of each keyword. In the third step, we visualize the semantic vector of each topic using matplotlib, and analyze the relationship between or among the topics based on the visualized result. The change of topic can be interpreted in the following manners. From the result of dynamic topic modeling, we identify rising top 5 keywords and descending top 5 keywords for each period to show the change of the topic. Existing many topic visualization studies usually visualize keywords of each topic, but our approach proposed in this study differs from previous studies in that it attempts to visualize each topic itself. To evaluate the practical applicability of the proposed methodology, we performed an experiment on 1,847 abstracts of artificial intelligence-related papers. The experiment was performed by dividing abstracts of artificial intelligence-related papers into three periods (2016-2017, 2018-2019, 2020-2021). We selected seven topics based on the consistency score, and utilized the pre-trained word embedding model of Word2vec trained with 'Wikipedia', an Internet encyclopedia. Based on the proposed methodology, we generated a semantic vector for each topic. Through this, by reflecting the meaning of keywords, we visualized and interpreted the themes by period. Through these experiments, we confirmed that the rising and descending of the topic weight of a keyword can be usefully used to interpret the semantic change of the corresponding topic and to grasp the relationship among topics. In this study, to overcome the limitations of dynamic topic modeling results, we used word embedding and dimension reduction techniques to visualize topics by era. The results of this study are meaningful in that they broadened the scope of topic understanding through the visualization of dynamic topic modeling results. In addition, the academic contribution can be acknowledged in that it laid the foundation for follow-up studies using various word embeddings and dimensionality reduction techniques to improve the performance of the proposed methodology.